<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sukriti Paul on Medium]]></title>
        <description><![CDATA[Stories by Sukriti Paul on Medium]]></description>
        <link>https://medium.com/@sukritipaul005?source=rss-18efe080f332------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*87qAlkZLjEOK4HFhMpIwgA.jpeg</url>
            <title>Stories by Sukriti Paul on Medium</title>
            <link>https://medium.com/@sukritipaul005?source=rss-18efe080f332------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 22:31:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sukritipaul005/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[My Second Quarter in the World of Data]]></title>
            <link>https://medium.com/@sukritipaul005/my-second-quarter-in-the-world-of-data-1661510ac2ce?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/1661510ac2ce</guid>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[career-change]]></category>
            <category><![CDATA[business-analytics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[careers]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Mon, 13 Jul 2020 14:31:01 GMT</pubDate>
            <atom:updated>2020-07-14T13:55:49.872Z</atom:updated>
            <content:encoded><![CDATA[<h4>Musings of a Newbie in a Corporate Data Science Team</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*onZ7ajVCIWUGuxOM" /><figcaption>Source: Franki Chamaki, Unsplash</figcaption></figure><p>Not long ago, I penned an <a href="https://medium.com/@sukritipaul005/my-first-quarter-in-a-world-of-data-8783cfffbb16">article</a> describing key takeaways in the first quarter of my new job. As anticipated, the learnings have transformed within a span of a quarter- I have a completely new list now! On interacting with my manager, team, and senior leaders, I’ve learned important lessons, which (I believe) <strong>could benefit every newbie pursuing a Data Analytics/ Data Science role.</strong> Therefore, I’ve compiled a list of learnings; <em>if you breathe data, this might be of use!</em></p><p><em>Note: This article does not contain any company-specific or confidential information. My views do not represent those of my workplace.</em></p><h3>Run It At One Go!</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*OAs8JNUL9LRx8u1R" /><figcaption>Source: Markus Spiske</figcaption></figure><h4>One Well-Planned (Minimally Iterative) Step</h4><p>Data and business use cases tend to become complex over time. As a consequence, there’s significant scope for errors and delays (even for something as simple as on outer join operation). <em>A change in logic could boost your code execution time by several business hours</em>. Inculcate the habit of <strong>allocating a considerable amount of time for building your logic</strong> to avoid repeated executions (I was advised to dedicate 2–3 hours on the logic before diving into any code).</p><ul><li><strong>Code optimization and logic optimization</strong> can be developed by brainstorming problems with teammates, and regularly referring to existing codes written by experienced team members. By no means does this imply that you should suppress the Internet Ninja in you — refer to Stack Overflow, Github Issues, etc.</li><li>While dealing with gargantuan databases, it becomes cumbersome and time-consuming to correct issues over many iterations<strong>. Be perceptive in assessing the glitches that could arise, and make pointers vis-à-vis crucial and commonly-used databases.</strong> <em>Imagine getting an ‘Out of Memory Error’ after running your code the entire night!</em></li><li><strong>Make note of ways to speed-up implementation</strong>. <em>For example, knowing that apply() or count() is way faster than any loop-based operation on a Pandas DataFrame, could save hours during any analysis, given that you’re dealing with data comprising millions of entries.</em></li></ul><h3>Permissions &amp; Documentation Everywhere!</h3><p>Unlike academia, <em>any incremental change in a learning model requires approvals across different verticals.</em> Any change that is anticipated to boost the model performance, must be backed by data (compliance may make things more eventful 😉). Gone are the days when one could make teeny, experimental changes in the model, to discover what clicked; sound research and analysis underlie any decision-making in model design. Although the breadth for experimentation may seem limited, it is a good practice<strong> to design effective tests and invest in extensive model testing</strong><em> </em>(which may include detailed reports comparing model architectures). Once the model is in production, trends like consumer behavior could take months to observe. In a way, rigorous inquiry assures that the product quality is enhanced while upgrading to challenger models.</p><h3>Data Pipeline Breaks: The Onus Isn’t on the Tech/Dev Team Alone</h3><p>As a part of a Data Science team, <em>one is equally responsible for coordinating and resolving breaks in the data pipeline.</em> Dig a little deeper: Understand the data flow; how online (implicit and explicit) data is collected from each channel; what kind of treatment or imputation is performed on raw data; how the data is stored and divided into training/testing data across several cells; how to detect anomalies in the ensemble logs; how the tech team optimizes the production code from their end; how the tech team’s tests vary from those of the DS teams; what are the plausible delays in production etc.</p><ul><li><strong>Documenting the data flow diagrammatically </strong>works wonders (given that the data could have a gazillion internal sources)!</li><li><strong>Clean code, crisp user stories, and comprehensive model documentation</strong> ensure that the tech team uses/reproduces the correct logic in the deployment code.</li></ul><h3>Juggling Tasks &amp; Projects</h3><p>Abrupt changes in business plans may call for an urgent analysis or a new use case; one’s current projects could get sidelined in all possibilities.<em> For example, boosting offline retail shopping recommendations on an e-commerce website may not be the best of ideas during a global pandemic.</em> It’s important to adapt to sudden high-priority requirements, and shuffle between older long-term projects resiliently.</p><ul><li><strong>Regular notetaking helps in summarizing progress in each project (i.e. jotting down impact, work done so far, requirements, timelines, etc)</strong>.</li><li>You can also maintain a separate label for e-mails regarding the ‘most recent project updates’.</li></ul><p>At times, one may abandon some analysis and revisit it later, plausibly due to reasons like data scarcity, a lack of campaigns, issues that need to be resolved among external teams (e.g. database decommission), or some behavior/metric that has to be observed over a longer duration of time. If you’re in the habit of working continuously on a project till its completion, you might want to work on need-based task prioritization.</p><h3>Your Data Should Tell a Story</h3><p>Here’s a great quote by Chris Arnold (commonly known as The Data Whisperer) stated during his talk ‘Data Viz in a Box’, held at IISc Bangalore:</p><blockquote>“ The thing with data analysts is that most of them get so engrossed in the complexities, that they fail to consider how their audience will comprehend the visual data presented. If the audience is finding it hard to understand something, it’s because the data analyst has not done a good job. It’s a case of work transfer, your viewer is doing what you should have done.”— Christopher Arnold Wells</blockquote><p>The minute you’re tasked with a presentation, ask yourself: <em>“Will the target audience infer fairly well without too many prompts, repetitions or back-and-forth”?</em><strong> Data storytellers connect the dots to make a convincing case in support of their conclusions- </strong>the content flows spontaneously, cogently, and systematically. They don’t dive into the details of each analysis; instead, they decide which piece of analysis helps in deriving<em> </em><strong><em>actionable insights</em></strong>!</p><ul><li>Mock presentations with team members and your manager can help identify the gaps (before facing a larger audience).</li><li>Keep your eyes peeled for the decks created by senior leadership.</li><li>Revisit and review your content after a time gap.</li></ul><h3>Who Is Your Audience?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*YDq2GrwQFSsJXdGn" /><figcaption>Source: Teemu Paananen (Unsplash)</figcaption></figure><p>Have you come across data analysts harping on, “knowing one’s audience”? I couldn’t stress enough that <strong>every unnecessary detail on your presentation comes at a cost.</strong> <strong><em>And the cost is the viewer’s attention</em></strong>. Remember to present only what is relevant to the audience. During the course of time, data findings may have to be presented across different levels of leadership. My manager elaborated on how she and I are much closer to the data; we could delve into the nitty-gritty of the analysis. However, presenting these findings to a director or VP comes with a level of <strong>abstraction</strong>: The senior leadership will most likely be interested in the <strong>insights/ highlights/ overviews</strong>, rather than data-related nuances! Undoubtedly,<em> learning that level of detail or abstraction comes with experience in the organization.</em></p><p>Data Science teams are constantly in touch with diverse teams (including tech, marketing, product development, data compliance, etc). In such a setting, it goes without saying that <strong>effective communication</strong> is an absolute necessity for getting one’s ideas/thoughts across. Many a time, <em>what you convey signals your clarity of thought, logic, and execution.</em></p><h3>“Everything is Available Because Y’all Have the Data”</h3><h4>Well, Is It?</h4><p>All things don’t come on a silver platter. On the contrary, there’s quite a bit of figuring out to do! Occasionally (or maybe even oftentimes) there are no readily available resources regarding testing processes, documentation, metrics, algorithms, what have you. We’re expected to research improvements to existing solutions, propose new use cases, enhance existing algorithms, and even provide ideas to the product team. Yes, we <strong>ideate and implement</strong>. <em>For example, your team may create a proprietary correlation metric, modify a loss function, or design slightly complex derived features! Or an e-commerce firm may try and find a new way to boost certain personalization-based capabilities for a handful of loyal customers.</em></p><h3>Time Management (Project Planning Phase)</h3><p>Ever underestimated the time required for a task, rather overambitiously? Or overcautiously allotted a lot of buffer days, should something go haywire? While playing with Big Data, it’s perfectly normal to face delays due to code execution, unexpected data ambiguities, processes depending on external teams, or the need for further analysis, etc. However, <strong>reasonable timelines are crucial in coordinating tasks across teams</strong>. <em>For example, the marketing team of an e-commerce firm may plan a campaign for X months. During this time, the learning model has to be designed and deployed — the tech team’s UAT will depend on the model creation and validation timeline.</em></p><ul><li>Walk your manager through a detailed timeline for the first few tasks. Ask her (1) the possible challenges that she can anticipate from her experience, and (2) if your buffer time is realistic.</li><li>Make note of all the instances when you crossed a deadline due to a factor that was beyond your control. This will help in future reference.</li><li>If the task is dependent on other employees, make sure to discuss how much time they would need for their contribution.</li></ul><h3>Learn &amp; Discover On the Go</h3><p>Learning is experiential and non-linear. After the first quarter, SQL, Python/R, and Excel become one’s bread and butter. By this time, the Data Science Tech Stack should be on one’s fingertips. As opposed to my experience in academia, I’ve seen that prolonged learning time may not count as productive business hours in an outcome-driven or result-oriented setting: <strong>Deliverables matter as much as your learning curve.</strong></p><ul><li>A curious mindset coupled with an open-minded attitude towards learning new technologies can go a long way!</li><li>Be flexible in adapting to new platforms. ‘Learning’ ultimately boils down to <strong>how efficiently one can apply any learning to a project.</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*ygbGyup1Y15iRqK2" /><figcaption>Source: Tim Mossholder, Unsplash.</figcaption></figure><h4>Keep Your Eyes Open</h4><p>In a recent leadership webinar, a senior Customer Data Science VP highlighted why it’s important to stay updated on how other businesses are leveraging technology to enhance their products and customer experience (ethically, without a breach of confidentiality of course). Actively engaging with similar products outside of work has helped me appreciate the potential of the field. Sky’s the limit when it comes to innovation; you never know how a product feature could get you thinking. <em>As a general user, if you observe a new feature on Spotify- one that collects user feedback in a unique way- there’s no harm in deriving some inspiration for explicit data collection for your use-case. </em>😉</p><p><em>Also, a shoutout to my managers, who raise the bar every single day, and help bring out the best in me.❤</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1661510ac2ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My First Quarter in a World of Data]]></title>
            <link>https://medium.com/@sukritipaul005/my-first-quarter-in-a-world-of-data-8783cfffbb16?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/8783cfffbb16</guid>
            <category><![CDATA[career-development]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[business-analytics]]></category>
            <category><![CDATA[career-change]]></category>
            <category><![CDATA[data-analytics]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Thu, 05 Mar 2020 18:20:43 GMT</pubDate>
            <atom:updated>2020-03-05T20:28:13.574Z</atom:updated>
            <content:encoded><![CDATA[<h3>My First Quarter in the World of Data</h3><h4>Musings of a Newbie in a Corporate Data Science Team</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/639/0*GR5q9uLnZXgGi0Xi.jpg" /><figcaption>Source: iStock, referenced via <a href="https://www.eecs.mit.edu/news-events/media/what-buyers-want">MIT EECS</a></figcaption></figure><p>Switching careers comes with a pinch of uncertainty, exhilaration, and changes that we homo sapiens can quickly adapt to. Embracing one such change wholeheartedly, I recently switched my job from a Research Assistant at a premier research college in India, to a role in the <strong>Enterprise</strong> <strong>Customer Data Science Team</strong> of a multinational financial services corporation. My journey has been enriching so far: there is scope for learning something new every day.</p><p>Over interactions with my team and senior leaders, I’ve learned important lessons, which (I believe) <strong>could benefit every newbie pursuing a Data Analytics/ Data Science role.</strong> Therefore, I’ve compiled a list of learnings; <em>if you breathe data, this might be of use!</em></p><p><em>Note: This article does not contain any company-specific or confidential information. My views do not represent those of my workplace.</em></p><h3>Brush up Your Tech Basics</h3><p>As you progress into the role, you will have more and more complex use cases to solve for. <em>Imagine your plight if you shoddily revise your basics, only to realise that you’re taking 3X the time to produce results after a year into the role!</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/0*74J4838z8yszmacS.png" /><figcaption>Source: A sample data science tech stack taken from <a href="https://dev.to/minchulkim87/my-data-science-tech-stack-2020-1poa">here.</a></figcaption></figure><p><strong>Utilize the first few months to strengthen your technical skills. </strong>Some skills can be learned on the go, however, there are a few theoretical concepts that need to be revisited now and then. Usually, every company specifies certain technical skills required for the role, based on its <strong>Data Science Tech Stack</strong>. The following list is generic:</p><ol><li><strong>SQL Knowledge is Sacrosanct: </strong>Remember, one needs to manage and work with data (possibly big data), irrespective of the application. Hence, it becomes imperative to revise your SQL basics. Fortunately, SQL is something that you can pick up rather quickly, irrespective of whether you come from a technical background.</li><li><strong>SQL Query Tools (or Querying Engines): </strong>If you’re working with Big Data, you’ll need query tools <em>depending on what your company uses for </em><strong><em>pulling data</em></strong><em>.</em> The tools include <strong>Apache Impala/Hive/Spark, Presto,</strong> what have you.</li><li><strong>Software Frameworks like <em>MapReduce</em>: </strong>Similar frameworks can help in processing large amounts of data.</li><li><strong>Implement those Learning Models</strong>: Familiarize yourself with Python libraries like <strong>Keras/ PyTorch/ Theano/ TF </strong>etc. and(or) deep learning frameworks like<strong> Caffe.</strong></li><li><strong>Working with Data</strong>: While pre-processing data or using it for a meaningful task, some corporates use <strong>Python/R</strong>. I also recommend getting accustomed to the shortcuts in MS Excel for drawing simple insights.</li></ol><h3>Understand the Business (Consumer) Landscape</h3><p>As a Data Scientist/Analyst, your findings will be used in business-related decision-making. <strong>To draw actionable insights from the data or study impact, you need to understand how things fit into the bigger picture! </strong>Figuring out solutions without any context can be likened to preparing a stand-up comedy script without knowing your target audience. Y<em>ou must connect the dots to envision what the entire picture will look like.