<?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[xplore.ai - Medium]]></title>
        <description><![CDATA[xplore.ai — Artificial Intelligence &amp; Experimentation company - Medium]]></description>
        <link>https://medium.com/xplore-ai?source=rss----8b2abc8b59a9---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>xplore.ai - Medium</title>
            <link>https://medium.com/xplore-ai?source=rss----8b2abc8b59a9---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 23:11:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/xplore-ai" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Principles, Vision & Hypothesis in Lean UX]]></title>
            <link>https://medium.com/xplore-ai/principles-vision-hypothesis-in-lean-ux-efe059fdd505?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/efe059fdd505</guid>
            <category><![CDATA[xplore]]></category>
            <category><![CDATA[lean-ux]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Niel S. Flamarich]]></dc:creator>
            <pubDate>Tue, 19 May 2020 07:09:43 GMT</pubDate>
            <atom:updated>2020-05-19T07:43:51.942Z</atom:updated>
            <content:encoded><![CDATA[<h4>How this methodology can solve problems and fuel innovation in your company</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*uMHRSVv7vqnGaKVf2CwSCQ.jpeg" /><figcaption>pexels</figcaption></figure><p>Lean UX was born out from Toyota’s manufacturing model. The innovation not only helped the japanese company improve its relationship with the customers, but it also<strong> improved the production process</strong>. It basically allowed them to have shorter production cycles by adding <strong>feedback from users</strong> on each one of those cycles. This resulted in an improvement of the company’s ability to respond to change.</p><p>While User Experience (UX) design has been around for quite a few years, it’s the advent and mixing of other methodologies what has created a mindset that can help companies improve products and user experience.</p><p>At its core, UX uses Human-Computer interaction design, but it also aims to extend it while embracing all aspects of product or service from the viewpoint of users. Lean UX is the <strong>process of explaining user behaviour with usability, usefulness and desirability in mind</strong>.</p><p>The development of <em>Agile</em> methodologies and <em>Lean Startup</em> make it possible to use a wide variety of tools in product design to improve user experience frameworks. What we basically pursue with Lean UX is blending the interaction design techniques with the scientific method to create beautiful and easy-to-use products, but also products that are <strong>measurably successful</strong>.</p><blockquote><a href="https://www.youtube.com/watch?v=EYPapE-3FRw">“If it disagrees with experiment, it’s wrong.</a>”</blockquote><blockquote>-Richard Feynman</blockquote><h3>What’s Lean UX</h3><p>Lean UX is a management method, but it’s also a process and a mindset. We aim to reduce waste from simple UX design processes, refocus software development to value, harmonize a structure of various roles that are present in all companies and perform an approach to problem solving, but above all, we are interested in <strong>shifting to a model based on experimentation</strong>.</p><p>One of the foundations of Lean UX is design thinking; allowing to use design methods to those unaccustomed to their intricacies, although that’s only a part. We also need to directly observe what people want and need, without forgetting <strong>what is technologically feasible</strong>. Agile methodologies fit perfectly in this mindset, but allowing freedom is also a must. “<strong>Build-measure-learn</strong>” is the motto, with special attention to customer communication (increasing the frequency of contact, for example). Perhaps the most important point to make is that in Lean UX, <strong>each design must be thought of as a hypothesis</strong>.</p><h3>Principles</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7KMV4jH4yQy7KAXfhcZpCw.jpeg" /><figcaption>pexels</figcaption></figure><p>The situation is this; most companies don’t know how to grow. They fail. Some die trying. We see it constantly. This is not a problem of designers, developers or any other specific roles. <strong>The problem is the systems we use to build companies</strong>. The shift needed to embrace Lean UX techniques can be summarized in many ways, but here are a few worth mentioning:</p><h4>1) From output to outcomes</h4><p>There needs to be a stop in how success is usually measured, because it’s meaningless, at least most of the time. Stop thinking about features completed and start thinking in terms of progress towards outcomes. Experiments should be acknowledged at every step, forgetting previous roadmaps. Hypotheses (designs, that is) must be built on three points; <strong>risk, feasibility and potential success</strong>.</p><h4>2) From limited roles to collaborative capabilities</h4><p>Roles in a team don’t need to be discrete. Think of them as a continuous function. <strong>Competencies and skills are above roles</strong>. Silos, whatever their kind, are inefficient. Collaborate with your team, and make sure the team has the tools to collaborate.</p><h4>3) Small, cross-functional teams</h4><p>Interaction amongst product managers, developers, QA engineers, designers or marketers should be on the same level, and no single discipline should dictate the others. <strong>Breaking down discipline-based boundaries creates cohesive teams </strong>and opens creative possibilities. This should be done in small teams working on smaller problems, making the effort more efficient.