</em> Take a minute to comprehend the stakes involved if your team decides whether a product/feature should continue or not, based on data!</p><h3>Embrace the Gargantuan Databases</h3><p>One of the best advice that I got, was to spend a good amount of weeks <strong>playing around with different corporate data sets</strong>. Defining your problem statements brings about a different thrill…<em>of discovering something new while drawing inferences.</em> <strong>Explore the gazillion attributes and data sources, find out which ones are important to your team</strong>. Y<em>ou grow with the databases</em>: as new tables are updated, you keep discovering new trends and links! While designing a learning model, you should be able to provide a well-thought-out data set. Creating attributes requires a solid meticulous understanding of the existing data!</p><h3>What is Your Team’s Workflow?</h3><p>As a Data Scientist/Analyst, it’s important to find out the <strong>synergy between different teams</strong> (i.e., how different teams work with yours). Understanding this pipeline is crucial in <strong>realising where you fit in, and how your team’s efforts are contributing to a larger cause.</strong> <em>For example, let us consider that you’re working for an online E-commerce firm. The Marketing and Product teams may design campaigns spanning different durations. After launching these campaigns, data may be collected and analysed by the Data Science team, who further design probabilistic or deterministic models for certain business use cases. Once the learning models are designed and implemented, the findings may be discussed with the Marketing/Product teams. Subsequently, the finalized model may be forwarded to the Tech/Dev team for production or optimization. </em>See how things make more sense when you have a defined purpose in the pipeline?</p><h3>Be a Content Master</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*RK_MibsHV7GKBogZ.png" /><figcaption>Design Credits: <a href="https://dribbble.com/shots/3112439-Lil-BFFs-Po-and-Master-Shifu">Dennis Salvatier</a></figcaption></figure><p>For starters, I’m quoting the “<em>content master</em>” phrase from a 1x1 session with a senior leader (phrase credits). Know your content inside out for a particular task. While analysing data in the initial stages, <em>chalk out the flow of steps that you wish to carry out, on paper. </em><strong>Avoid repeating executions and looping back to previous steps due to flawed logic or ambiguities in the initial problem statement. </strong>Big Data tasks can take hours and days; teeny changes to your code will not reflect in a jiffy!</p><h3>Clarify the Data Flow at the Beginning</h3><p>If you’re tasked with creating new features for a data set, discuss the <strong>requirements and database design flow with your manager</strong>- <em>the immediate goal should be crystal clear. </em>When it comes to<strong> feature engineering</strong>, one must consider that campaigns can run for months. If your algorithm was trained and is being tested on select attributes, you possibly cannot decide to pop in an extra attribute on the control group, as and when! Modifying the data provided to a model is far easier, timewise, in academia.</p><h3>Be Mindful of the Testing Process (Especially for Online Learning Tasks)</h3><p>Just like the Dev team, Data Science teams also have extensive testing performed over prolonged periods. It is worthwhile to <strong>enquire about the different test cells</strong>, including the <strong>control</strong>. <em>How are data points divided into these cells? What are the durations for the same? What if a model is not performing up to the mark, midway? Which cells should the model be designed for? How to avoid overfitting on the optimal test cell?</em> In addition, gain a general idea about <strong>performance metrics and their values for the existing model</strong>(s). On getting the feel of ML for enterprise right after my experience in academia, I observed a considerable difference in what was deemed to be a <em>well-performing algorithm: I experienced what I call a ‘reality versus expectation’ scenario w.r.t metric values and thresholds. </em>😛</p><h3>Learn How to Communicate Findings Effectively</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zImt20nxGvblKnrZ.jpg" /><figcaption>Source: A Nature jobs <a href="http://blogs.nature.com/naturejobs/2017/11/20/why-scientists-should-communicate-science-getting-to-the-heart-of-the-matter/">blog article.</a></figcaption></figure><p>In my second review-meeting, I presented a few results to my leaders.<em> Little did I know that each word would count towards making the team interpret my results. </em>Fortunately, my Team Lead corrected me when there was an ambiguity in what I did versus what I had verbally expressed. Thereon, I realised the importance of <strong>clearly communicating the data flow to a large group of members. </strong><em>A good way to get the terms right is via mock discussions- before the meeting- where you could </em><strong><em>explain the flow to your teammates, and ask them if there’s a better way of organizing your content!</em></strong></p><h3>Count on Your Intuition (Not Fully Though)</h3><p>Oftentimes, managers can view the data and assess if the numbers are what they expected, or if there are far too many outliers. Unless you’ve been a part of your corporate ecosystem for a good amount of time, it becomes challenging to develop such intuition. However, <strong>glaring results will play on your intuition early on.</strong> If you feel that the numbers are not right, <strong><em>run your logic though a more experienced team member. </em></strong>Let’s revisit the E-commerce firm example.<em> If the number of customer transactions is far lesser than the number of merchants, then this could indicate that you’ve gone wrong somewhere (or the product/feature has flopped, which is unlikely). </em>Common sense goes a long way, sometimes.</p><p><em>Note: This is my first attempt at writing a corporate job-related article. Do comment below, if you have any feedback! Thanks for reading the article. :)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8783cfffbb16" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Together We Can Do So Much]]></title>
            <link>https://medium.com/@sukritipaul005/together-we-can-do-so-much-2c4b559c550a?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c4b559c550a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[collaboration]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Tue, 13 Aug 2019 20:14:33 GMT</pubDate>
            <atom:updated>2021-07-31T10:53:09.990Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part 1: My Takeaways from Field Research and Medical Collaboration</h4><p>Before I dive into the subject matter, let me give you some context. I’m a research assistant, working on <strong>designing Deep learning (DL) models and developing fully-automated AI-based diagnostic tools in the area of biomedical imaging</strong>. In short, I architect DL algorithms and pipelines under image classification, object detection, semantic segmentation, instance segmentation and deep clustering.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XrU4XT_aWP2iuIawpTyMIw.png" /><figcaption>Fig.1: Healthworkscollective (Source)</figcaption></figure><p>By now, you must have inferred that I do not have an iota of ‘medicine’ or ‘bio sciences’ in my curriculum vitae . Truth be told, Biology was a subject that I absolutely dreaded at school. You never know what Fortuna and her Wheel have in store for you<em>. Before long, I was making multiple field visits to our partner hospital for clinical validation and data collection</em>.<em> </em>A <strong>perspective shift</strong> was all it took, to embrace the idea of working on a biomedical problem statement.</p><p><em>Note: Details pertaining to collaborative projects at the Spectrum Lab, IISc, are confidential. Therefore, I do not take any direct mention of the problem statement. The motive for drafting this article is solely knowledge enrichment via experiential learning . The following sections highlight my learnings- I hope my experiences enlighten you! </em>🌟</p><h3>Appreciating Diverse Perspectives</h3><blockquote>“The best way to ‘know’ a thing is in the context of another discipline”.- Leonard Bernstein</blockquote><p>A biomedical problem statement can be viewed from a <strong>medical</strong> standpoint and a<strong> technical</strong><em> </em>standpoint. To understand this, let us consider an example of malignant tumour detection. A group of DL enthusiasts could brainstorm ways to optimize object detection algorithms for detecting malign/benign tumours, whilst oncologists use their expertise for diagnosis. Imagine a scenario where both the entities do not work in tandem. <em>It can be likened to designing a new fighter aircraft model without any interaction between pilots (target users), aircraft designers and OEMs of aircraft parts.</em></p><p>What better way to detail multiple outlooks on a topic, than by discussing challenges faced by the parties involved? Within a few visits, I realised that <em>our technical challenges and the physicians’ challenges were as different as chalk and cheese! </em>Fig.2 summarises my observations in respect of a diagnostic software tool.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*IAoDsH59p_j0rxRKoGC1DQ.png" /><figcaption>Fig.2</figcaption></figure><h3>Data Science Cognoscenti vs a Collaborative Team</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/548/0*XfzZO1zuSQM0DE35.jpg" /><figcaption>Fig.3</figcaption></figure><p>At this point, we’re going to consider the following questions from the viewpoint of a data scientist (Ollie).<em> Would Ollie be able to answer the following independently?</em></p><ul><li><strong>What are the different image acquisition devices used for Ollie<em>’s</em> problem?</strong> If the equipments vary i.e.the sensors have different specifics- the<strong> </strong>data distributions will differ. Simply put, if Ollie trains her predictive model with data from distribution-1, and validates the model against data from distribution-2,<em> she’s in for some erratic results</em>.</li><li><strong>Is image enhancement required? </strong>Ollie<strong> </strong>may need to incorporate data cleaning and de-noising techniques based on image clarity.</li><li><strong>How severe is the condition or pathology?</strong> Serious diseases (like cancer) call for a<em> lower margin of error, false positives and false negatives.</em></li><li>Although DL models retrieve high-level and low-level features, <strong>should Ollie consider any additional handcrafted features?</strong></li><li><strong>How does the doctor make the right diagnosis? </strong>Let us consider that<strong> </strong>Ollie<strong> </strong>designs a DL model assuming that oncologists bank on visual perception mainly i.e <em>colour, texture and tumour depth.</em> <em>What if the tumour location and vitals play an important role too?</em></li><li><strong>What are the risks involved and how much of risk is acceptable? </strong>Indisputably, medical risks may have dire consequences<strong>. </strong>Therefore Ollie should enquire about additional symptoms, outliers, performance standards, whether time-critical diagnosis is required etc. <em>What seems like a polypoid could actually be a malignant tumour </em>— imagine getting that wrong!</li><li><strong>How many cases are encountered on an annual basis? </strong>This would give Ollie an estimate of the number of data samples that she can collect. Large numbers can strengthen an argument for fully automated diagnostic tools or other automation platforms.</li><li><strong>What are the durations and costs of procedures?</strong> Ollie<strong> </strong>would want to build a <em>robust, reliable, computationally efficient, affordable and accurate</em> technology solution (hardware/software) that is cheaper than the existing ones in the market.</li><li><strong>What are the challenges faced by the doctors?</strong> Ollie should contemplate if the doctors are <em>actually</em> going to utilize any assistive technology. For instance, <em>a doctor may not need a diagnostic tool to detect skin ulceration</em>, since these are easy to identify.</li><li><strong>Would a real-time solution be preferred?</strong> If an ICU patient has an active bleeding region in his stomach, the doctor is bound to take immediate action like cauterizing affected areas to mitigate bleeding. <em>There’s no way he would use a diagnostic app where he has to manually feed separate images to validate his diagnosis.</em></li><li><strong>How would clinical validations be carried out</strong>? In certain cases, doctors may not be able to validate Ollie<em>’s</em> findings based on visual perception solely. For example, the oncologist may demand a second step of validation (biopsy) carried by a histopathologist- to determine if a tumour is malign.</li><li><strong>Are there any performance standards?</strong> For procedures like Glaucoma, severity screening is carried out in accordance with ICD — 9 and DDLS standards.</li></ul><h3>Data Collection is a Herculean Task For Biomedical Imaging Applications</h3><p>The lack of annotated data can be crippling when it comes to training deep neural networks. One cannot simply scrape a couple of websites, label data manually without an expert, or carry out a survey- to create a rich medical data repository (no offence meant to NLP aficionados). To top it all, medical experts may not be able to mark the affected regions accurately, at times.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/433/1*DX88tDh2ZrgOMdMNOkvTFg.png" /><figcaption>Fig 4. (a) Histopathology data for breast cancer detection (c) annotation sample (marked by a doc) [1]</figcaption></figure><p>What happens if you try annotating a few samples yourself? I embarked on a data annotation spree, keeping in mind what the doctors had taught me. To my dismay, the <strong>boundaries of the lesions</strong> were ambiguous and there were <strong>multiple variants</strong> of the same abnormality. Furthermore, I misidentified the normal region (large glaring cherry red scary striations)and abnormal regions (a minuscule mild purplish mass at the periphery). 😝</p><blockquote><strong><em>Food for Thought: </em></strong>How would a model recognise an abnormality if it had multiple variants like clusters, patchy, granular and viscous masses? What if the model starts learning the larger red regions instead of the purple one?</blockquote><p>When my ninja annotation skills turned against me, I resorted to weakly semi-supervised learning, and artificial data synthesis alternatives using GANs/ texture transfer. I obtained a few artificial samples rather optimistically, only to realise that the<em> target distribution learnt by my model did not yield </em><strong><em>visually realistic images.</em></strong></p><blockquote>“How can blood be transparent? We should not be able to see the tissue beneath the viscous mass. ”- Doctors</blockquote><p>Subsequently, I also learned that the <strong>background </strong>(normal region) of an image mattered just as much as the foreground (abnormal region), since some lesions are <strong>location-specific</strong>. This was quite the bummer.<em> Sometimes it’s best to bank on medical experts</em>- even if it takes hours of field work- to obtain reliable manual annotations.</p><h3>Selling an Idea to a Doctor</h3><p>To begin with, <em>one cannot get away frivolously by giving a weak proof of concept, purpose or rationale</em> to the doctor (just to draw some research publications out of the project). Experienced specialists can see through your solution before you know it. Be specific about the outcomes, <strong>detailing how they would benefit both, doctors and patients</strong>. At the same time, <strong>avoid complicated algorithm descriptions and technical jargons</strong>- these will only vex physicians. Here are some hacks that I picked up, along the way.</p><h4>Simplify the Concept</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FopqIa5Jiwuw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DopqIa5Jiwuw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FopqIa5Jiwuw%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/146c72fd02f6329fc26330354561df63/href">https://medium.com/media/146c72fd02f6329fc26330354561df63/href</a></iframe><p>First off, picture explaining the application, novelty, scope and performance of the proposed solution to a newbie. The key is to start with a high level overview/ <strong>‘the big picture’</strong> and slowly dive into the low-level specifics (only if the doctor is inquisitive and asks you to elaborate on a particular topic). While compiling presentation content, <strong>put yourself in the doctors’ shoes</strong>. Subsequently,<strong> ask any of your medical friends what they would find relevant</strong>, should you present your problem statement to them.</p><h4>Leverage Visual Learning Tools</h4><p>Remember how your high school Biology prof would use animated videos to delineate different systems and bodily functions? Ever used Khan Academy or TED-Ed videos to satisfy your curiosity? These platforms exemplify the power of visual learning.</p><p><strong>Using animated videos, infographics or concept maps can help explain complex pipelines and workflows.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FnmcDnEhZTJM%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnmcDnEhZTJM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnmcDnEhZTJM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/78a70e089c1ae6310b44cedb17e7c4ac/href">https://medium.com/media/78a70e089c1ae6310b44cedb17e7c4ac/href</a></iframe><p>Take a few minutes to watch the video above. <em>Ask yourself how you would explain this to a technical and medical tyro,</em><strong><em> within 2 minutes</em></strong><em>.</em> Note that your strategy should be one where the person can remember the project for a few days at the least. After this, need I justify why one should be creating a healthcare-product prototype video?</p><h4>Doctors Care about Performance Metrics</h4><blockquote>“Which performance metric did you use and why?”</blockquote><p>Just telling the doctor that I was using precision/recall/F-Score/ AUROC- because relevant academic literature mentioned those- was clearly not enough. My state-of-the-art results barely counted if I could not provide a strong case for choosing certain metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/330/1*QpIZ5ekWBLn0BrD8qqckrg.png" /><figcaption>Fig.5 : (a) IDRiD dataset annotation (ground truth) (b) predicted segmentation map</figcaption></figure><p>A simple example can underscore the need for correct performance parameters. Fig.5(A) represents an annotation of micro-aneurysms in a fundus image. If a segmentation model wrongly predicts a segmentation map 5(B), very high accuracy values (&gt;98+% ) will be observed. Unfortunately, <em>not a single micro-aneurysm is detected. Here’s where s</em><strong><em>ensitivity</em></strong><em> comes to the rescue!</em></p><h4>The Big Question</h4><p><em>Very few doctors will entertain the idea of having technology that replaces their expertise and professional duties</em>. If you go to a doctor stating that your software will interpret ultrasound images better than the doctor himself, there’s a high likelihood that you will receive condescending remarks throughout the session.</p><blockquote>“What can this tool do, that I or my existing equipment cannot do efficiently?”- Doctors</blockquote><p>This question should be of paramount importance if one wishes to work on a healthcare application. For any collaborative venture, doctors would ideally look for<strong><em> a mutually- beneficial, impactful solution which is either novel or substantially better than the existing ones</em></strong><em> (for all you know, they too follow the 10X Rule unknowingly). </em>Either ways, <strong><em>one needs to make the doctor feel involved and her expertise/experience seem valued at all times</em></strong>.<em> Mentioning something as simple as, “assistive technology that reduces diagnosis time” would be a far better description than stating, “this tool can reduce the need for medical specialists”.</em></p><p>For starters, make a list of how your hardware/software would benefit the local medical community. Your solution may focus on:</p><ol><li>Designing better image acquisition devices.</li><li>Causing less patient discomfort.</li><li>Reducing expert review time.</li><li>Re-designing the hardware to make affordable equipment.</li><li>Providing assistive technology to reduce diagnostic errors.</li><li>Developing prescreening tools for a particular pathology.</li><li>Providing simulation/ data visualisation tools.</li><li>Designing a multi-purpose patient app.</li></ol><p>(By no means is this list exhaustive).</p><p>It’s worthwhile to spend some time reading about successful medical and technical collaborative projects, to understand the factors that made them stand out for funding (99Dots is a good place to start).</p><p><a href="https://www.99dots.org/Home/Index#idea">99DOTS</a></p><h4>Draw Smart Parallels</h4><p>It’s natural to observe a gap between your healthcare app functionalities and the doctor’s perception of the app. A physician <em>may take time to visualise the app design and usage,</em> unless you show him a <strong>prototype</strong>. To bridge this gap, I resorted to UI mockups/prototyping tools to get my ideas across.</p><p>When the doctors still had ambiguities in what the end product would look like, I realised that a <strong>relatable analogy</strong> would result in a singular line of thought.</p><p>Doctors can relate to medical terminology; <strong>find an existing medical procedure/tool that conveys your idea.</strong> For example, <em>we found similarities between our application and an ECG report. The latter consisted of patient demographics, timestamps and predictions- all of which were some of the fundamental features of our application.</em></p><h4>What Could Go Wrong</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-eUoEflNLCjvrL9j.jpg" /></figure><p><strong><em>Be transparent</em></strong><em> while explaining any </em><strong><em>limitations</em></strong><em> of the proposed solution</em> so that the doctor is aware of the risks involved. What could be an error due to model under-fitting, could result in serious consequences (due to misdiagnosis).</p><blockquote>Only the Doctor should decide if the risks are worth taking.</blockquote><p><strong>Clarify the gold standard and acceptable margin for errors</strong> while targeting any medical problem. Remember that surpassing the state of the art in some AI conference does not guarantee that your solution will be accepted by the medical community.</p><h4>Describe Your Sample Dataset (Even if it is Publicly Available)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*f17TNPSGAkxxy-GT.gif" /></figure><p>One of the perks of working with senior doctors is that you learn to face unexpected queries. During a visit, <em>I was asked to describe patient demographics</em>, based on the publicly available dataset repository used.</p><blockquote>“What are the number of patients, were multiple samples drawn from them?”</blockquote><blockquote>“What are the acquisition device specifications and the ethnicity of the people?”</blockquote><p><em>Little did I know that the doctors were trying to envision how our solution would help the </em><strong><em>local masses</em></strong><em>.</em></p><p>When I enquired about the necessity of these details, the doctors explained how 4<em>0% of the pathologies in my dataset were extremely rare in India and common in the western countries.</em> Therefore,<strong> designing a predictive model for such rare abnormalities would not be of much use, in the Indian context. </strong>Secondly, the image acquisition devices used in the western countries were far more advanced and expensive when compared to the ones used in local hospitals; <strong>training our DL models on high resolution images from such devices, may not help the medical community in India.</strong></p><h4><strong>Prospects for Medical Research</strong></h4><p>I must say that I have been extremely lucky to work with such senior doctors in one of the largest government hospitals in my country. Since we had to propose our solution to external government agencies, they taught me an important lesson first-hand.</p><blockquote>“Your (tech) team has to offer some carrots to convince higher authorities for external collaboration”.</blockquote><p>Simply put, it is essential to highlight all the benefits offered to the other party in a collaborative setting. Justify what their community would gain by making time for your project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3mLtqEexHlhAu3g4.jpg" /></figure><p>In this context, <strong>we elaborated on how our research would help the doctors’ research. </strong>Just like scientists, doctors publish their work to renown medical conferences and journals. While AI enthusiasts publish their research on fully automated solutions, their medical partners may simultaneously write about discovery of a protein, findings of a case-study or a first-of-its-kind procedure! <em>Doctors would certainly like to explore an area which is </em><strong><em>intellectually stimulating and has a good scope for research</em></strong><em>.</em></p><p>Give the doctors a reason to spend valuable time from their jam-packed schedule (comprising continuous shifts and sleepless nights interspersed with unanticipated cases) for the project.</p><p><em>Note: I’m grateful to my professor, for guiding me through my project. Part 2 will be out, soon! Hope ya’ll liked this. Feel free to comment. :D</em></p><p><em>[1] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation.” International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c4b559c550a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Whimsy Bookworm Musings]]></title>
            <link>https://medium.com/series/whimsy-bookworm-musings-a68c835a9263?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/a68c835a9263</guid>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Wed, 26 Dec 2018 23:08:00 GMT</pubDate>
            <atom:updated>2019-01-18T08:01:42.091Z</atom:updated>
            <content:encoded><![CDATA[<h3>A List of Books That Made Me Go Wow</h3><h4>(Recommended Reads)</h4><h3>Introduction</h3><h4>Early Days</h4><p>My journey with novels and other story books started right from pre-school. For some reason, my relatives and friends knew that I’d turn out to be a voracious reader (or maybe they wanted to inculcate a reading habit). Either ways, my birthday gifts comprised books like Arabian Nights, Folk Tales and a LOT of Enid Blyton. As I stepped into primary school, my mother made it a point to stock my bedroom with series of <em>Famous Five, Secret Seven, Nancy Drew, Hardy Boys, Agatha Christie and P.G. Wodehouse.</em> By now, I’m sure that some of your pasts resonate with mine.</p><h4>Classics Never Grow Old</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/661/1*bVJQPir9YmDSABR48JRXTQ.png" /></figure><p>When you’re brought up in a family of Bengali bibliophagists, there’s hardly anything that you can do, to escape the stream of books that flood your room. My current family tree starts with my great grandma, who bagged a scholarship for English Language. My grandmother, a college-level English professor, played a crucial role in introducing me to several classics. Although I found my first read (David Copperfield) slightly monotonous, I was hooked on to subsequent classics. <em>The Tale of Two Cities, Little Women, Treasure Island, Adventures of Huckleberry Finn, Robinson Crusoe, Animal Farm, Matilda, The Adventures of Tom Sawyer, Kidnapped, Alice in Wonderland, Heidi, Black Beauty, The Adventures of Sherlock Holmes, Jane Eyre, The Prince and the Pauper, The Secret Garden, Pygmalion, Pride and Prejudice, King Lear and Oliver Twist</em> -to name a few. As I grew up (the eager beaver that I am), I explored books like <em>The Godfather, Lolita, Macbeth and Gone with the Wind</em>.</p><h4>The Harry Potter Fandom</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/347/1*DyPtD3Ld4fD53M20mf_4FA.jpeg" /></figure><p>The Harry Potter craze hit me during my 5th Grade (like it did, to many other middle and high school students). I must confess that I was so obsessed with the series, that I used to visit a Crossword Book store after class, and sit for hours together- resolute to finish the first 6 books that were available, devouring every bit of the unexpected that J.K Rowling had intricately weaved in her spell-binding novels. Unfortunately, my mom caught me red-handed and fortunately, she enrolled me at a neighbourhood library. Towards the end of my 7th grade, I had topped Math in my batch; I used my cash prize to buy <em>The Deathly Hallows</em>- one of the best gifts that I could give myself. I could go on about the Potter series, but let’s keep that for another day :)</p><h4>Transitioning</h4><p>My High School days were about<em> Robin Cook, David Baldacci, Jeffrey Archer, John Grisham, Lee Child, Christopher Paolini, C. S. Lewis and Dan Brown </em>novels. Alongside, I was glued to the <em>‘A Song of Ice and Fire’ s</em>eries, thanks to a close friend at secondary school (Pinko ❤).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*R07OfOzTb9NAjlDoG2Lbng.jpeg" /></figure><p>Over the years, I’ve come to realize the importance of <strong>switching genres</strong> (a mantra to creative thinking and enhancing one’s general knowledge). Hence, you’ll find my choice of books fairly random. <strong>After all, if each book is a journey in itself, wouldn’t you like to travel across different worlds?</strong></p><h4>About the Series</h4><p>You will find a list of reads that I deem impactful and compelling. Note that this is subjective, since I’ve compiled the list based on my personal reading escapades. There maybe a LOT many others; please feel free to write to me in case you have any suggestions. Since a lot of my fellow bookworms have influenced my reading choices, I’ve included credits too. At the moment, I have made an attempt at penning my reviews for two novels. In the near future, I certainly hope to review the rest elaborately :)</p><h4>The Concept</h4><p>You must be wondering why I chose to share my book list. For starters, take a second to contemplate the number of hours you spend in reading a book to completion. And, when I say reading, this does not imply browsing or cursory glances. In this regard, I believe<em> </em><strong><em>each book is a valuable investment of time</em>. </strong>Thence, <strong>one must choose novels judiciously.</strong></p><p>From my experience, <strong>personal recommended reads </strong>have worked out in my favour, when compared to the generic Google guru. For instance, I’d choose books from Gates Notes, Oprah’s Book Club Lists and my aunt’s recommended reads over a thousand random reviews. Therefore, I thought of compiling my list of fairly good reads. I really hope that you gain inspiration from this and share your list with me too!</p><p><em>Note: Look out for </em>⭐<em>.</em></p><h3>Oprah’s Book Club List</h3><p>By far, this has been one of the best compilation of suggested reads. Most of the novels are etched in my hippocampus. The following are must-reads!</p><h4>The Sun Does Shine: How I Found Life and Freedom on Death Row ⭐</h4><p>Author: Anthony Ray Hinton<br>Pages: 255<br>Release Date: March 2018</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*mv8198CgMiHyEvyXl8XgTw.jpeg" /></figure><p>I began my 2019 by experiencing feelings of hope, sympathy, admiration, grief, utter disappointment, joy and relief in 48 hours. All thanks to Mr. Hinton! From his his darkest times during 30 years on death row, to times where he found hope and freedom using his imagination- this simple narrative highlights a true story with underlying themes of humanity, optimism, benevolence and friendship.</p><h4>The Invention of Wings ⭐</h4><p>Author: Sue Monk Kidd<br>Pages: 384<br>Release Date: January 2014</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/772/1*E73w87WFtdmkhHRTIH_-LA.jpeg" /><figcaption>Credits: My Cousin Brother (Rajat)</figcaption></figure><h4>The Goldfinch ⭐</h4><p>Author : Donna Tartt<br>Pages: 784<br>Release Date: September 2013</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F8O1FUPXPlZeNA4tRYElEQ.jpeg" /></figure><h4>A Fine Balance</h4><p>Author : Rohinton Mistry<br>Pages: 624<br>Release Date: 1995</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/326/1*y9aIZTo8VY_xNOKdIUncPA.jpeg" /></figure><h4><strong>The Secret Life of Bees</strong></h4><p>Author : Sue Monk Kidd<br>Pages: 336<br>Release Date: November 2001</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*JMY8NZMcVJenixE07mjbvA.jpeg" /></figure><h3>Books By Indian Authors</h3><p>My experiences with Indian novels have been varied. From surges of emotions to spells of monotony- I’ve encountered books covering a large spectrum of moods.</p><h4>The Shiva Trilogy</h4><p>Author: Amish Tripathi<br>Books: (The first one was my favourite)</p><ul><li>The Immortals of Meluha</li><li>The Secret of the Nagas</li><li>The Oath of the Vayuputras</li></ul><p>Release Dates: 2010–2013</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ISFNfBTjuiyrRm8nvDNBaw.jpeg" /></figure><h4>The God of Small Things ⭐</h4><p>Author: Arundhati Roy<br>Pages:321<br>Release Date: April 1997</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/339/1*NnrNM9GQdNKC-OdmiTQDbg.jpeg" /><figcaption><em>Credits: My aunt’s book collection ❤</em></figcaption></figure><h4>I Too Had a Love Story</h4><p>Author: Ravinder Singh<br>Pages: 199<br>Release Date: 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*volvLcc5fJfRvcVikOPU-g.jpeg" /><figcaption>Credits: Pranshu, a senior from college :)</figcaption></figure><p>Although this is a light read, the book encompasses both, euphoric and heart-wrenching moments. I usually don’t venture into romance novels, but this one is worth a read.</p><h4>Books by Chetan Bhagat</h4><p>I am not a huge fan of the grammatical constructs in his novels. However, the simplicity of middle-class India reverberates in his works, thereby engaging the readers. Out of the lot, I liked two novels.</p><p>Books: Five Point Someone, The 3 Mistakes of My Life<br>Release Dates: 2004, 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*wSUV1m8IufKx6U5niEdVJQ.jpeg" /><figcaption>Credits: From daddy’s collection ❤</figcaption></figure><h4>The Krishna Key</h4><p>Author: Ashwin Sanghi<br>Pages: 475<br>Release Date: 2012</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/327/1*XEWOECTOtzw2hgWKG7Ggeg.jpeg" /><figcaption>Credits: Gifted by a close friend (Mayank) :)</figcaption></figure><h4>Wings of Fire</h4><p>An autobiography of ‎A P J Abdul Kalam, a former President of India‎.<br>Pages: 180<br>Release Date: 1999</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/319/1*UtDV2GFjsILqTvVI5JL1_g.png" /><figcaption>Credits: From daddy’s collection ❤</figcaption></figure><p>Although I haven’t read any novel by Jhumpa Lahiri, some of her works like <em>The Lowland</em> are renown.</p><h3>Teacher’s Corner</h3><p>Here’s a list of books (recommended by my school teachers) which had a profound influence on me. If you’ve spent 14 years of your life in a Convent School, you will have invariably come across some of the finest English teachers and sisters.</p><h4>Tuesdays with Morrie</h4><p>Author: Mitch Albom<br>Pages: 192<br>Release Date: 1997</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WCKZDt3CwkfU65cyG9PDvA.jpeg" /><figcaption>Credits: Mrs. Dilruba, my ICSE English teacher ❤</figcaption></figure><h4>The Eagle Has Landed</h4><p>Author: Jack Higgins<br>Pages: 352<br>Release Date:1975</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/313/1*RKBKNxctFJ5_n76nCfevxQ.jpeg" /><figcaption>Credits: Mrs. Jayshree Jacobs, ISC English Prof. ❤</figcaption></figure><h4>The Emperor of All Maladies ⭐</h4><p>Author: Siddhartha Mukherjee<br>Pages: 592<br>Release Date: November 2010</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/305/1*DFjw40Mcpg5E_dZLL5wlxA.jpeg" /><figcaption>Credits: My Uncle &amp; Mrs. Veera Menon (ISC Chemistry Prof)❤</figcaption></figure><p>Note that this book makes use of gruesome imagery and personification. Almost like a macabre. In short, this book is not for the faint-hearted.