</p><h4>4) Third party</h4><p>Whether it’s software building or data projects, these enterprises necessitate iterative learning to thrive. That learning process means change. Initial plans will change, often unexpectedly, so <strong>you need to be ready for that change</strong>. Third party solutions can help in this dynamic. Plan for change, and structure third party resources around it.</p><h4>5) Problem solving</h4><p>The Lean UX approach brings a lot of change; <strong>be prepared</strong>. In the process of designing, the ability to think fast is extremely important. Build a shared understanding in the team to minimize execution time.</p><h4>6) Aesthetics</h4><p>Ideas and processes must be edited down to the core. Working quickly creates many artifacts, and these artifacts are a transient part of the project. <strong>Get it done, get it out there, discuss, move on</strong>. The effort at the visual refinement stage should be done at the end, not before.</p><blockquote>“Get it done, get it out there, discuss, move on.”</blockquote><h4>7) Embrace UX debt</h4><p>There needs to be a strong commitment to <strong>continuous improvement</strong>. This UX debt should be treated as technical teams treat technical debt. Capture the gap and analyze where the experience is today and where you want it to be tomorrow.</p><h4>8) Manage up and out</h4><p>Stepping away from the roadmap mindset also means giving freedom to pursue effective solutions, but this <strong>freedom comes with the responsibility of communicating effectively that agenda</strong>. Reach out to members of the team, including those who aren’t involved in your work, make sure they are aware of what’s going on.</p><h3>Vision, Framing, Outcomes</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*Q9yTu6IX57ZN4AT_f2aiyg.jpeg" /></figure><p>While the principles of Lean UX may be thought of differently, the main idea here is to treat hypotheses as a starting point, a way to express assumptions in a testable form. Remember the <strong>goal is not to just create a deliverable, but to create an outcome</strong>.</p><p>In practice, this can be achieved following these steps:</p><ol><li><strong><em>Declare assumptions</em></strong>. Listen to everyone’s ideas, let them be expressed. Then, go through a ruthless priorization.</li><li><strong><em>Hypotheses statements</em></strong>. Break down hypotheses into smaller parts, and test the assumptions. Settle benchmarks and keep in mind; it’s outcomes what we are pursuing.</li><li><strong><em>Personas</em></strong>. Work on simple models representing users, profiles. Build a template and think about who is that person, what will she/he need, what are the pain points, and how can your product help.</li><li><strong><em>Features</em></strong>. Once a list of outcomes is set and the team is focused in a group of users, it’s time to think about the features. Start a brainstorming process.</li></ol><blockquote>“Remember the <strong>goal is not to just create a deliverable, but to create an outcome.”</strong></blockquote><h3>Collaborative Design</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*1ymZ5ilvEEOFo-hK5WEcoA.jpeg" /></figure><p>The design studio phase should encourage informal meetings and sessions. Here, we should break down organizational silos and forum. A way to achieve this is starting meetings with a <strong>problem definition intro</strong>, then individual <strong>idea generation</strong>, presentation and critique where all team members have a say. This should be <strong>constructive</strong>, up to the point of iterating ideas and counter-arguments. Then it’s time for the team idea generation, the fruit of all the effort.</p><blockquote>“This should be constructive, up to the point of iterating ideas and counter-arguments.”</blockquote><h3>MVPs and Experiments</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*MKq7Mp5-xNqz36s89YpK0A.jpeg" /></figure><p>Creating an MVP is the next step. <strong>An MVP is the smallest thing we can create to learn if a hypothesis is valid</strong>. Is there a need for the solution we are designing? Is there value in the features? Is the solution usable? These questions are all that matters in this phase. <strong>But be clear, concise</strong>. Measure behaviour, be functional and integrate with existing analytics. Be consistent with the rest of the application, and build the smallest MVP possible.</p><p>Prototyping, approximating the UX to allow simulating the usage, is next. Specify the intended audience and focus on the core workflows that illustrate the MVP. This can be worked on with clickable demos and previews for teammates.</p><blockquote>“An MVP is the smallest thing we can create to learn if a hypothesis is valid.”</blockquote><h3>Feedback and Research</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*N4mySec_VKqcGWLr4PZ5HQ.jpeg" /></figure><p>The aim of this whole process is to be able to begin a validation process. Collaborative design techniques build shared understanding, but this last step is where all should end up. Informal qualitative research studies are welcome in every iteration, as continuous research is what we’re interested in. Which artifacts to test? What results to expect? <strong>Here’s where A/B testing comes in.</strong> We also need to listen to the voice of the customer.</p><p>Build <strong>research activities into every sprint</strong>, and don’t rely only on specialized work to deliver learning. Research activities and responsibilities should at this point be shared across the entire team. Remember this is a <strong>team effort to test ideas in the market</strong>. Review and decide as a team. Mix smaller teams’ roles into pairs and make them work on a version of the MVP. Conduct interviews and change roles, it’s surprising how many new ideas can come up by embracing this dynamic.