</p><h4>Gem in the Lotus: The Seeding of Indian Civilisation</h4><p>Author: Abraham Eraly<br>Pages: 496<br>Release Date: 2000</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/475/1*sJ9HZY_RgnfB7nEtJipmsQ.jpeg" /><figcaption>Credits: Mrs. Pratima Rao (ICSE Math and CS prof)</figcaption></figure><h4>The Final Diagnosis</h4><p>Author: <a href="https://www.goodreads.com/author/show/71955.Arthur_Hailey">Arthur Hailey</a><br>Pages: 333<br>Release Date: 1959</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/308/1*rBmUIr82nqZ6U9SPbrak2A.jpeg" /><figcaption>Credits: Mrs. Sheela Rani (ICSE Bio prof.)</figcaption></figure><h4>The Five People You Meet in Heaven</h4><p>Author: Mitch Albom<br>Pages: 194<br>Release Date: September 2003</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/529/1*bi438cOh12e1i4TU_Ez4DA.jpeg" /><figcaption>Credits: Mrs. Dilruba, my ICSE English teacher ❤</figcaption></figure><h3>Self-help</h3><p>Mental health and attitude are important aspects of an individual. Such evergreen books certainly help over-thinkers like me, in dealing with real-life situations.</p><h4>Deep Work ⭐</h4><p>Author: Cal Newport<br>Pages: 265<br>Release Date: Jan 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/324/1*wbXlI9uDF-CkT7fwOX-2Fg.jpeg" /><figcaption><a href="https://medium.com/u/cfd46a6e3b12">Swatiii</a>’s recommendation ❤</figcaption></figure><h4>Tough Times Never Last, but Tough People Do!</h4><p>Author: Dr. Robert H. Schuller<br>Pages: 240<br>Release Date: 1983</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/296/1*VJ-12RFHOBXcgHHqwU9nNQ.jpeg" /><figcaption>Credits: Thanks Mom ❤</figcaption></figure><h3>The Power of Positive Thinking</h3><p>Author: Norman Vincent Peale<br>Pages:128<br>Release Date: 1952</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M8EysXRuHiWTVWLX4giTHA.jpeg" /><figcaption>Credits: Thank you Sukul ❤</figcaption></figure><h4>The Monk Who Sold His Ferrari ⭐</h4><p>Author: Robin Sharma<br>Pages: 198<br>Release Date: 1996</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vYD_Su0NIejlf2fgDKV8Dg.jpeg" /><figcaption><em>Credits: My aunt’s book collection ❤</em></figcaption></figure><h4>The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life</h4><p>Author: Mark Manson<br>Pages: 224<br>Release Date: September 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lNs-D0ZvA2OH3Rt83cL0yg.jpeg" /><figcaption>Credits: A gift from Sukul ❤</figcaption></figure><h4>Outliers ⭐</h4><p>Author: Malcolm Gladwell<br>Pages: 304<br>Release Date: November 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/325/1*SPorLXoPNAgA1youb_2kJg.jpeg" /><figcaption>Credits: A birthday gift from a close friend, Mayank :)</figcaption></figure><h4>The Law of Success</h4><p>Author: Napoleon Hill<br>Pages: 502<br>Release Date: 1925</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uK9gWC1ZgKukcAv104ojTg.jpeg" /><figcaption>Credits: Mom’s recommended reads ❤</figcaption></figure><h4>StrengthsFinder 2.0</h4><p>Author: Tom Rath <br>Pages: 192<br>Date Released: 2007</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gH8fJFtHCAGKVXW742d7JQ.jpeg" /><figcaption>Credits: From daddy’s collection ❤</figcaption></figure><h4>The Secret</h4><p>Author: Rhonda Byrne<br>Pages: 198<br>Date Released: November 2006</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*4HNTc10tMrIvpLbSE2gHnQ.jpeg" /></figure><h3>Biography &amp; Motivation</h3><p>From stories of escape and survival to those of unique and paranormal events.</p><h4>Life of Pi</h4><p>Author: Yann Martel<br>Pages: 354<br>Date Released: September 2001</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/274/1*w12dW-FxCy9sEELtF0qqig.jpeg" /></figure><p>If you like the usage of imagery and metaphors, this book is for you!</p><h4>Papillon</h4><p>Author: Henri Charrière<br>Pages: 516<br>Date Released: 1969</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/180/1*HI1VGmByCg7lW7dHx0iY0w.jpeg" /><figcaption>Credits: My Cousin Brother (Rajat) ❤</figcaption></figure><h4>The Story of a Shipwrecked Sailor</h4><p>Author: Gabriel García Márquez<br>Pages: 135<br>Date Released: 1970</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*N1LEq-z3aKbjco5x2LVRZQ.jpeg" /></figure><h4>Many Lives, Many Masters</h4><p>Author: Brian Weiss<br>Pages: 217<br>Date Released: July 1988</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*zUf-36JfBG74B2V5xbPbAA.jpeg" /><figcaption>Credits: My Uncle’s book list ❤</figcaption></figure><p>Note that this one is a little puzzling and supports notions pertaining to karma, rebirth and evolution of one’s soul.</p><h4>Shantaram</h4><p>Author: Gregory David Roberts<br>Pages: 936<br>Date Released: 2003</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/326/1*Y7K9qYTFhmemJXma5Y5OXw.jpeg" /><figcaption><em>Credits: My aunt’s book collection ❤</em></figcaption></figure><p>Regrettably, this is one of the very few books that I could not finish till the last page. The content was too intense, back in secondary school. Nevertheless, I enjoyed reading most of it. If you prefer light reading, this ain’t the book for you.</p><h4>Steve Jobs</h4><p>Author: Walter Isaacson<br>Pages: 656<br>Date Released: October 2011</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/319/1*95lUNS9vVYbl_EbGIA6ufw.jpeg" /></figure><h4>Shoe Dog: A Memoir by the Creator of Nike</h4><p>Author: Phil Knight<br>Pages: 386<br>Date Released: April 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/270/1*jFa4rSe53ZIIPsx0CST5pQ.jpeg" /></figure><h4>The Last Lecture</h4><p>Author: Jeffrey Zaslow, Randy Pausch<br>Pages: 224<br>Date Released: April 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/326/1*I8Vq2zZunzVrtOpwdrrjiw.jpeg" /></figure><p>Despite having watched Randy Pausch’s original lecture before reading the book, I wouldn’t mind doing the latter a hundred times over!</p><h4>Kafka on the Shore ⭐</h4><p>Author: Haruki Murakami<br>Pages: 615<br>Date Released: September 2012</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/258/1*m-x2Lqf7Jwz0Rq1Y8FwqwA.jpeg" /><figcaption>Credits: Recommended by a close friend, Suvo</figcaption></figure><p>Remember that feeling of bewilderment after watching Inception? Ring a bell? This book evokes a similar feeling; there’s no ‘correct’ ending per se. The complexity of the plot, personifications and juxtapositions make this a worthwhile read. Not to forget the quotes- some of them are important life lessons :’)</p><h4>Lean In: Women, Work, and the Will to Lead</h4><p>Authors: Sheryl Sandberg, Nell Scovell<br>Pages: 387<br>Date Released: March 2013</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/303/1*QxHPacESC_3SWkxXLA7-hQ.jpeg" /><figcaption>Credits: Lakshya ❤</figcaption></figure><p>With over 4.2 million copies sold worldwide, do I really have to say much?</p><h3>Fiction</h3><p>You must be wondering why this section is called ‘fiction’ although some of the books falling under this genre are mentioned elsewhere. To be honest, this is an observable commonality among the remaining books that I wish to suggest. Pardon me for the ambiguity.</p><h4>A Case of Exploding Mangoes</h4><p>Author: Mohammed Hanif<br>Pages: 336<br>Date Released: 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/325/1*I424iMb5ATmuDo-SR17Efw.jpeg" /><figcaption><em>Credits: My aunt’s book collection ❤</em></figcaption></figure><h4>The Kite Runner ⭐</h4><p>Author: Khaled Hosseini<br>Pages: 372<br>Date Released: May 2003</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/325/1*5oBGsPBhwRtJw6l1W1Z8JQ.jpeg" /><figcaption>Credits: My Roomie, Abby ❤</figcaption></figure><h4>Local Girl Missing</h4><p>Author: Claire Douglas<br>Pages: 368<br>Date Released: 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nMNfY8aodeZ51dScwfYNzQ.jpeg" /></figure><p>Note that the ending may send chills down your spine. This book is full of surprises!</p><h4>Property of a Noblewoman</h4><p>Author: Danielle Steel<br>Pages: 336<br>Date Released: February 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/327/1*lto0wdGFN7rHvqE6gNeazQ.jpeg" /></figure><h4>The Spy</h4><p>Author: Paulo Coelho<br>Pages: 186<br>Date Released: September 2016</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*peeD5FPdYskjwwoaJuG8Sg.jpeg" /></figure><h4>The Girl with the Dragon Tattoo</h4><p>Author: Stieg Larsson<br>Pages: 465<br>Date Released: August 2005</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/322/1*1e-Kcgt5wPzH8gZxFEmE_A.jpeg" /></figure><h4>Memoirs of a Geisha</h4><p>Author: Arthur Golden<br>Pages: 448<br>Date Released: September 1997</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/360/1*9od0-m9CA-CQDb4txJvj8Q.jpeg" /><figcaption>Credits: My bench-partner at high school, Ayusshi ❤</figcaption></figure><h3>Authors That Inspire</h3><p>Now that you’ve read the heading and sub-headings, you definitely know that these authors require a separate section. What best way than to categorise a medley of books based on their authors?</p><h4>John Grisham</h4><p>From the books that I have read, I vividly remember <em>The Racketeer, Camino Island, A Time to Kill and Bleachers.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tr6WgI-oMdREhIcCqsPW8Q.jpeg" /></figure><h4>Dan Brown</h4><p>Ironically, my grandpa introduced me to <em>The Da Vinci Code </em>before<em> Angels &amp; Deamons.</em> After a point, the story lines of the different novels become predictable. Nevertheless, I’d vouch for the books over the movies, any day. Rich in historical and biblical references, the Robert Langdon series could be one of the most informative thrillers that you’d find.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*q4iAATwf18OPf-7gQVxONQ.jpeg" /></figure><p>I have read <em>The Da Vinci Code, Angels &amp; Deamons, The Lost Symbol, Deception Point, Digital Fortress and Inferno. </em>Out of the lot, <em>The Da Vinci Code</em> continues to remain my favourite.</p><h4>Jeffrey Archer</h4><p>To begin with- I am a huge fan of Jeffrey Archer. Back in high school, my aunt introduced me to Kane and Abel. Since then, I’ve been lucky enough to ransack her bookshelf and catch hold of novels like<em> A Prisoner of Birth, The Prodigal Daughter, Only Time Will Tell and Best Kept Secret.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/665/1*eD2hy5lafAsN-l1a-FItvA.jpeg" /></figure><h4>Robin Cook</h4><p>All credits to my mother, for gifting me my first set of Robin Cook books. In my opinion, <em>Coma </em>is his best novel. I’d put <em>Blindsight, Outbreak, Mortal Fear, Brain</em> and <em>Contagio</em>n as some of my best reads too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/298/1*ePatcilv3XEN8GIOs7yoBA.jpeg" /></figure><p>I didn’t quite enjoy Biology classes in secondary school. Looking back, mumma reveals that this was a tactic to get me interested in biosciences :P Today, most of my work deals with deep learning and computer vision for biomedical imaging applications. You never know where life takes you! (Never dislike a subject).</p><h3>Others</h3><p>There are books which top the ‘best-sellers list’ and then you have what I call the <em>hidden oyster pearls</em> ( jewellery-grade pearls, of course). These are gripping books which aren’t as popular as other works of their authors.</p><h4>The Quantum and the Lotus</h4><p>Author: Matthieu Ricard<br>Pages: 320<br>Date Released: 2000</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/204/1*jMVckB5WUCEDJA7FSKIbjg.jpeg" /></figure><p>This books follows an interview format. If you’re not accustomed to reading dialogues in plays, you may not appreciate the construct.</p><h4>61 Hours</h4><p>Author: Lee Child<br>Pages: 448<br>Date Released: March 2010</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*J8A0nVawhGf3oyQ3Ur7sLw.jpeg" /></figure><h4>The Winner Stands Alone</h4><p>Author: Paulo Coelho<br>Pages: 375<br>Date Released: 2008</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/309/1*kJ-f13JTCfcXEvBBF8aHIQ.jpeg" /><figcaption>Credits: My Uncle’s book list ❤</figcaption></figure><h4>The Importance of Being Earnest</h4><p>Author: Oscar Wilde<br>Pages: 80<br>Date Released: 1985</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/179/1*luW12W-UzUGoCBlO47NpIA.jpeg" /><figcaption>Credits: Gifted by a close friend (Mayank)</figcaption></figure><p>This is an ideal recipe for a rotten mood. Definitely a light read, with multiple twists and turns.</p><h4>Compromised Data: From Social Media to Big Data</h4><p>Publisher: Bloomsbury Publishing<br>Pages: 257<br>Date Released: July 2015</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/420/1*MRBYqE0iWabWlTtw2uxABQ.jpeg" /></figure><p>The case studies are an eye-opener!</p><p><em>By no means is this list exhaustive. Stay tuned for more :)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a68c835a9263" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fast and Scalable Estimation of Uncertainty using Bayesian Deep Learning]]></title>
            <link>https://medium.com/lean-in-women-in-tech-india/fast-and-scalable-estimation-of-uncertainty-using-bayesian-deep-learning-e312571042bb?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/e312571042bb</guid>
            <category><![CDATA[uncertainty]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[bayesian-machine-learning]]></category>
            <category><![CDATA[seminar]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Sat, 03 Nov 2018 12:53:04 GMT</pubDate>
            <atom:updated>2018-12-29T04:30:56.577Z</atom:updated>
            <content:encoded><![CDATA[<h4>An IISc Talk by Dr. Emtiyaz Khan</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/954/1*5qCpNUKy5HxY1IkM47CPWA.png" /><figcaption>Fig.1 Credits: Slides provided by Dr. Emtiyaz[2]</figcaption></figure><p>Recently, I attended an hour-long presentation by Dr. Emtiyaz Khan, the team lead of the Approximate Bayesian Inference (ABI) Team at RIKEN Center for Advanced Intelligence Project (Tokyo). His talk revolved around<em> </em><strong><em>learning variance by natural gradients. </em></strong>Systematically, he explained why it is challenging to compute the uncertainty and how they took inspiration from the Adam optimizer, in their <a href="https://arxiv.org/pdf/1806.04854.pdf">latest publication at ICML’18 </a>(which beats the state of the art). Overall, he walked us through variational inference, Bayesian models, natural gradients and fast Gaussian approximation for deep learning models, cogently.</p><blockquote>“My main goal is to understand the principles of learning from data and use them to develop algorithms that can learn like living beings.”- Dr.Khan</blockquote><h3>Uncertainty</h3><blockquote>“Say you have to detect animals like cats and dogs. It is very common because there are a lot of images available. Assume that you are asked to detect a cow from rural India- the number of images are far lesser. Hence, the <strong>confidence</strong> in predicting cats and dogs is more than that of cows.”- Dr.Khan</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/663/1*Y67Iyj2DmABaPzSfK5cInA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/597/1*ywJz1sMtTnjS4xfhq5bGgg.png" /><figcaption>Fig.2 (a) and (b) Credits: Slides provided by Dr. Emtiyaz[2]</figcaption></figure><p><strong>Uncertainty helps in estimating the confidence in the predictions of a deep learning system.</strong> To explain variance and why it’s important to know what we don’t know, he gave us a simple example of Japan earthquake data. The green squares in Fig.2(a) represent data points. After giving us a brief introduction to the problem statement, he asked the attendees whether the blue dashed line or the continuous red line would be a better fit. While a majority voted for the red line, his response took us by surprise- both the lines were correct. In such a scenario, how does one decide the confidence of the prediction? The observation was the variance (green region) increases with an increase in the magnitude. It implies that earthquakes with lower magnitudes can predict with higher confidence and that uncertainty rises with size.</p><p>To visualize uncertainty, he quoted an example of image segmentation. Fig.1 represents the ground truth, prediction and risks of several cityscapes. The uncertainty map plots the variance in segmentation. A cursory glance at the results will show that lighter colored regions or uncertain regions in the uncertainty maps correspond to the areas which have not adequately segmented in the predicted maps. We can obtain more robust predictions if uncertainty is on top of a deep learning system. About the current example, imagine the risks involved if driverless car ventures into footpaths (areas of low confidence) because the prediction maps are not segmented properly!</p><ul><li>Computing the uncertainty helps if the data is unreliable, scarce or missing.</li><li>It is difficult to compute the variance or confidence for massive data and large models.</li><li>Using the concepts of Bayesian statistics, optimization and information geometry, they have worked on a fast computation of uncertainty.</li></ul><h3>Why is it Difficulty to Compute Uncertainty?</h3><p>If you’re a deep learning enthusiast, you’re no stranger to the Naive approach in eqn.1. <em>θ </em>is generated from a prior distribution <em>p(θ)</em>(a Gaussian prior in this case). D refers to the data set and <em>θ </em>represents the parameters. Also, <em>x_i </em>and <em>y_i</em> are the input and output terms respectively. The function <em>f </em>is a neural network.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*FraYI6rNSvsGvuL0KNKKew.png" /></figure><p>Fig. 2(b) shows the variance of the draws from a known distribution. We randomly draw out f_θ(x_i), compute ‘p’ and rank the data subsequently. <strong>The grey lines show that we can have multiple best fits in the adjacent green region.</strong> If there are numerous neural networks are used to average them. <strong>How narrow is the spread? The narrowness of the range is the uncertainty (green region).</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/625/1*iiAUIuv-hwUWA8Y1t2uqoQ.png" /></figure><p>Via Baye’s rule(Eq.2), we obtain the posterior distribution. And, in turn, helps us in finding the mean and the variance of the posterior distribution. However, t<strong>he intractable integral is a normalisation constant which is hard to calculate.</strong> Due to a large number of samples and parameters, the computation as mentioned above is intensive. Additionally, <strong>the problem of integration is much harder than that of optimisation. </strong>Keeping this challenge in mind, his team used natural-gradients for a fast approximation of the integral term.</p><h3>Variational Inference with Gradients</h3><p>Assume a normal distribution <em>N, </em>with a<em> mean (</em>μ)<em> and a variance (</em>σ²). The goal is to find the values of μ and σ² such that <strong>the distance between the posterior distribution and <em>N </em>are minimized </strong>(eqn.3). But here’s the catch. What if μ and σ² are 10⁶ dimensional vectors?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*m15Tvlzs0Yjn-ygw1Bl6SQ.png" /></figure><p>Intuitively, the KL divergence (eqn.4) is a means by which the similarity or match between two distributions can be measured. Using this, we can select a neural network out of many candidate neural networks, based on the ability of the neural network to explain the data well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/639/1*W1sBgEKbJfikGKh3mpDIBw.png" /></figure><p>Now that we’ve defined the KL divergence, what is variational inference? In short, <strong>the variation inference is used to approximate the intractable posterior distribution with tractable variational distribution for the KL divergence. </strong>Traditionally, a gradient descent approach is used to update μ and σ, based on a fixed step size or learning rate(ρ). Eqn.5,6 represent the same.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*Jf7XltkL6-fkrhyvhV2Uow.png" /></figure><h4>Natural Gradient Descent</h4><p>By now, you must be contemplating if they have tweaked the gradient descent significantly. Ironically, a single Fisher Information Matrix term does the trick! Let’s begin with the Fisher Information Matrix given by eqn.7. It can be perceived as the variance of the observed information and is used to compute the covariance matrices associated with maximum-likelihood estimates.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/642/1*l3dP5xhHMx6mDybd7BXSSg.png" /></figure><p>Why do we need the Fisher Information Matrix? You will find your answer in a few lines. Eqn.8 represents the update step of gradient descent, while eqn.9 gives the natural-gradient descent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*JMqANuyQF4lnQ6-J0x18uA.png" /></figure><p>Based on the following relation, we can infer that t<strong>he natural gradients (L.H.S) are an approximation of the gradients in gradient descent (R.H.S)</strong>. Furthermore, Emtiaz et.al. propose that if we find the right parameters for the Fisher Information Matrix, the natural-gradients can be computed quickly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/153/1*kf47CmmxY6i5_eK1WZORRQ.png" /></figure><ul><li><strong>Natural-gradients are defined on a manifold</strong> and not in the Euclidean space.</li><li>Euclidean distance is not an apt measure. Assume two Gaussian distributions with mean values of 0 and 25 respectively. Although a fixed Euclidean distance of 25 exists between them, i.e. between the means, they can have <strong>different variances</strong>. If they have smaller variances, the distributions won’t overlap. However, larger variances may lead to overlapping regions.</li></ul><h4>Variational Inference with Natural Gradient Descent</h4><p>The Natural-Gradient variational inference can be found by the equations given below. Eqn.10,11 show the update steps for the mean and the variance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/502/1*NtXf0FT6R9FV3ckMPM0O8w.png" /></figure><ul><li>The updates can be derived when<em> </em><strong><em>q</em> belongs to an exponential family. </strong>Methods like Kalman filters can be generalized by this.</li><li><strong>The learning rate (β) is scaled by the variance </strong>(in blue). Based on the uncertainty, the system can decide whether to take longer or smaller steps.</li></ul><h3>Fast Computation of Uncertainty</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*H9tU9F6MVz4OHnTqpMpkcQ.png" /><figcaption>Fig.3. Difference between the proposed method and Adam. Credits: Original Paper [1]</figcaption></figure><blockquote>“We approximate by a Gaussian distribution, and find it by “perturbing” the parameters during backpropagation”.- Dr.Khan</blockquote><p>At an abstract level, the quotation as mentioned earlier can be understood by assuming that relevant noise is added before back-propagation. This way, variances about different noises can be obtained. Visually, the ultimate goal is to get the bounds for the spread invariance (green region). This approach is analogous to <strong>Adam optimisation </strong>in the sense that the initial performance (variance spread) is relatively similar in both cases. You’re probably racking your brains to understand the Vadam (Variational Adam)algorithm. Wait for a second; there’s an easy way out! Dr Khan summarised Vadam in 5 simple steps :)</p><h4>Vadam (Variational Adam)</h4><p>At this point, you may want to re-visit the Naive approach and Bayesian inference mentioned earlier. A Gaussian distribution (standard normal distribution) is approximated. We assume the following.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/97/1*KAlNY5GFsjv6pavABa0kpg.png" /></figure><p>After that, the steps in Fig.4 are performed. A variety of neural network models can consider different local minima or global minima. If you’re still asking yourself why Vadam is significant- it helps us in answering what other models we could have used, to get a better fit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*MRZvc95VNEwRwDfmazR1Lg.png" /><figcaption>Fig.4. The 5-step method (Vadam). Credits: Seminar Slides[2]</figcaption></figure><p>Lastly, the gradient step is given by eqn.12, and the modified part is highlighted in blue. See how simple the update is?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/464/1*wgIgCHI6QTbLj8i1xm4UxA.png" /></figure><ul><li><strong>Vadam converges to the same variance as Adam. </strong>The variance is found during the training stage and not after it.</li><li>Once the bounds on the uncertainty are obtained, this can be <strong>transferred </strong>to a problem statement with similar data distribution.</li><li><strong>Vadam avoids using local minima </strong>and reduces over-fitting.</li><li>Replacing gradients with natural-gradients is a<strong> faster </strong>and more<strong> robust </strong>approach.</li></ul><h3>Conclusion</h3><p>The following graph shows that Vadam has more stability while converging. If you pay close attention to the first 2000 iterations, it’s evident that the drop in loss is more gradual in the case of Vadam.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/887/1*SA4FMcoRPQ2UB-t7l9mEtg.png" /><figcaption>Fig.4. Vadam results (iteration vs log loss). Credits: Seminar Slides[2]</figcaption></figure><p><strong><em>References:</em></strong><em><br>[1] </em><a href="https://arxiv.org/pdf/1806.04854.pdf"><em>Fast and Scalable Bayesian Deep Learning by Weight-Perturbation in Adam</em></a><em>, M.E. Khan et.al.,Thirty-fifth International Conference on Machine Learning, 2018.<br>[2]IISc presentation slides on </em><a href="https://emtiyaz.github.io/index.html"><em>Dr. Khan’s website</em></a><em>.<br>[3] Fast yet Simple Natural-Gradient Descent for Variational Inference in Complex Models,M.E. Khan et.al.,ISITA 2018.</em></p><p><em>Note: This summary is based on both, the notes taken by me and the original seminar slides. Neither has the proposed work been done by me nor do the images belong to me. Feel free to point out any corrections in case of any ambiguity :)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e312571042bb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/lean-in-women-in-tech-india/fast-and-scalable-estimation-of-uncertainty-using-bayesian-deep-learning-e312571042bb">Fast and Scalable Estimation of Uncertainty using Bayesian Deep Learning</a> was originally published in <a href="https://medium.com/lean-in-women-in-tech-india">Lean In Women In Tech India</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An IISc Seminar: The Beginning of VNIR]]></title>
            <link>https://medium.com/acm-w-manipal/an-iisc-seminar-the-beginning-of-vnir-ef5eadaf90d?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef5eadaf90d</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[seminar]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Thu, 25 Oct 2018 12:35:45 GMT</pubDate>
            <atom:updated>2018-10-25T17:45:58.256Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*acC67nwOj6OTCCnn36kJ1g.jpeg" /></figure><p>You know that a talk has left its mark in your memory lane when you can vividly reminisce every detail that was discussed, a month after. The startup centric talk conducted by Dr. Meher Prakash is an exemplar, in this regard. From the word go, the narrative style of presentation captivated the attendees. Furthermore, it was an eye-opener in the sense that there was a realisation that whole startup culture could not be stereotypically attributed to a business perspective alone. The very fact that a postdoctoral student from CalTech and ETH Zurich ventured into establishing two startups is reason enough to validate the same. This article delineates the instances and points underscored by the speaker.</p><h3>The Importance of Having a Vision</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GLb3fGUbzGms_RtOaSOFrw.png" /><figcaption>VNIR: Vision</figcaption></figure><p>Before venturing into a startup , <strong>convincing yourself of the vision</strong> and the impact is sacrosanct.<strong> Is the idea unique? Is it potentially scalable?</strong> Sky is the limit when it comes to dreaming big; he wanted to create a global impact with regard to cellular imaging. Surprisingly, their team proposed a name for the startup, only to realise that 40 other startups with similar names already exist! As inconsequential as it sounds, having 20 iterations of rejections and re-naming, spoke multitudes of the competition in their area of interest.</p><p>When their team decided on designing a product, they had to consider a few aspects.</p><ul><li><strong>Imaginable Scalability</strong>: Their business was bulk-scale. In short, they made high profits from a few target customers.</li><li><strong>Target</strong>: These were researchers primarily.</li><li><strong>Uniqueness</strong>: For a mutual inertia amongst users, he spoke of the 10X rule which has been elaborated on, later.</li><li><strong>Sensing the Market</strong>: Aggressive promotion and advertisement was essential.</li><li><strong>Product Usage and Scope</strong>: To decide if their product was for diagnosis, early diagnosis or prognosis.</li></ul><blockquote>“The startup experience is dynamic and once you step into it, it becomes addictive.”- Dr.Prakash</blockquote><h3>The 10X Rule</h3><p>With the increasing competition, are there times when you’ve wondered if your product or business will sustain? Is there any mantra in assessing if the product or idea is impactful? He highlighted the importance of this rule, in most aspects of his decision-making as an entrepreneur. <strong>If you are designing a product, see to it that it’s 10X better than the state-of-the art</strong> (in terms of application and results). Or fix the price to be 10X cheaper than the cost of similar products. It’s as simple as that! Incremental steps won’t help. <strong>Unless you’re going to have these humungous 10X leaps, your product will not draw the attention of potential customers. </strong>In short, he emphasised the application of this rule in product scalability, accuracy, uniqueness, pricing and usage. Also, this assures that you’re way ahead of your competitors. <a href="https://thenextweb.com/entrepreneur/2015/07/13/the-10x-rule-for-great-startup-ideas/">Here’s</a> a great article on this.</p><blockquote>“Basically, the 10X rule is an idea that every project you do will take 10 times more time, energy, money, and effort than assumed, so set the much more difficult expectation, and if it happens to take less, then great. Too often, people set out to accomplish challenging goals (e.g. I want to build a successful business, I want to lose 20 pounds, etc.) and don’t set even remotely realistic expectations. However hard you think it is, set expectations for it to be 10 times harder and then you’ll be in the ballpark. Put another way, most people don’t reach their goals because they didn’t work hard enough at it.” - David Cummings</blockquote><h3>Time is an Illusion</h3><p>When a customer tells you that they will get back to you in a day, they will take a week at the least. And if they say they need a week’s time, you can rest assure that they will take a month to respond. Likewise, a month should be interpreted as a few months or a year. See the trend? His venture involved a few customers who would buy the product (near infra red fluorescence probes) in bulk. Despite targeting a few potential customers, <strong>expecting timely responses and sticking to rigid deadlines turned out to be impractical</strong>.</p><h3>Early Startups Must Have Investors</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ymRNrQ39CWobWB1F38VSgg.jpeg" /><figcaption>Angel Investors</figcaption></figure><p>Before completing his postdoctoral studies, Dr. Prakash was the founder of another startup which made a non-corrosive dye for metal surfaces. Since they had to sell samples, the operation costs were minimal. Although the first startup could not sustain the desired business growth, the entire experience was a cornerstone in the ideation phase of VNIR, his second start-up venture. This time, he decided that the he would not be the CEO.</p><blockquote>“Before I stepped into the world of startups, I would always criticize overblown pricing. For instance, if a plate of Biryani could be bought for Rs.50, then why did food stalls charge Rs.250 for the same? <strong>It later dawned on me that they have to compensate for the bad days</strong>. The days where they have low sales, fewer customers, increasing costs of raw materials (ingredients) or lesser employees.”<br>-Dr. Prakash</blockquote><p><strong>Inspite of being approached by many angel investors, he didn’t approve of them. This was one of his biggest mistakes. </strong>He confessed that he was hell-bent on approaching them later, when his company made <em>significant profits</em>. And once his company did make reasonable profits, this was tabled for a future date when he would make <em>some more profit</em>. But that never happened.</p><p>When we fast forward to a few years later, he invested 1.5 crores in VNIR and started looking for angel investors well in advance. Planning for the next 1–2 years of operation was a priority.</p><h3>Incubators and Government Funding</h3><p>A business incubator is an established institution that provides mentorship, initial funding, work space, investment and networking opportunities. At present, VNIR has its base at the Jawaharlal Nehru Centre for Advanced Scientific Research (JNCASR), Bangalore. According to him, <strong>the benefits of such incubators must be leveraged</strong>. If you’re looking for startup incubators in India, Bangalore is teeming with them!</p><p>What about government funds? Well, you could certainly use them…if they are adequate. He recollected an incident where they wanted to avail 50 lakhs via government funds. The Indian government supported them with 20 lakhs. Guess that’s better than nothing. Having said that, <strong>one cannot solely depend on government funds to support their startups</strong>.</p><h3>Does a Startup Need a Salesperson?</h3><blockquote>“Most of the people we approached, gave us advice, schemes and plans on how to increase profits. But none of them spoke of ways in which we could make profits while starting at ground zero, with no revenues.” — Dr. Prakash</blockquote><p>This depends on what the startup is about and its scope. They did not find the need of a salesperson or someone with a degree in Marketing, since their business was small-scale. In fact, what they required was a <strong>financial advisor</strong>. Their financial advisors had a better market awareness and helped in making business-decisions, so that they could scale up their business.</p><p>In this context, he mentioned the concept of a sales funnel. <strong>Creating a sales funnel helped them analyze the path taken by prospects in becoming customers.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/748/1*8luH-e5jorPCyBoojiItsA.png" /><figcaption>Sales Funnel</figcaption></figure><p>Lastly, a common notion would be to assume that the founder is a know-it-all. <strong>Ironically, he or she is ‘figuring out’ what works best for the startup company- just like any other employee.</strong></p><blockquote>“My job was to make my research partner realise that he has the ability to discover something new. I didn’t discover it. I motivated him to.”- Dr. Prakash</blockquote><h3>The Random Walk</h3><blockquote>“Without unpredictability, there is no thrill.”- Dr. Prakash</blockquote><p>His journey was full of ups and downs. Initially, his first startup venture saw 6 months of success. His bank accounts were brimming with cash. However, things started taking a downward spiral eventually. He realised that his product could not sustain the market. This led to fewer customers and that, in turn, led to lower revenues. It was a domino effect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/650/1*7BxCFuhZcS6bb6X0iG1c_g.jpeg" /><figcaption>Mitochondria Strain (VNIR Biotechnologies pvt ltd)</figcaption></figure><p>In the midst of such a downfall, he decided to pursue his postdoctoral research. However, his startup craze did not die down. Along with his colleague at JNCASR, he came up with the idea of infra red fluorescence probes for cell imaging.</p><p>Generally , a blood sample is smeared on a glass microscope slide and stained subsequently. Following this, an expert analyzes the slides. They wanted to make chemical probes that would help in diagnostics by lighting the regions under study instead of existing stains. To this end, <strong>they offered a fluorescence molecule with far-reaching applications.</strong> For instance, this could also be used with the cerebrospinal fluid obtained during a lumbar puncture procedure, in order to detect Alzihmer’s disease. Was their idea unique? Have a look at the figure and think for yourself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MhlL1Vh5isW_kT6CfnFeeQ.png" /><figcaption>The uniqueness of their product</figcaption></figure><h3>Conclusion</h3><p>Their immediate goals are as follows:</p><ul><li>To find ways in which their product can be modified for prognosis.</li><li>To scale-up their venture to a corporate level. This would ensure better revenues.</li></ul><p>VNIR has several patents to their name and continues to make significant contributions to the research sector. You can visit their <a href="http://vnir.life/">website</a> to learn about their current work.</p><p><em>Note: This is my first attempt at writing a startup related article. Any suggestions are welcome. Also, this article certainly doesn’t do justice to the talk.</em></p><p><em>If you’re heading a startup, do check out the </em><a href="https://developers.google.com/programs/launchpad/accelerators/"><em>Launchpad Accelerator Program </em></a><em>by Google Developers’ Group. Additionally, if you’re a female entrepreneur, make sure that you apply for the </em><a href="https://ghcindia.anitab.org/women-entrepreneur-quest-2018/"><em>AnitaB.org Women Entrepreneur Quest </em></a><em>:)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef5eadaf90d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/acm-w-manipal/an-iisc-seminar-the-beginning-of-vnir-ef5eadaf90d">An IISc Seminar: The Beginning of VNIR</a> was originally published in <a href="https://medium.com/acm-w-manipal">ACM-W Manipal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An IISc Lecture: Deep Learning Research- Representation Learning]]></title>