</p><blockquote>“The aim of this whole process is to be able to begin a validation process. Collaborative design techniques build shared understanding, but this last step is where all should end up.”</blockquote><p>Lastly, you need to make sense of the research. Look for patterns, verify with other sources and test everything.</p><p>___________________________________________________________________</p><p>For those of us steeped in traditional management techniques, this new way of working can seem a little disorienting, but what we mainly intend with these summarized techniques is to <strong>learn as soon as possible what’s needed to improve user experience</strong>. For this to work, we need to be close to our potential customers, learn from them and improve the product. Some of the great advantages of Lean UX include minimizing the risk of the projects, being certain of their duration, building up cohesive teams and <strong>measure their success objectively</strong>.</p><p>Sources:</p><ul><li><em>Applying Lean Principles to Improve User Experience</em>, by Jeff Gothelf and Josh Seiden</li><li><em>The Design of Everyday Things</em>, by Donald A. Norman</li><li><em>Sketching User Experiences: Getting the Design Right and the Right Design, </em>by Bill Buxton</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=efe059fdd505" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/principles-vision-hypothesis-in-lean-ux-efe059fdd505">Principles, Vision &amp; Hypothesis in Lean UX</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 steps to AWS Machine Learning Specialty Certification, Made Easy]]></title>
            <link>https://medium.com/xplore-ai/aws-machine-learning-specialty-certification-made-easy-fe72a7a3095a?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/fe72a7a3095a</guid>
            <category><![CDATA[xploreai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[aws]]></category>
            <dc:creator><![CDATA[Niel S. Flamarich]]></dc:creator>
            <pubDate>Mon, 04 May 2020 07:10:54 GMT</pubDate>
            <atom:updated>2020-05-04T09:22:41.285Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How to prepare for the exam in 5 steps and what not to do.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KcU0bnB7JjUPwUlLLgyUoQ.jpeg" /></figure><p>The <strong>AWS Machine Learning certification differs a bit from all the other certifications</strong>; you need to have a good grasp on all things Amazon Web Services, but also understand the basics of Machine Learning and Deep Learning. This may not be an easy task depending on your background.</p><p>Many different profiles try this exam, <strong>from engineers to cloud experts, but also data scientists and analysts</strong>. Of course, the ideal in terms of having a good chance of passing the exam would be to be an experienced jack-of-all-trades that can feel confortable in all those topics. <strong>This is rarely the case</strong>, as people usually have strong knowledge of their field, but weaker knowledge when it comes to things they haven’t studied or worked with. <em>We all know unicorns don’t exist.</em></p><p>To help with this situation, this article will show you:</p><ul><li>Even without experience in all the topics, you can catch up.</li><li>Resources exist, but you’ll also need to play around with AWS services on your own.</li><li>Nothing beats first hand experience when it comes to understand AWS.</li><li>Time management in the exam is extremely important.</li><li>Total ignorance or insufficient knowledge about Machine Learning and Deep Learning is a ticket to failure.</li></ul><h3>Strategy</h3><p>Summarized, here are the 5 points that helped me pass. We’ll go through them below:</p><p>1- <strong><em>Read the whitepapers</em></strong><em>, especially Sagemaker’s, but ideally all those from the main AWS services.</em></p><p>2- <strong><em>Try as many AWS services as you can</em></strong><em> if you haven’t worked with them in the past. Play with them, feel comfortable.</em></p><p>3- <em>Forget AWS official content outline percentages; keep in mind it’s </em><strong><em>50% ML/DL, 25% Sagemaker and 25% other AWS services</em></strong><em>.</em></p><p>4- <em>Don’t get stuck and/or think about how much time you have left; </em><strong><em>180 minutes is enough, but use that time wisely</em></strong><em>.</em></p><p>5- <strong><em>Don’t be fooled by the exam’s constant tricks</em></strong><em> that try to confuse you. Once you read a difficult question properly, you’ll notice it wasn’t that hard.</em></p><h3><strong>1- Read the Whitepapers</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XdZ9FZ7BxIvX098J1tB3YQ.jpeg" /></figure><p>Although you can find online resources and paid courses, the best way to understand AWS is to read the <a href="https://docs.aws.amazon.com/">documentation of the services</a>. Yes, it is a lot. But as you go through them, you’ll be able to connect the dots and acquire a general and much needed understanding. Amazon’s free training courses can help you too, but some of them are just simple tutorials that won’t give you a deep knowledge of the subject (they’re usually internal presentations and formations made public). <strong>The most important service is Sagemaker</strong>, and that’s where your effort should be. Pivot from there.</p><h3><strong>2- Go play with AWS</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*1vg-dTcGerHuBjbqqZFFqA.jpeg" /></figure><p>If you don’t have it already, <strong>open an AWS account and start using its Machine Learning services</strong>. The <a href="https://aws.amazon.com/free/?all-free-tier.sort-by=item.additionalFields.SortRank&amp;all-free-tier.sort-order=asc">Free Tier</a> offers hours to compute, and lets you use more than 60 services for free for 12 months. There are great demos in there, and free of charge tools that will help you understand usage and case scenarios and to try pre-trained algorithms by yourself. This is priceless.