            <link>https://medium.com/@sukritipaul005/an-iisc-lecture-deep-learning-research-representation-learning-62176d3bc66?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/62176d3bc66</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[iisc]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Mon, 18 Jun 2018 18:15:48 GMT</pubDate>
            <atom:updated>2018-06-18T18:47:11.238Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8fIxT9kd0ZZZYdtryimfaQ.png" /><figcaption>Fig.1 : Explainable AI</figcaption></figure><p>In his hour long lecture pertaining to cutting-edge research in deep learning(DL), <a href="https://cedar.buffalo.edu/~srihari/">Professor Sargur Narasimhamurthy Srihari </a>touched upon various topics including the<em> significance of representation, representation learning methods, transfer learning, disentangling variables and explainable artificial intelligence (AI)</em>. Prof. Srihari is a SUNY Distinguished Professor in the Department of Computer Science and Engineering at the University at Buffalo. A pioneering researcher, his research efforts have resulted in the first large-scale <a href="http://www.cedar.buffalo.edu/papers/articles/HWAI-RCR97.pdf">handwritten address interpretation systems</a> in the world (deployed by the IRS and USPS), post-Daubert court acceptance of handwriting testimony based on <a href="http://www.cedar.buffalo.edu/~srihari/papers/JFS-2002.pdf">handwriting individuality assessment</a>, a software system in use by <a href="http://www.cedar.buffalo.edu/~srihari/papers/JFS2008-color.pdf">forensic document examiners</a> worldwide, statistical characterization of uncertainty in <a href="http://www.cedar.buffalo.edu/~srihari/papers/IEEE-Spectrum.pdf">impression evidence</a> and first characterization of <a href="http://www.cedar.buffalo.edu/~srihari/papers/DocumentImageAnalysis.pdf">document image analysis</a> as a sub-field of pattern recognition. I strive to highlight important points of the session.</p><p><em>Note: Most of the images have been taken from sir’s lecture slides and are available on his website.</em></p><h3>The Importance of Representations</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/740/1*fnXfMZMzB3EJhEiAr4ZCKA.png" /><figcaption>Fig.2 : Changing the coordinate system for a given data-set</figcaption></figure><p>Synonymous with the concept of having different metrics to measure distances (like miles and km), having good representations for our problem statements are vital for enhanced and effective learning. You must be wondering how representations matter to a computer science student. Via a simple example of inserting a number in a sorted list versus inserting a number into a list which is represented as a red-black tree, it can be shown that the complexity varies significantly. <em>The former has a complexity of O(n), whereas the complexity of the latter is O(log n).</em></p><p>But what has this got to do with machine learning (ML)? Is there an ML-centric example which depicts the importance of representations? Fig.2 depicts how changing the coordinate system ((x,y)-&gt;(r,θ)), can influence the ease with which classification is performed. In the second case, a linear classifier can be used effectively.<em> Representations should be disentangled, noise invariant, informative and uncomplicated </em>(when it comes to working with them). Representations do matter- we aim at learning representations after all!</p><h4>Role Played by Deep Learning When it Comes to Representations</h4><p>To cut things short, <strong>deep learning finds representations</strong>. The dark regions imply that the corresponding components can learn from data. Extending the concept of representation learning to multiple layers gives rise to deep learning frameworks, as shown below (Fig.3).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*5FmyJ6H1SaEXDeS7nTgb2g.png" /><figcaption>Fig.3: Representation Learning</figcaption></figure><p>If you’re a deep learning enthusiast, it is plausible that you’re au fait with the next few topics. But that’s not why I’ve penned down my learnings; this article doesn’t focus on brushing up your basics or going into the nitty gritties. It’s more about facts, clearing ambiguities and discussing current research work.</p><p>Representation learning methods comprise the following:</p><ol><li>Autoencoders</li><li>Convolutional Neural Networks (CNN)</li><li>Recurrent Neural Networks (RNN)</li></ol><h3>Autoencoders</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*VYH3i2-2CZ6Fyd7Bv9UHFw.png" /><figcaption>Fig.4: An autoencoder (<a href="https://www.researchgate.net/publication/320658590_Deep_Clustering_with_Convolutional_Autoencoders">Deep Clustering with Convolutional Autoencoders</a>)</figcaption></figure><p>Autoencoders are unsupervised learning algorithms. The following facts hold true for <strong>deterministic</strong> autoencoders. Encoders are designed to obtain <strong>compressed and highly informative representations</strong> of the input. Compressed forms hold relevant information and data properties pertaining to the task . Decoders are expanded/converted to reconstruct the input. Sparse autoencoders penalize the compressed representations (in the hidden layer), via a sparsity penalty.</p><blockquote>“The model is forced to prioritize which aspects of input are copied”.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*-1AmDPoSm-Emk1WD0RKS6A.png" /><figcaption>Fig.5: An example of a 1-D manifold</figcaption></figure><p>Another notable fact would be that<strong><em> autoencoders learn manifolds.</em></strong> In layman’s terms, imagine a manifold to be a subset of the vector space (ambient space) of the original data i.e. a <strong><em>low dimensional space </em></strong>associated with the compressed representation. <strong><em>A manifold can have several dimensions.</em></strong> <strong>An interesting observation would be that any atypical behaviour can be observed if the input is not on the manifold.</strong> Essentially, a manifold can be perceived as a projection onto a lower dimensional space (which is why sir has quoted an example of PCA).</p><blockquote>“Autoencoders learn a representation function that maps any point in ambient space to its embedding”.</blockquote><p>An embedding is a low dimensional vector or a point on the manifold. Fig.6 depicts embeddings on a manifold of familiar classes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1014/1*gKiHeQlIL8R1895KNb_M8g.png" /><figcaption>Fig.6: Embedding on a manifold</figcaption></figure><h4><strong>How is the compression different from JPEG or Mp3 compression and how do we harness the power of these representations?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/459/1*CMqK9hRgkm9BEqzziJAhqQ.png" /><figcaption>Fig.7: Deepstyle example</figcaption></figure><p>Autoencoders are <strong><em>lossy and learnt</em></strong>. Furthermore,<strong><em> the intermediate compressed representations can be modified in the encoded space, to obtain outputs that have never encountered. </em></strong><a href="https://arxiv.org/pdf/1801.03002.pdf"><strong>DeepStyle</strong></a><strong> </strong>is an example of the same.</p><blockquote>“Using one input image and changing values along different dimensions of feature space you can see how the generated image changes (patterning, color texture) in style space.”</blockquote><p>Another instance where compression can play an important role is in the case of analyzing tweets; a researcher is known to have analyzed 198 billion tweets with 2.8 billion words in the training set where a 30,000 dimensional One hot vector can be represented by a compressed vector of 300 dimensions (rich with information). A recent paper termed, <strong>‘Embedding Historical Linguistics Data’ </strong>makes use of words shared across various languages to obtain a hyperbolic geometry (in the compressed form) as shown in Fig.9. This is an ICML’18 proceeding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/890/1*lE_QM-AbNIVzaDv-Cp1lAw.png" /><figcaption>Fig.9: Embedding Historical Linguistics Data</figcaption></figure><h3>The Underlying Concept Behind CNNs</h3><h4>Understanding Sparse Connectivity</h4><p>Sparse interactions are a <strong><em>consequence of the kernel size being smaller than the input image size.</em></strong> For simplicity sake, imagine a hidden unit in layer L , as the input to the hidden units in layer L+1. In a traditional feed forward neural network, the current hidden unit will serve as an input to <strong>ALL the hidden units of layer L+1</strong> (unless you apply dropout regularization). The crux in designing CNNs, lies in the fact that the<strong><em> current hidden unit will impact or effect only k hidden units of layer L+1</em></strong>, where k is the kernel size. Thus, this is an implication of sparse connectedness.</p><h4>CNNs and Matrix Multiplication</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/433/1*xzsfhEUxuUVlXi66KO98XA.png" /><figcaption>Fig.10: CNNs vs Matrix Multiplication</figcaption></figure><p>Operations associated with CNNs are <strong>not 1–1 with mathematical convolution</strong>, owing to sparsity. CNNs ignore the effects of one layer on another. Fig.10 represents CNNs versus traditional matrix multiplication. The shaded regions s_x correspond to the units that are impacted by x_3.</p><h3>Computational Graphs for RNNs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*79Er8f7OPu39M3f1KR2YKw.png" /><figcaption>Fig.11: RNN-Unfolding a graph</figcaption></figure><p>An RNN can be perceived as an <strong><em>unfolded form of a graph of RNN, over various time instances.</em></strong> Fig.11 represents how a single computational graph can be replicated over several time instances. Cycles signify how the current value of a variable impacts its future value, at a particular time step.<em> Let us assume that a person discusses about Kalman filters, what she ate for breakfast and returns to the topic of Kalman filters. Although she has made a connection of her references to something that she quoted earlier, how does a system correlate current context with prior information ?</em><strong> Contextual information is pivotal in such a scenario and the concept of ‘larger memory over a longer time’, is sacrosanct in such NLP related tasks.</strong> This gave rise to <strong><em>LSTMs</em> </strong>where the <em>cell state or conveyor belt</em> is crucial in transferring previous information.</p><h3>Transfer Learning</h3><p>If you’re a newbie to DL or if you don’t have the required hardware to train deep networks, this is an ideal first assignment! We use transfer learning when:</p><ul><li>Two or more tasks have to be performed.</li><li>We can safely assume that the factors describing the variations in task X , are relevant to the variations that are to be considered for learning task Y.</li></ul><p>For example, a DL framework designed for the recognition of handwritten English numerals, can be used for Indic script numeral recognition. Transfer learning is characterised by <em>shared representations </em>and is known to improve the performance of DNNs.</p><h3>Understanding an Optimal Representation for ML</h3><p>A good representation will make the classification task easier as we progress through successive layers. It is plausible that classes which are not linearly separable in input features, become linearly separable in the final layer.<strong><em> The entire task can be viewed as learning a single function f(x)</em></strong>, which comprises functions like f(1), f(2) and f(3). <strong><em>At each stage of solving for f(x), the classification becomes easier, ideally.</em></strong></p><p>f (x)= f (3) [ f (2) [ f (1)(x)]]</p><blockquote>“An ideal representation is one that disentangles the underlying causal factors of variation that generated the data, especially those that are relevant to our application. Most strategies for representation learning are based on introducing clues that help the learning to find these underlying factors of variation.”</blockquote><h3>Latest Buzz</h3><p>There are a few more concepts that I will be covering in subsequent blog articles. Here are a few ongoing areas of research in DL, that you could explore.</p><ol><li><strong>Disentangling or Untangling Variables.</strong></li><li><strong>Semantic Manifold and Manifold Learning </strong>: Prof. Sargur used an example of handwritten text to obtain images which were generated from a manifold. The core idea behind this was to travel along a dimension in latent(hidden) space, ensuring that the other dimensions wereunaltered. His PhD students are working on deep learning for handwritten comparison; unfortunately, its details are unavailable as it’s an ongoing project using variational autoencoders for forensic comparison. Incase you’re interested in handwritten recognition, give <a href="https://cedar.buffalo.edu/papers/articles/Individuality_Handwriting_2002.pdf">this</a> a read .</li><li><strong>Capsule Networks </strong>: It’s rather ironical that <a href="https://www.youtube.com/watch?v=rTawFwUvnLE">Geoffrey Hinton expresses his dismay</a> in using CNNs, stating that the pooling operation is a blunder and its apt working could be cataclysmic for the DL community. Recently, he proposed <strong>CapsNet</strong> to overcome the drawbacks held by CNNs. <a href="https://medium.com/ai%C2%B3-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b">Here’s</a> a concise blog article regarding the same. (I was introduced to this, by a friend).</li><li><strong>Explainable Artificial Intelligence</strong>: Ever wondered why the answer given by a DNN, is the answer? The algorithm certainly isn’t the answer and it’s high time we contemplate something that is in-between the input and the output. Researchers are working at incorporating an <strong><em>explanation interface </em></strong>(as shown in Fig.1) to the learning system. This will help in answering some of the following questions:</li></ol><ul><li>Why was X performed?</li><li>Why wasn’t Y or any other method performed?</li><li>What are the scenarios where this will be correct (or incorrect)?</li><li>How can mistakes pertaining to X be corrected?</li><li>When can we trust this? (If you’re working on detecting malignancy in images or deciding the execution of a criminal, errors are unacceptable).</li><li>Will this cheat, in any scenario? (For example, a recent experiment associated with recognizing horses, worked incorrectly despite optimal results, as each of the input images pertaining to horses had a copyright symbol. The classifier used the copyright symbol as an implication of a horse :P).</li></ul><blockquote>“Most of AI is a black-box when it comes to learning”.</blockquote><p>5. <strong>Extending concepts of Variational Autoencoders (VAE) and Generative adversarial networks (GANs) to explainable AI</strong>.</p><h3>Open-ended Questions by the Audience</h3><ol><li>Does explainable AI aim at improving existing frameworks or is it a new and unconventional approach altogether?</li><li>As representation learning advances, feature representations will change. Can the current DL models handle this? (Apparently Kalman filters and LSTMs can handle this).</li><li>Will there be a point where DL beats classical ML in all scenarios?</li><li>Can transfer learning be conceived as a regularizer?</li></ol><p><em>Note: This is an overview of the seminar. I shall post a few more articles in detail. None of the above-mentioned facts are my findings; I have merely jotted down points that were discussed in the lecture. Feel free to correct me or add a few more ongoing research areas in DL (there are a plethora of them, by now) :)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62176d3bc66" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learn How to Train U-Net On Your Dataset]]></title>
            <link>https://medium.com/coinmonks/learn-how-to-train-u-net-on-your-dataset-8e3f89fbd623?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e3f89fbd623</guid>
            <category><![CDATA[convolution-neural-net]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Fri, 08 Jun 2018 22:36:31 GMT</pubDate>