</p><h3><strong>3- Content Outline Percentages</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BJUbi5K2u5WyPPYuS26v5g.jpeg" /></figure><p>Amazon provides a list of domains you’ll need to know, from <em>Data Engineering</em> and <em>ML Implementations and Operations</em> (each representing 20% of the exam), <em>Exploratory Data Analysis</em> (24%) and <em>Modelling</em> (36%). My view after passing the exam was that they didn’t make much sense. A more real partition would be this:</p><p><strong><em>Machine Learning and Deep Learning</em></strong><em> (50% of the exam)</em></p><p>You should be comfortable with the whole ML cycle; from data collection and data preparation to exploratory data analysis and modelling. <strong>A good grasp on how to formulate problems and measure success is paramount</strong>. Which metrics to use play an important role in the exam. What types of data (structured, unstructured) exist and what makes good data in a machine learning project.</p><p>Data preparation is a key step in data science, and you should expect a few questions on missing value handling, categorical encoding, imputation and other feature engineering steps. Pay attention to scaling (<em>normalization</em> and <em>standardization</em>) for numerical data, <em>n-grams</em> and <em>bag-of-words</em> for text data, as well as <em>Term Frequency — Inverse Document Frequency (tf-idf)</em>, as you may have to solve simple corpus data problems. Also, keep in mind how and why <em>recordIO protobuf</em> format will work better in AWS jobs.</p><p>Some questions may present you visualizations of data and ask you something about that data. This is clearly an attempt to test your skills in data analysis. <strong>You should be comfortable knowing the different types of graphs and their aim (comparison, composition, relationships and distributions)</strong>. Be aware of how they are used and how they can help in exploratory data analysis.</p><p>Model design is obviously present in the exam too. How to select a good model, what ML approach is better in a situation (regressions, classifications) and what metrics and strategies should be used. About the algorithms, <strong>you should be comfortable with all the classics</strong> (K-Means and its differences with K-Nearest Neighbours, Random Forests and Decision Trees) as well as with <strong>Convolutional Neural Networks</strong>.</p><p>You’ll need to <strong>understand the concepts of training, test and validation data</strong>, identify potential biases introduced by insufficient splits, and additional measures that could be used to increase data value. What kind of generalization are we looking for in a ML process, how it will be consumed (real time, batch, API applications), and <strong>how to tell if the generalization is working (accuracy tests) is an all-present topic throughout the exam</strong>. You’ll need to be able to explain how Confusion Matrices, Recall, Precision and False Positive Rate work.</p><p>About evaluating models, you’ll need to understand offline and online validation, and at least conceptually know about canary deployments. <strong>Underfitting and overfitting and how to overcome them</strong>, regression accuracy (RMSE), histograms and skewness in them, AUC metrics in classification, and all the trade-offs in evaluation that might call for different optimizations are all very important in the exam. Also, <strong>model tuning specifics</strong> and at least some grasp in <strong>Bayesian optimization</strong> will help a lot.</p><p><strong><em>Sagemaker</em></strong><em> (25% of the exam)</em></p><p>While knowledge of Machine Learning is expected from you in the exam, <strong>the star of the game is Sagemaker, AWS’ central and pivotal service for all things AI</strong>. Sagemaker helps you through the whole ML cycle from start to finish, and you’ll probably have to explain how AWS services interact with it. Keep in mind that it will appear in questions quite often, and you need to know its intricacies when it comes to training models (how to create a training job API, how to set up an Elastic Container Repository, inferences).</p><p>Another important thing to know are <strong>Sagemaker built-in algorithms</strong>. From Linear Learners to Factorization Machines, Image Analysis and Anomaly Detection, and differences amongst text analysis algorithms (managing and explaining when to use <em>LDA</em>, <em>Neural Topic Modeling</em>, <em>Seq2Seq</em> and <em>Blazing Text</em> can and probably will appear often). As a widely used algorithm both in regression and classification problems, <em>XGBoost</em> shows up too, as in comparisons with other algorithms. Keep in mind that <strong>knowing the main built-in algorithms hyperparameters and metrics is a must</strong>.</p><p>Last but not least, knowing <strong>how to take Sagemaker models to production</strong> and how Jupyter notebooks work is needed. Sagemaker hosting services (how to create endpoint configurations) and inference pipelines to chain together algorithms, docker containers and elastic inference can pop up in the questions. You should also know how automatic scaling works.</p><p><strong><em>Other AWS services</em></strong><em> (25% of the exam)</em></p><p>Here’s where the exam gets similar to other AWS certifications. If you already have a solutions architect or big data certification (I didn’t), all this will surely be familiar. The most important services here are <strong>AWS Glue and Athena</strong> (for ETL jobs), the <strong>Kinesis family</strong> (for streaming data), <strong>S3</strong>, <strong>RDS</strong>, <strong>DynamoDB</strong> and <strong>Redshift</strong> (as data stores) and the Hadooq cluster ecosystem <strong>Elastic Map Reduce</strong> (EMR). You should understand in depth how they all work, what are their strengths and why should be used in specific situations. The white papers for these services are quite good, but ideally you should have some experience working with them. If that’s your case, you are covered.</p><p>Some questions may make you choose the correct order of services to use in a given business situation. You will need to know each service, <strong>its purpose and use-cases that will be presented to you in the questions</strong>. Keep in mind all monitoring and evaluation tools for deployments AWS has to offer. I’d recommend making an effort in understanding <strong>CloudWatch</strong> and <strong>CloudTrail</strong> in detail if you’ve never used them before.</p><p>There’s a few AI developer services that you must know, but the good news is that they’re quite simple: <em>Forecast</em>, <em>Lex</em>, <em>Personalize</em>, <em>Polly</em>, <em>Rekognition</em>, <em>Transcribe</em> and <em>Translate</em> are all easy-to-use services available in the console that you just need to try yourself to understand.</p><p>For deployments outside Sagemaker, AWS offers some things you should check out: Elastic Container Service, EC2 AMIs, Elastic Map Reduce and on-premises options (MXNet and TensorFlow frameworks).</p><h3>4- Time</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z0U86_6ptXmlul6bysZjfg.jpeg" /></figure><p>While you’ll have a bit less than three minutes per question (which is more than enough), I find the best strategy is to <strong>answer first all those questions you know</strong>, -or have little doubt about- and leave the harder ones for later.</p><p>This means going through the questions once and mark the ones that make you doubt. Then, dedicate the bulk of your time to hard questions. If you do this right you may even have the last 20–30 minutes of the exam to make a last check and get an intuition of how you just did.</p><h3>5- Exam Logic</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ki5B_gxNIMi9zSwR71MABA.jpeg" /></figure><p>You’ll notice some verbose questions in the exam, but after reading them, you’ll see they are just trying to confuse you. <strong>Read the answers carefully. Even before the questions themselves</strong>. They usually point to the service or solution better than the question<em>.</em></p><p>This is an important point: the exam will try to trick you, almost constantly. It’ll try to find where your knowledge is thinner with very similar answers that may superficially look the same. <strong>If a question seems complicated, leave it for later</strong>. The worst thing you can do is to get stuck in a question and spend 10 minutes there, because you’ll probably need them at the end of the test.</p><p>There’s also a chance that you’ll need to do some calculations. Ask for paper and a pen, or whatever option is allowed, in your examination center. This may seem silly, but having the opportunity to write down doubts and scribble simple math could be a game changer in the exam, and who doesn’t want all the odds in her/his favour?</p><p>Best of luck!</p><p><em>Original article </em><a href="https://medium.com/@danielsflamarich/a-different-take-on-aws-machine-learning-specialty-certification-a1bea3ee9ec9"><em>here</em></a><em>.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/727/0*F75lZoVOHV3nB27N" /><figcaption>Daniel S Flamarich is a AWS certified Machine Learning Specialist</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe72a7a3095a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/aws-machine-learning-specialty-certification-made-easy-fe72a7a3095a">5 steps to AWS Machine Learning Specialty Certification, Made Easy</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Implementing RoI Pooling in TensorFlow + Keras]]></title>
            <link>https://medium.com/xplore-ai/implementing-attention-in-tensorflow-keras-using-roi-pooling-992508b6592b?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/992508b6592b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[keras]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Jaime Sevilla]]></dc:creator>
            <pubDate>Mon, 01 Apr 2019 18:15:39 GMT</pubDate>
            <atom:updated>2019-09-04T13:12:56.660Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<p><em>by </em><a href="https://twitter.com/jsevillamol"><em>Jaime Sevilla</em></a><em> @xplore.ai</em></p><p>In this post we explain the basic concept and general usage of <strong>RoI (Region of Interest) pooling</strong> and provide an implementation using Keras layers and the TensorFlow backend.</p><p>The intended audience for this post are people familiar with the <a href="https://course.fast.ai/">basic theory of (Convolutional) Neural Networks</a> and who are capable of building and <a href="https://keras.io/#getting-started-30-seconds-to-keras">running simple models using Keras</a>.</p><p>If you are here just for the code, serve yourself from <a href="https://gist.github.com/Jsevillamol/0daac5a6001843942f91f2a3daea27a7">this gist</a> and do not forget to like and share the article!</p><h4>Understanding RoI Pooling</h4><p>RoI Pooling was proposed by Ross Girshick in the <a href="https://arxiv.org/abs/1504.08083">Fast R-CNN paper</a> as part of his object recognition pipeline.</p><p>In the general use case for RoI Pooling we have an image-like object, and multiple regions of interest specified via bounding boxes. We want to generate an embedding out of each RoI.