            <atom:updated>2022-06-24T17:44:41.047Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sj5U51StQ16r_x38t7yHuA.png" /><figcaption>Fig.1 : A test image along with its label (semantically segmented output)</figcaption></figure><p>With the aim of performing semantic segmentation on a small bio-medical data-set, I made a resolute attempt at demystifying the workings of U-Net, using Keras. Since I haven’t come across any article which explains the training steps systematically, the thought of documenting this for other deep learning enthusiasts, occurred to me. Some of you must be thinking if I’ve covered the theoretical aspects of this framework. Although my primary focus is to elaborate on the implementation, I try to include details relevant to its understanding as well. Most of my references include<strong> </strong><a href="https://github.com/zhixuhao"><strong>zhixuhao</strong></a><strong>’s unet repository</strong> on Github and the paper, ‘<strong>U-Net: Convolutional Networks for Biomedical Image Segmentation’ </strong>by<strong> </strong><a href="https://arxiv.org/search?searchtype=author&amp;query=Ronneberger%2C+O">Olaf Ronneberger</a> et.al.</p><h3>About U-Net</h3><p>The U-net architecture is synonymous with an encoder-decoder architecture. Essentially, it is a deep-learning framework based on <a href="https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf">FCNs</a>; it comprises two parts:</p><ol><li>A contracting path similar to an encoder, to <strong>capture context via a compact feature map.</strong></li><li>A symmetric expanding path similar to a decoder, which allows <strong>precise localisation</strong>. This step is done to retain boundary information (spatial information) despite down sampling and max-pooling performed in the encoder stage.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*TXfEPqTbFBPCbXYh2bstlA.png" /><figcaption>Fig.2: Architecture of U-Net based on the paper by <a href="https://arxiv.org/search?searchtype=author&amp;query=Ronneberger%2C+O">Olaf Ronneberger</a> et.al</figcaption></figure><h4>Advantages of Using U-Net</h4><ol><li>Computationally efficient</li><li>Trainable with a small data-set</li><li>Trained end-to-end</li><li>Preferable for bio-medical applications</li></ol><p>Wait, what on earth are semantic segmentation and localisation? In simple terms, they refer to <strong>pixel-level labelling </strong>i.e. each pixel in an image is provided with a class label. You get segmentation maps, which look like the one in Fig.1. Prior to this, bio-medical researchers followed two approaches:</p><ol><li>Classifying the image as a whole (malign or benign).</li><li>Dividing the images into <strong>patches </strong>and classifying them.</li></ol><p>Owing to an increased data-set size, patching was certainly better than a whole image classification, however, there were a few drawbacks pertaining to the same. Smaller strides or patching with a lot of overlapping is both, computationally intensive and results in redundant (repetitive) information. Secondly, <strong><em>a good tradeoff between context information and localisation is vital</em></strong>.<strong> Small patches result in loss of contextual information whereas large patches tamper with the localization results.</strong> Lastly, non-overlapping patches result in a loss of context information. Based on prior observations, an encoder-decoder architecture yields much <strong>higher intersection over union (IoU) </strong>values than feeding each pixel to a CNN for classification.</p><h3>Preparing the dataset</h3><p>Let’s dive into preparing your data-set! I will be writing a separate article covering the nitty-gritties of U-Net, so don’t rack your brains and get bogged down with its architecture.</p><ol><li>Divide your original dataset and corresponding annotations into two groups, namely, <em>training and testing</em>(validation, to be more precise) sets. The <strong>original images are in RGB</strong> while their<strong> masks are binary</strong> (black and white).</li><li>Convert the all of the image data to <strong>.tif.</strong></li><li>You won’t be needing the image class labels or the annotations of the test set (this is not a classification problem).</li></ol><p><strong>NOTE:</strong> The image size should be selected, such that the consecutive convs and max-pooling yields <strong><em>even values of x and y </em></strong>(i.e. width and height of the feature map)<strong><em>at each stage</em></strong>. Although my images were 360X360 pixels, I resized them to 256X256. Crop out the borders to get an appropriate image size. I have included the code, for this.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/74ffeb9d4b4b06904451e53bad032678/href">https://medium.com/media/74ffeb9d4b4b06904451e53bad032678/href</a></iframe><p>You must be wondering what to name the folders, yada yada yada..wait for a few more minutes and you’ll know exactly where to place them!</p><h3>Pre-requisites</h3><ul><li>Tensorflow</li><li>Keras &gt;= 1.0</li><li>libtiff(optional)</li><li>OpenCV 3 (if you’re a MAC user, follow <a href="https://www.codingforentrepreneurs.com/blog/install-opencv-3-for-python-on-mac/">this</a> for the installation)</li></ul><p>Also, this code should be compatible with Python versions 2.7–3.5.</p><h3>Training &amp; Data-augmentation</h3><p>You can rotate, reflect and warp the images if need be. <a href="https://github.com/zhixuhao">zhixuhao</a> uses a deformation method available <a href="https://github.com/cxcxcxcx/imgwarp-opencv">here</a>. Follow the next few steps carefully; missing a step can make you go bonkers for hours! I have divided the training stage, into two parts- Part A and Part B, for simplicity.</p><h4>Part A- Modifying data.py</h4><p>1.Clone <a href="https://github.com/zhixuhao">zhixuhao</a>’s repository.</p><pre>$ git clone <a href="https://github.com/zhixuhao/unet">https://github.com/zhixuhao/unet</a></pre><p>2.Enter the <strong><em>image</em></strong> folder, which lies within the train folder (<em>../unet/data/train/image</em>).</p><p>3.Include the training images in the image folder. <strong>Each of the images should be in the .tif form, named consecutively, <em>starting from 0.tif, 1.tif…and so on.</em></strong></p><p>4.Get into the <strong><em>label </em></strong>folder which lies within the train folder (<em>../unet/data/train/label</em>). Include the corresponding train-image annotations. <strong>Each of the images should be in the .tif form, named consecutively, <em>starting from 0.tif, 1.tif…and so on. Labelling must correspond to the training image-set</em></strong>.</p><p>5.Enter the test folder which lies within the data folder (<em>../unet/data/test</em>).</p><p>6.Create a folder called <strong><em>npydata</em></strong> within the data folder (<em>../unet/data/npydata</em>). Let this remain empty; the processed data-set will be saved in this, as 3 .npy files, subsequently.</p><p>7.Open the <strong><em>data.py</em></strong> file in the unet folder (<em>../unet/data.py</em>). Steps 8,9, 10 and 11 refer to the changes that you will have to make in this file, for RGB images. The regions in bold correspond to the changes made by me.</p><p>8.Modify <em>def create_train_data(self) </em>as shown below.</p><pre>def create_train_data(self):<br>...<br>  imgdatas = np.ndarray((len(imgs),self.out_rows,self.out_cols,<strong>3</strong>),     dtype=np.uint8)<br>  imglabels =  np.ndarray((len(imgs),self.out_rows,self.out_cols,1),   dtype=np.uint8)<br>...<br>  img = load_img(self.data_path + &quot;/&quot; + midname) <strong>#Removed grayscale</strong><br>  label = load_img(self.label_path + &quot;/&quot; + midname,grayscale = True)<br>#Correspond to lines 159-164</pre><p>9.Modify def <em>create_test_data(self) </em>as shown below.</p><pre>def create_test_data(self):<br>…<br> imgdatas = np.ndarray((len(imgs),self.out_rows,self.out_cols,<strong>3</strong>), dtype=np.uint8)<br> …<br> img = load_img(self.test_path + &quot;/&quot; + midname) <strong>#Removed grayscale<br></strong>#Correspond to lines 188 and 191</pre><p>10.Change<strong><em> imgnum</em></strong> to the size of your train-set.</p><pre>def doAugmentate(self, img, save_to_dir, save_prefix, batch_size=1, save_format=&#39;tif&#39;, <strong>imgnum=26</strong>): #I&#39;ve considered 26 training images</pre><p>11. Modify the following in class dataProcess(object) by providing the correct path details. I changed the <strong><em>data_path , label_path, test_path and npy_path</em></strong> to the correct path (corresponding to the directories in my system). You can try editing these in line 138, in data.py. Incase a few errors pop up, go though my answers to<a href="https://github.com/zhixuhao/unet/issues/40"> issue #40</a> on Github. Please see to it that the path to the npydata folder is not wrong (this is a common mistake).</p><pre>def __init__(self, out_rows, out_cols, <strong>data_path</strong> = &quot;/Users/sukritipaul/Dev/newcvtestpy2/unet2/data/train/image&quot;, <strong>label_path</strong> = &quot;/Users/sukritipaul/Dev/newcvtestpy2/unet2/data/train/label&quot;, <strong>test_path</strong> = &quot;/Users/sukritipaul/Dev/newcvtestpy2/unet2/data/test&quot;, <strong>npy_path</strong> = &quot;/Users/sukritipaul/Dev/newcvtestpy2/unet2/data/npydata&quot;, img_type = &quot;tif&quot;):<br>#Corresponds to line 138</pre><p>12.Run data.py</p><pre>$python data.py</pre><h4>Part A- Verification</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/432/1*wV9UpIBrGq2cXjGaMdhbGw.png" /><figcaption>Fig 3: Output obtained on running data.py</figcaption></figure><p>Your output should match the output in Fig.3. I’ve printed the sizes of the training images, training annotations and test images. If your terminal shows the correct number of training and testing images (8 and 4 in this case),then you’re good to go! If it shows 0, then your files have not been included in the .npy files. Check if you have the following files in <em>/unet/data/npydata</em>:</p><ol><li>imgs_mask_train.npy</li><li>imgs_test.npy</li><li>imgs_train.npy</li></ol><p>Wohooo! You’re done with the data preparation bit :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/346/1*xWYQ_4yvBkVDm0na5ykKlg.png" /><figcaption>Fig4: Result on the ISBI cell tracking challenge- input and cyan mask</figcaption></figure><h3>Part B- Modifying unet.py</h3><ol><li>Create a <strong><em>results</em></strong> folder in the unet folder (../unet/results). If you’re thinking why you’ve created this- you’ll know why, in a jiffy!</li></ol><p>2. Open <strong><em>unet.py</em></strong> (.<em>./unet/unet.py</em>)</p><p>3. Edit the number of rows and columns in the following line. My images had dimensions of 256X256.</p><pre>def __init__(self, img_rows = 256, img_cols = 256):<br> #Corresponds to line 13</pre><p>4. Modify the following for a 3 channel input, in <em>get_unet(self)</em>, as shown below.</p><pre>def get_unet(self):<br> inputs = Input((self.img_rows, self.img_cols,3))<br> #Corresponds to line 27</pre><p>5. Modify the following lines in <em>train(self).</em></p><pre>def train(self):<br>...<br> np.save(&#39;<strong>/Users/sukritipaul/Dev/newcvtestpy2/unet2/results/imgs_mask_test.npy</strong>&#39;, imgs_mask_test)<br>...<br>#Note that the address to the <strong>results </strong>directory must be provided<br>##Corresponds to line 164</pre><p>6. Modify <em>def save_img(self), </em>keeping in mind the <strong><em>address of the results directory, as specified in step 4.</em></strong></p><pre>def save_img(self):<br> imgs = np.load(&#39;/Users/sukritipaul/Dev/newcvtestpy2/unet2/results/imgs_mask_test.npy&#39;) <strong># Use the same address as above</strong><br> img.save(&quot;/Users/sukritipaul/Dev/newcvtestpy2/unet2/results/%d.jpg&quot;%(i)) #Saves the resulting segmented maps in /results<br>##Corresponds to lines 169 and 173 </pre><p>7. Run <strong><em>unet.py</em></strong> and wait for a couple of minutes (your training time could take hours depending on the dataset size and system hardware).</p><pre>$python unet.py</pre><h4>Part B- Verification</h4><p>Visit <em>/unet/results</em> and viola! You can find the generated image masks or segmented feature maps in grayscale :)</p><p>In conclusion, I got an overall accuracy of 90.71% on 65 training images and 10 validation images of size 256X256. They have used binary cross-entropy as the loss function. It took me approximately 3 hours to train on an 8 GB MacBook Air with a 1.8 GHz Intel Core i5 processor (CPU).</p><p><em>Feel free to include your queries as comments! :)</em></p><blockquote>Join Coinmonks<a href="https://t.me/coincodecap"> Telegram Channel</a> and<a href="https://www.youtube.com/c/coinmonks/videos"> Youtube Channel</a> get daily <a href="http://coincodecap.com/">Crypto News</a></blockquote><h4>Also, Read</h4><ul><li><a href="http://Top 4 Telegram Channels for Crypto Traders">Crypto Telegram Signals</a> | <a href="https://medium.com/coinmonks/crypto-trading-bot-c2ffce8acb2a">Crypto Trading Bot</a></li><li><a href="https://medium.com/coinmonks/top-10-crypto-copy-trading-platforms-for-beginners-d0c37c7d698c">Copy Trading</a> | <a href="https://medium.com/coinmonks/crypto-tax-software-ed4b4810e338">Crypto Tax Software</a></li><li><a href="https://coincodecap.com/grid-trading">Grid Trading</a> | <a href="https://medium.com/coinmonks/the-best-cryptocurrency-hardware-wallets-of-2020-e28b1c124069">Crypto Hardware Wallet</a></li><li><a href="https://medium.com/coinmonks/crypto-exchange-dd2f9d6f3769">Crypto Exchange</a> | <a href="https://medium.com/coinmonks/buy-bitcoin-in-india-feb50ddfef94">Crypto Apps in India</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e3f89fbd623" width="1" height="1" alt=""><hr><p><a href="https://medium.com/coinmonks/learn-how-to-train-u-net-on-your-dataset-8e3f89fbd623">Learn How to Train U-Net On Your Dataset</a> was originally published in <a href="https://medium.com/coinmonks">Coinmonks</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Graph Convolutional Neural Networks- A Talk at IISc]]></title>
            <link>https://medium.com/acm-w-manipal/graph-convolutional-neural-networks-a-talk-at-iisc-502aa258bfdc?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/502aa258bfdc</guid>
            <category><![CDATA[cnn]]></category>
            <category><![CDATA[deep-neural-networks]]></category>
            <category><![CDATA[graph-nn]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[acm-w]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Wed, 21 Feb 2018 17:00:13 GMT</pubDate>
            <atom:updated>2018-02-21T17:10:49.167Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>“ What if we could input a computer network as a graph, directly, and use neural networks to predict..maybe the top few computers that are prone to attacks by hackers?”<br>- Dr. Tijmen Tieleman</blockquote><p>Being an intern at IISc has its own perks, one of them being the talks given by researchers in trending technological domains. I was fortunate to attend a talk recently, organised by the IEEE Signal Processing Society, Bangalore Chapter and the Electrical Engineering department at IISc. Current CTO of mind.ai and a former graduate from Geoff Hinton’s lab (University of Toronto), Dr. Tijmen Tieleman introduced us to the concept of graph convolutional neural networks within an hour! I’ve tried to pen down the remnants of the session, from my perspective. This one’s for the AI &amp; Deep Learning enthusiasts.</p><p>We are accustomed to focussing on grid-based data, in the Euclidian domain, with defined neighbours. For example, in case of a traditional CNN, we know the neighbourhood of an image pixel, especially while applying filters in the convolutional or pooling layers (I’ll elaborate on this a little later). Imagine a scenario where we wish to input data in the form of a graph, with all the neighbourhood connections maintained. Many scenarios require data to be analysed in the non-Euclidean space, as graphs. Computer networks, social media networks, studies pertaining to genetics and chemical bonds- all of these contain data which are in the form of graphs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3ku00_hhs3kbjJl6amChLw.png" /><figcaption>About the talk</figcaption></figure><h3>Graph CNNs versus Grid CNNs</h3><blockquote>“Graph CNNs fall under bleeding edge research, when it comes to deep learning.” <br>- Dr. Tijmen Tieleman</blockquote><p>The traditional grid CNNs have two characteristics which pose to be challenges in case of graph CNNs. It’s difficult to assign weights to nodes when the relative location of the nodes are not in the grid format. <em>How does one know which node to assign w1, w2 or w3 to, without a 3x3 or NxN spatial structure? </em>The following represent the challenges faced while designing a graph CNN.</p><ul><li><strong>No defined neighbourhood </strong>(What are the neighbours?)</li><li><strong>The weight vectors pertaining to neighbours are unknown</strong> (Which weight should be assigned to which neighbour?)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VDQyn-GoBhectN-fy0VT7A.jpeg" /><figcaption>Traditional Grid CNN</figcaption></figure><p>Hence, researchers started looking at localisation techniques. There is an inherent tension between semantics and location in semantic segmentation: <strong>local information resolves <em>where</em> while global information seeks to resolve <em>what</em>.</strong> <em>1x1 CNN or Fully Convolutional Networks combine shallow, appearance information with deep, coarse semantic information. </em>Hence, although using a 1X1 filters may seem a little absurd owing to the fact that the neighbouring pixels are not considered for convolution, converting fully connected layers to convolutional layers with this concept proves to be effective in terms of localised information. The method proposed by Jonathan Long et al. learns end-to-end. Addressing the two issues earlier, the following architecture was proposed:</p><ul><li><strong>Use the same weight <em>w</em> for a particular layer, for a particular graph node</strong>(instead of w1, w2, w3…). This resolves location issues.</li><li><strong>Each layer will have the same architecture in terms of node connections. </strong>This solves the issue of neighbourhoods.