</p><p>For example, in the R-CNN set up we have an image and a proposal mechanism that produces bounding boxes for potentially interesting parts of the image. Now, we want to embed each suggested patch of the image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*A2IfHwmF8MT_kmwE4DmnDA.jpeg" /></figure><p>Simply cropping each suggested region will not work, because we want to stack the resulting embeddings on top of each other, and the suggested regions do not necessarily have the same shape!</p><p>So we need to come up with a way of transforming each patch in a way that results in a embedding of a predefined shape. How can we do that?</p><p>A standard way in Computer Vision of reducing the shape of images is using some sort of pooling operation.</p><p>The most common one is <strong>max pooling</strong>, where we divide the input image in (usually non-overlapping) areas of equal shape, and form the output by taking the maximum value found in each area.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*krZsn7QzWQ9Oux6MraFjAw.jpeg" /><figcaption>The maxpool operation divides each region into pooling areas of the same size</figcaption></figure><p>This does not directly solve the problem we have — patches of different shapes will be divided into a variable number of areas of equal shape, producing an output of variable shape.</p><p>But this inspires an idea. What if we divide each region of interest into an equal number of areas of variable different shape, and take the maximum of each of them?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*IPoKQl9x7qcAOwX2Q0UJLg.jpeg" /><figcaption>The ROI Pooling operation divides all regions into equivalent grids of pooling areas</figcaption></figure><p>And that is exactly what the ROI Pooling layer does.</p><h4>Use cases for RoI pooling</h4><p>RoI pooling is a fairly general tool. It usually shows up in conjunction with RoI proposal mechanisms, bridging the gap between region proposals and their embedding. We will see two concrete examples to illustrate its potential.</p><p>Firstly, in the context it was developed (object recognition) it allows us to split the task pipeline into two parts (region proposal and classification of regions) while retaining an end-to-end single-pass differentiable architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zw4YAo0OVFQdU_Fux-03-Q.png" /><figcaption><a href="https://arxiv.org/abs/1504.08083">Fast R-CNN</a> architecture showcasing RoI pooling, by Ross Girshick</figcaption></figure><p>Thus in the R-CNN model we first have a component in the model that proposes a fixed number of RoIs. RoI pooling allows us to run a CNN classifier on all the proposed RoIs and pick as our output the one with the highest target class probability.</p><p>Secondly, together with region proposals, RoI Pooling can also be used to implement <strong>visual attention</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Rtz-VF_G9TK47upKOjSJSQ.png" /><figcaption><a href="https://arxiv.org/pdf/1702.01478.pdf">Attentional Network for Visual Object Detection</a> showcasing RoI pooling, by Hara <em>et al</em></figcaption></figure><p>As an example in <a href="https://arxiv.org/pdf/1702.01478.pdf">Attentional Network for Visual Object Detection</a> we see how Hara <em>et al. </em>implement attention using iterative RoI proposals and RoI pooling. They first generate a first proposal (t=1) which the RoI pooling adapts to the Fully Connected layers hereunder. The result is used as an input to the Glimpse component (t=2) to generate a new region proposal that is again embedded using RoI pooling. The process is repeated T times.</p><h4>Typing it</h4><p>Before we dive into the implementation let’s pause for a minute to think about the type signature of the ROI layer.</p><p>It takes in two tensors:</p><ul><li>A batch of <strong>images</strong>. In order to be able to process them together, all images must have the same shape. The resulting shape of the tensor will be <em>(batch_size, img_width, img_height, n_channels)</em>.</li><li>A batch of <strong>Region Of Interest (ROI) proposals</strong>. If we want to stack them together in a tensor, the number of proposed regions must be fixed for each image. Since each bounding box must be specified with 4 coordinates, the shape of this tensor will be <em>(batch_size, n_rois, 4)</em>.</li></ul><p>And it must output:</p><ul><li>A list of <strong>embeddings</strong> for each image, codifying the regions specified by each ROI. The corresponding shape should be <em>(batch_size, n_rois, pooled_width, pooled_height, n_channels)</em>.</li></ul><h4>The code in Keras</h4><p>Keras allows us to easily implement custom layers via inheritance of the base Layer class.</p><p>The <a href="https://www.tensorflow.org/api_docs/python/tf/keras/layers/Layer">tf.keras documentation</a> recommends implementing the <em>__init__</em>, <em>build</em> and <em>call</em> methods for our custom layer. However, since the purpose of the <em>build</em> function is to add the weights of the layer and our layer has no weights, we do not need to override that method. We will also implement the convenience method <em>compute_output_shape</em>.</p><p>We will code each part separately and then put all the pieces together in the end.</p><pre>def __init__(self, pooled_height, pooled_width, **kwargs):<br>    self.pooled_height = pooled_height<br>    self.pooled_width = pooled_width<br>    super(ROIPoolingLayer, self).__init__(**kwargs)</pre><p>The <em>constructor</em> of the class is quite simple to understand. We need to specify the target height and width of the embeddings we are producing. In the last line of the constructor we call the parent constructor to initialize the rest of the class attributes.</p><pre>def compute_output_shape(self, input_shape):<br>    &quot;&quot;&quot; Returns the shape of the ROI Layer output<br>    &quot;&quot;&quot;<br>    feature_map_shape, rois_shape = input_shape<br>    assert feature_map_shape[0] == rois_shape[0]<br>    batch_size = feature_map_shape[0]<br>    n_rois = rois_shape[1]<br>    n_channels = feature_map_shape[3]<br>    return (batch_size, n_rois, self.pooled_height, <br>            self.pooled_width, n_channels)</pre><p><em>compute_output_shape </em>is just a nice utility function that will tell us what the output of the layer will be for a particular input.</p><p>Next we have to implement <em>call</em>. The <em>call</em> function is where the logic of the layer lives. This function should take as input the two tensors that hold the input to the ROI Pooling Layer, and should output the tensor with the embeddings.</p><p>Before we implement that, we need to implement a simpler function that will take a single image and a single ROI and return the corresponding embedding.</p><p>Let’s do that step by step.</p><pre><a href="http://twitter.com/staticmethod">@staticmethod</a><br>def _pool_roi(feature_map, roi, pooled_height, pooled_width):<br>  &quot;&quot;&quot; Applies ROI Pooling to a single image and a single ROI<br>  &quot;&quot;&quot;</pre><pre># Compute the region of interest        <br>  feature_map_height = int(feature_map.shape[0])<br>  feature_map_width  = int(feature_map.shape[1])<br>        <br>  h_start = tf.cast(feature_map_height * roi[0], &#39;int32&#39;)<br>  w_start = tf.cast(feature_map_width  * roi[1], &#39;int32&#39;)<br>  h_end   = tf.cast(feature_map_height * roi[2], &#39;int32&#39;)<br>  w_end   = tf.cast(feature_map_width  * roi[3], &#39;int32&#39;)<br>        <br>  region = feature_map[h_start:h_end, w_start:w_end, :]</pre><pre>...</pre><p>The first six lines of the function are computing where the region of interest starts and finishes within the image.</p><p>We have chosen as a convention that the coordinates of each ROI should be specified in relative terms, as numbers between 0 and 1. Concretely, each ROI is specified by a 4-dimensional tensor containing four relative coordinates <em>(x_min, y_min, x_max, y_max)</em>.</p><p>We could have decided to identify it in absolute terms, but this is generally worse, since a common pattern is to pass the input image through some convolutions that change the shape of the image before feeding it to the ROI Pooling layer, which would force us to keep track of how the image shape changes to scale the ROI bounding boxes properly.</p><p>The seventh line just crops the image down to the region of interest using the super powerful tensor slicing syntax TensorFlow provides us.</p><pre>...</pre><pre># Divide the region into non overlapping areas<br>region_height = h_end - h_start<br>region_width  = w_end - w_start<br>h_step = tf.cast(region_height / pooled_height, &#39;int32&#39;)<br>w_step = tf.cast(region_width  / pooled_width , &#39;int32&#39;)<br>        <br>areas = [[(<br>           i*h_step, <br>           j*w_step, <br>           (i+1)*h_step if i+1 &lt; pooled_height else region_height, <br>           (j+1)*w_step if j+1 &lt; pooled_width else region_width<br>          ) <br>          for j in range(pooled_width)] <br>         for i in range(pooled_height)]</pre><pre>...</pre><p>In the next four lines we compute the shape of each area within the ROI that is going to be pooled.</p><p>Afterwards we create a 2D array of tensors where each component is a tuple indicating the coordinates of the start and end of each of the areas we are going to take a maximum.</p><p>The code that generates the grid of area coordinates seems overly complicated, but notice that if we just divide the region of interest into areas of shape <em>(region_height // pooled_height, region_width // pooled_width)</em> there would be some pixels of the ROI that would not fall within any area.</p><p>We fix that by extending the right and bottom most areas to encompass the remaining pixels that would not fall in any area by default. This is realized in the code via conditional declarations of the max coordinates of each bounding box.</p><p>The result is a 2D list of bounding boxes, and we move on to the next part.</p><pre>...</pre><pre># Take the maximum of each area and stack the result<br>def pool_area(x): <br>  return tf.math.reduce_max(region[x[0]:x[2],x[1]:x[3],:], axis=[0,1])<br>        <br>pooled_features = tf.stack([[pool_area(x) for x in row] for row in areas])</pre><pre>return pooled_features</pre><p>These lines above do all the magic. We define an auxiliary function <em>pool_area</em> that takes as input a bounding box specified by a tuples such as the ones we just created, and outputs the maximum of each channel in the area.</p><p>We map <em>pool_area</em> over every area we have declared using a list comprehension.</p><p>By this point we return a tensor of shape <em>(pooled_height, pooled_width, n_channels)</em>, holding the result of pooling one region of interest of a single image.</p><p>Next step is pooling many ROIs from a single image. This is straightforward to implement using an auxiliary function and <em>tf.map_fn</em> to produce a tensor of shape <em>(n_rois, pooled_height, pooled_width, n_channels)</em>.</p><pre><a href="http://twitter.com/staticmethod">@staticmethod</a><br>def _pool_rois(feature_map, rois, pooled_height, pooled_width):<br>  &quot;&quot;&quot; Applies ROI pooling for a single image and varios ROIs<br>  &quot;&quot;&quot;<br>  def curried_pool_roi(roi): <br>    return ROIPoolingLayer._pool_roi(feature_map, roi, <br>                                     pooled_height, pooled_width)<br>        <br>  pooled_areas = tf.map_fn(curried_pool_roi, rois, dtype=tf.float32)<br>  return pooled_areas</pre><p>Lastly we need to implement the batch level iteration. If we pass to <em>tf.map_fn </em>a sequence of tensors (such as our input <em>x</em>) it will take care of zipping it under the hood for us.