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*84xjZjmasxw5GSRzuBzVmA.jpeg" /><figcaption>Graph CNNs with the same weight</figcaption></figure><p>Using concepts of graph theory can be useful in several cases. Most of the current research work focuses on the <strong>spatial domain</strong>, however, results in the spectral domain also prove to be efficient and comparable to state-of -the art results. For instance, average pooling can be substituted by graph clustering!</p><p>A modification of this would be to have the same weights for immediate neighbours and a different weight for a loop, as shown. Furthermore, a graph network within a graph network, can store the local node information per node. This is inspired by the Network in Network architecture where each node is another smaller network, instead of simply computing a convolution followed by a non-linear activation at once! Following this, the audience had many doubts; a plethora of them are open-questions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t2wyIZikl-8gVOEqKp6QYg.jpeg" /><figcaption>Modified Graph CNNs</figcaption></figure><h3>Questions from the audience</h3><ol><li>Why can’t we use an adjacency matrix to represent a graph and then use a standard 2D CNN to it?<br>Graph CNNs are designed to also restore connectedness across various layers. Not only will the matrix be extremely sparse, information regarding connectedness will be lost over further layers. Also, this doesn’t help in a one-shot representation of a node, say Node A, connected with Node B and Node C simultaneously.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JfzXPJidK1voaWz4j2ucrw.jpeg" /><figcaption>Flattening of nodes</figcaption></figure><p>2. Can graph CNN concepts be used for Grid CNN concepts?<br>Yes. Essentially, grid representations can be viewed as a simple form of graph representations with known and structured neighbours. A good suggestion to this was that a graph representation can be viewed as a stacked or flattened grid representation where each layer can represent a central node with a weight of <strong><em>w</em></strong> for the spatial locations where other neighbouring or connected nodes are present (as shown).</p><p>3. What if 2 graph nodes are not directly connected?<br>ConvNets always follow a hierarchy, when it comes to passing information. So incase two nodes are not directly connected, their information will be combined, with corresponding weights, over deeper layers.</p><p>4. What if we have to consider weighted graphs?<br>Currently, they are working on techniques where the link weights are represented as nodes.</p><p>5. What if 2 graphs have identical relative degrees?<br>The corresponding weights will vary because two things are taken into consideration for such an ideology- <strong>connectedness </strong>(which will be the same in this case) and <strong>individual node identity</strong>(which will vary, based on the information).</p><p>6. If we use the concept of same weights to grid layers, will it be effective?<br>No. Having lesser weights reduces synapses and learning capacity. Graph CNNs use it as a novel approach, but ideally, different weights would prove to be a better approach.</p><h3>Open-ended Questions</h3><ul><li>Can we extend this concept to RNNs?</li><li>Can we extend this concept to Directed Graphs?</li><li>Can we generalise this for both, Graph CNNs and Grid CNNs?</li><li>How can we represent link information? (For example, whether it is a single, double or triple bond between atoms).</li><li>How can we distribute distinct weights better?</li><li>GPU and other hardware, to increase the computational efficiency.</li></ul><p><em>Note: This is a summary of the hour long session and none of the methodologies are proposed by me. Feel free to correct me or provide your suggestions in the comments’ section. Also, excuse my drawings :P</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=502aa258bfdc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/acm-w-manipal/graph-convolutional-neural-networks-a-talk-at-iisc-502aa258bfdc">Graph Convolutional Neural Networks- A Talk at IISc</a> was originally published in <a href="https://medium.com/acm-w-manipal">ACM-W Manipal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Yikes! It’s the Research Bogeyman]]></title>
            <link>https://medium.com/acm-w-manipal/yikes-its-the-research-bogeyman-f5225fbce982?source=rss-18efe080f332------2</link>
            <guid isPermaLink="false">https://medium.com/p/f5225fbce982</guid>
            <category><![CDATA[research-and-development]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[publication]]></category>
            <category><![CDATA[acm]]></category>
            <category><![CDATA[research-paper]]></category>
            <dc:creator><![CDATA[Sukriti Paul]]></dc:creator>
            <pubDate>Fri, 12 Jan 2018 17:08:19 GMT</pubDate>
            <atom:updated>2018-01-18T20:14:35.303Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*bwPfaIoeL9d832QrmCkG-A.jpeg" /></figure><p>Learnt a new technical tool or software recently? Let’s use it to work on a small data-set or implement a mini-project.Why not publish it, by using flowery language, resplendent images, a couple of haphazardly chosen citations and some jargon, overnight? Viola! You have one of the shallowest papers in terms of depth, analysis and content. <em>There is no gainsaying the fact that ‘Research’ is a LOT more than just publishing a paper or two by mere implementation, via a technological tool. </em>It doesn’t happen overnight and deadlines are never concrete, nothing is immutable or foreseeable at times. Here’s why I’m writing this one- because I’ve done it too. Here are some resources and tips that could help you sail through your research ventures as an undergrad.</p><blockquote>If we knew what it was we were doing, it would not be called research, would it? Albert Einstein</blockquote><h3>1. Be a Sponge</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/259/1*TVfsgCAWxZCSn-EyBj1YaA.png" /></figure><p>Pretty out of the blue eh? If you’re an aspiring researcher, well, let’s just say that it’s crucial for you to be aware and up-to-date regarding the ongoing work in your domain of interest.<em> Keep your eyes open for recent publications, news, research groups in your field of interest, corresponding conferences, research labs that offer internships, erudite professors and prominent individuals who have notable works and stellar publications in a particular domain. </em>You must be wondering how to be on the lookout for information and where to get resources- don’t worry, we have a few references to get you started!</p><h4>Research Groups , University Websites and Blogs</h4><p>There are a plethora of domain-specific research groups in universities and private sectors that you could follow. Here’s a list of a few prominent groups:</p><ol><li>CSAIL at MIT : The latest projects and impact areas are available on their <a href="https://www.csail.mit.edu/">website</a> and <a href="https://twitter.com/MIT_CSAIL?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5Eauthor">Twitter</a> page.</li></ol><p>2. <a href="https://graphics.stanford.edu/">Stanford Computer Graphics Laboratory</a> and <a href="https://www-cs.stanford.edu/research">other labs </a>at Stanford</p><p>3. <a href="http://research.baidu.com/baidu-tech-blog/">Baidu Research Blog</a> is centred around computer vision, speech recognition and synthesis, deep learning, artificial general intelligence, high performance computing and natural language understanding.</p><p>4. <a href="https://news.stanford.edu/section/science-technology/">Stanford News</a> and <a href="https://www.technologyreview.com/">MIT Technology Review</a></p><p>5.BAIR (Berkeley AI Research) T<a href="https://twitter.com/berkeley_ai?lang=en">witter</a> page and <a href="https://www2.eecs.berkeley.edu/Research/Areas/CS/">other research areas</a> at Berkeley</p><p>6. <a href="https://insights.sei.cmu.edu/">SEI Insights-</a> Carnegie Mellon University</p><p>7. <a href="https://research.google.com/">Google Research</a> and <a href="https://deepmind.com/blog/">Google Deepmind</a></p><p>8.<a href="https://www.microsoft.com/en-us/research/search/?from=https%3A%2F%2Fresearch.microsoft.com%2Fapps%2Fcatalog%2Fdefault.aspx%3Ft%3Dpublications">Microsoft Research </a>(patents are super useful) and the <a href="https://research.fb.com/publications/">Facebook AI Research blog (FAIR).</a></p><p>9. Labs at the IITs and IISc : It’s best to follow<a href="https://www.linkedin.com/school/164340/"> IISc on LinkedIn</a> for their current research ventures. On the other hand, The Hindu has articles on the latest research projects at the IITs. In addition to this, lab websites notify students about the upcoming events and projects; you never know when one of these may have openings for interns!</p><p>10. Individual blogs: <a href="http://www.computersciencedegreehub.com/top-30-computer-science-programming-blogs-2014/">Here’s a list of blogs</a> that could help you hone your technical knowledge, in general.</p><h4><strong>Follow Professors and Researchers</strong></h4><p>Keeping tabs on the publications and research work of an acclaimed professor always helps. For example, I follow <a href="https://twitter.com/drfeifei?lang=en">Dr. Fei-Fei Li</a> (Stanford Vision Lab), <a href="https://scholar.google.com/citations?user=oY9R5YQAAAAJ">Jitendra Malik </a>(UCB), <a href="http://talukdar.net/">Partha Pratim Talukdar</a> (IISc) and <a href="http://www.cs.cmu.edu/~awb/">Alan W Black (CMU).</a> Once you figure out where your interest lies, start connecting with professors beyond the scope of our college. Use Google Scholar to your advantage.</p><h4>Become an IEEE and ACM Member</h4><p>Being a member will help you leverage resources like <a href="https://cacm.acm.org/">Communications of the ACM</a>, ACM Webinars and <a href="https://cacm.acm.org/research">latest research</a>, ACM Digital Library, ACM TechNews, IEEE Xplore Digital library and IEEE Spectrum. Moreover, you’ll get to know about conferences dates and volunteering opportunities for the same. Student memberships have their own perks after all!</p><h3>2.Read, Read and Read!</h3><blockquote>“Research methodology is 85 % conceptual, mathematical and theoretical while implementation holds 15% and comes right in the end, according to my research practices” — Dr.Sanjay Singh</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/236/1*PHcozbtI6pjsoVcDUXDPOA.jpeg" /></figure><p>Reading relevant papers is of paramount importance, before jumping into conclusions or implementation. Starting from choosing the right papers to comprehending the concepts and the Math, from correlating current works with previous works to scrutinising limitations- this step covers it all! You’d be surprised at the number of people who skim through the <em>Results and Observations Section</em>, when the heart of a research paper lies in the very same; as boring as they look, the tables, diagrams, readings and stats are super important! <a href="http://blizzard.cs.uwaterloo.ca/keshav/home/Papers/data/07/paper-reading.pdf">‘How to Read a Paper’</a>, by S.Keshav provides a 3 phase approach to reading a paper effectively. There are resources on <a href="http://ento.psu.edu/graduateprograms/handbook/degree-information/degree-requirements/phd/CriticallyReadingJournalArticles1.pdf">reading journal articles</a> and <a href="http://www.sciencemag.org/careers/2016/03/how-seriously-read-scientific-paper">Sciencemag</a> provides reading strategies used by researchers. You could go through the <a href="https://blog.acolyer.org/2015/03/30/end-of-term-and-how-i-read-a-paper/">Morning Paper</a> and a <a href="https://www.eecs.harvard.edu/~michaelm/postscripts/ReadPaper.pdf">Harvard article</a> to boot, in order to plan your next read!</p><h3>3. Pen Down your Inferences- Get to Writing</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*sC1dwUoY44cTqlAgwsfKdw.png" /></figure><p>Assuming that you’ve worked on a project/domain/research statement and have obtained the desired results, presenting your results is vital to the conference or journal, where you decide to submit your paper. <a href="https://www.forbes.com/sites/quora/2018/01/08/what-is-the-best-way-to-write-a-computer-science-research-paper/#290ceb615265">David J. Malan’s</a> techniques mention a few pre-requisites to this step, simplistically. Furthermore, ‘<a href="http://www.cis.famu.edu/~cen5055joe/Administrative/HowToWrite_ResearchPaper.pdf">Writing Good Software Engineering Research Papers</a>’ by Mary Shaw, ‘<a href="http://dbis.rwth-aachen.de/~derntl/papers/misc/paperwriting.pdf">Basics of research paper writing and publishing</a>’ by Michael Derntl, <a href="https://watermark.silverchair.com/mzh053.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAAakwggGlBgkqhkiG9w0BBwagggGWMIIBkgIBADCCAYsGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMO4h_cLbMP6Kewms9AgEQgIIBXF50-QQ8zgG8wuhdmGmGdGAJjz6wA3KewQfim9X3BYEd8XSJnXC1f7gGUDIPDyNUeYBrPbvdtsibCRBGN3cltc042qpForVKYS8YoLVaKCa69yslieoiYpZK_9nfFI-BwVwX1ChUTUAqEr3CXq-EsEGNXyR7-lc2sx4e3kcVFww1PHz9PHdh5N-dGZSGzjDVEfEd7bg5fGCB40HYy1nR8Ljd3BemYusGCmk7hmQdLJdsnirdvRlr_uOWVLfX-oziYPs1DJlhYHW0V23ajJixl5ukdv3Q2GX3G1gVGYw1aPVFuNuNZxFli2hU24Udx6BL3L5_yKa8QAxcJuUzvle5pUXwLOqGuPXEl800sYVVm-9N9okjP_zZX6J7vxmaWT4Ft6fowEUvvISsBC8JcsFUXJPIG1GzVlaTqbuLSEuhHqDFDJSohDIowX9gLItXjHY065e1622DHRdIaimbow">Writing a research article: advice to beginners</a> and ‘<a href="http://www.ruf.rice.edu/~bioslabs/tools/report/reportform.html">Writing Research Papers</a>’ by the Rice University, are a few good reads. It’s best to get your paper reviewed by your colleagues and professors, once you’re done; revisions enhance precision and content.</p><h3>4. Ask Away</h3><blockquote>“Asking a question is embarrassing for a moment, but not asking is embarrassing for a lifetime”.- Haruki Murakami, Kafka on the Shore</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/271/1*km_FfSkLf4pzvSOMNNOKPA.jpeg" /></figure><p>Whether it’s about impactful conferences or trending research topics, I’ve personally found the advice provided by my professors, mentors and lab-peers invaluable. Find out what your course-mates are working on and approach the concerned faculty, for ideas and suggestions (in case you already have a problem statement in mind). Working under professors worked out for me, since they provided me with abundant internship opportunities, recommendations, connections, and personalised resources (pertaining to their research work). Never hesitate to ask!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/749/1*BSBHJEydWKLP5E4dsb3-Ng.png" /><figcaption>A glimpse into the List of Journals provided by Dr.Sanjay Singh</figcaption></figure><h3>5. Types of Papers</h3><blockquote>“A survey paper is a paper that summarises and organises recent research results in a novel way that integrates and adds understanding to work in the field. A survey article assumes a general knowledge of the area; it emphasises the classification of the existing literature, developing a perspective on the area, and evaluating trends.”- ACM Computing Surveys</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/892/1*ajIukgTzTCI7dCvOHnI0vA.png" /><figcaption>Types of papers</figcaption></figure><p>Review, survey and research papers are a few categories of conference papers. Although review and survey papers do not focus on original research work, they form valuable scientific literature by summarising pre-existing works. Research publications are culminated domain-wise, so that readers gain a general idea and overview without having to brood over reading a gargantuan number of articles. A research paper focuses on original study, collection and analysis of raw data, experimentation and survey.</p><blockquote>“A research paper <strong>analyses a perspective </strong>or <strong>argues a point</strong>. Regardless of the type of research paper you are writing, your finished research paper should present your own thinking backed up by others’ ideas and information.”- Suny Empire State College</blockquote><h3>5. Hacks to Present a Paper</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/217/1*8HpZ1Z3z_MDSmcdVTebcpA.png" /></figure><p>Conferences provide approximately 15 minutes in presenting important aspects of your paper . But what we’re addressing deviates from this context in the sense that these were the steps I was asked to follow, by Sanjay Sir, while presenting an existing paper.Once you’ve read through a paper extensively, start by making a presentation of 5 slides. Each slide should cover the following:</p><ol><li>Problem Statement- What is intended? What is the primary focus of the paper?</li><li>Previous Work</li><li>Recent Work</li><li>Why Do I Care?- Mention what interests you</li><li>Critical Analysis and Future Work</li></ol><p>You can view a sample presentation <a href="https://drive.google.com/file/d/0B_swnsCcJGMpbGp1R1B2NmNmUlh0Mi1hOFhsYUx6M0hYQ0Vn/view?usp=sharing">here.</a></p><h3>6. Summing it Up</h3><p>It’s needless to say that everyone follows a different strategy. However, a few guidelines won’t hurt. This one’s helpful for those who are intimidated to kick-start and venture into research; don’t let the bogeyman intimidate you! :) Let’s wrap this up with a simple flow, for newbies (shown below). Ideally, you should have at least three revisions, including one by a professor, team-mate or colleague.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1022/1*MHu5_Wdq_otlT2kKYauPxQ.png" /></figure><p>By now, the Research Bogeyman has fled hopefully :P Don’t hesitate; start with your first research venture- it’s not rocket science; domain depth is the key. <em>Note: The article comprises resources and tips based on my personal experiences. Feel free to recommend resources and provide suggestions as comments :)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f5225fbce982" width="1" height="1" alt=""><hr><p><a href="https://medium.com/acm-w-manipal/yikes-its-the-research-bogeyman-f5225fbce982">Yikes! It’s the Research Bogeyman</a> was originally published in <a href="https://medium.com/acm-w-manipal">ACM-W Manipal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>