</p><pre>def call(self, x):<br>  &quot;&quot;&quot; Maps the input tensor of the ROI layer to its output<br>  &quot;&quot;&quot;<br>  def curried_pool_rois(x): <br>    return ROIPoolingLayer._pool_rois(x[0], x[1], <br>                                      self.pooled_height, <br>                                      self.pooled_width)<br>        <br>  pooled_areas = tf.map_fn(curried_pool_rois, x, dtype=tf.float32)<br>  return pooled_areas</pre><p>Notice that we must specify the <em>dtype</em> parameter of <em>tf.map_fn</em> every time its expected output does not match the data type of its input. In general it is good practice to specify it as often as possible to make explicit how the types are changing through our Tensorflow computation graph.</p><p>Let’s put everything together:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a674093f5564aaae741c4f7d2065e981/href">https://medium.com/media/a674093f5564aaae741c4f7d2065e981/href</a></iframe><p>Let’s test our implementation! We are going to suppose a 100x200 1-channel image and we will extract 2 RoIs using 7x3 pooling patches. Images can have regions classified in a maximum of 4 labels. Example feature map is all 1s but a single value of 50 placed at (<em>height-1, width-3</em>).</p><pre>import numpy as np</pre><pre># Define parameters<br>batch_size = 1<br>img_height = 200<br>img_width = 100<br>n_channels = 1<br>n_rois = 2<br>pooled_height = 3<br>pooled_width = 7</pre><pre># Create feature map input<br>feature_maps_shape = (batch_size, img_height, img_width, n_channels)<br>feature_maps_tf = tf.placeholder(tf.float32, shape=feature_maps_shape)<br>feature_maps_np = np.ones(feature_maps_tf.shape, dtype=&#39;float32&#39;)<br>feature_maps_np[0, img_height-1, img_width-3, 0] = 50<br>print(f&quot;feature_maps_np.shape = {feature_maps_np.shape}&quot;)</pre><pre># Create batch size<br>roiss_tf = tf.placeholder(tf.float32, shape=(batch_size, n_rois, 4))<br>roiss_np = np.asarray([[[0.5,0.2,0.7,0.4], [0.0,0.0,1.0,1.0]]], dtype=&#39;float32&#39;)<br>print(f&quot;roiss_np.shape = {roiss_np.shape}&quot;)</pre><pre># Create layer<br>roi_layer = ROIPoolingLayer(pooled_height, pooled_width)<br>pooled_features = roi_layer([feature_maps_tf, roiss_tf])<br>print(f&quot;output shape of layer call = {pooled_features.shape}&quot;)</pre><pre># Run tensorflow session<br>with tf.Session() as session:<br>    result = session.run(pooled_features, <br>                         feed_dict={feature_maps_tf:feature_maps_np,  <br>                                    roiss_tf:roiss_np})<br>    <br>print(f&quot;result.shape = {result.shape}&quot;)<br>print(f&quot;first  roi embedding=\n{result[0,0,:,:,0]}&quot;)<br>print(f&quot;second roi embedding=\n{result[0,1,:,:,0]}&quot;)</pre><p>The lines above define a test input for the layer, build the corresponding tensors and run a TensorFlow session so we can check its output.</p><p>Running the script will result in the following output:</p><pre>feature_maps_np.shape = (1, 200, 100, 1)<br>roiss_np.shape = (1, 2, 4)<br>output shape of layer call = (1, 2, 3, 7, 1)<br>result.shape = (1, 2, 3, 7, 1)<br>first  roi embedding=<br>[[1. 1. 1. 1. 1. 1. 1.]<br> [1. 1. 1. 1. 1. 1. 1.]<br> [1. 1. 1. 1. 1. 1. 1.]]<br>second roi embedding=<br>[[ 1.  1.  1.  1.  1.  1.  1.]<br> [ 1.  1.  1.  1.  1.  1.  1.]<br> [ 1.  1.  1.  1.  1.  1. 50.]]</pre><p>We can check that the tensor shapes match our expected results. The resulting embeddings are all 1s except in the area where we introduced a 50 feature.</p><p>It seems to be working!</p><h4>Conclusion</h4><p>And that is it for today folks!</p><p>Today we have learned what the ROI Pooling Layer does, and how we can use it to implement attention. Furthermore we have learned how to extend Keras to implement custom layers without weights, and gave an implementation of the aforementioned ROI Pooling Layer.</p><p>I hope this was useful for you, do not forget to share the article and leave a comment if it was!</p><p><em>Thank you to Ari Brill, Tjark Miener and Bryan Kim for feedback on the article.</em></p><h4>References</h4><p>Ross Girshick. <a href="https://arxiv.org/abs/1504.08083">Fast R-CNN</a>. Proceedings of the IEEE International Conference on Computer Vision. 2015.</p><p>Kota Hara, Ming-Yu Liu, Oncel Tuzel, Amir-massoud Farahmand. <a href="https://arxiv.org/pdf/1702.01478.pdf">Attentional Network for Visual Object Detection</a>. 2017.</p><p>Check our services and explorations with Deep Learning, Machine Learning, Computer Vision and GANs on our <a href="https://www.linkedin.com/company/xplore-ai/">LinkedIn page</a>, <a href="https://twitter.com/XploreAi">Twitter</a>, <a href="https://www.instagram.com/xplore.ai">Instagram</a> and don’t forget to follow us on <a href="https://medium.com/xplore-ai">Medium</a> to not miss any post like this in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=992508b6592b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/implementing-attention-in-tensorflow-keras-using-roi-pooling-992508b6592b">Implementing RoI Pooling in TensorFlow + Keras</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nvidia eGPU + MacOS+ TensorFlow-GPU? The Definitive Setup Guide to Avoid Headaches]]></title>
            <link>https://medium.com/xplore-ai/nvidia-egpu-macos-tensorflow-gpu-the-definitive-setup-guide-to-avoid-headaches-f40e831f26ea?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/f40e831f26ea</guid>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[external-gpu]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Uxío Piñeiro]]></dc:creator>
            <pubDate>Wed, 30 Jan 2019 15:31:30 GMT</pubDate>
            <atom:updated>2019-09-04T13:09:50.668Z</atom:updated>
            <content:encoded><![CDATA[<h3>Nvidia eGPU + MacOS + TensorFlow-GPU? The Definitive Setup Guide to Avoid Headaches</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WEfCQtWK33BeJjiMZ7A9OA.jpeg" /><figcaption>One of our eGPUs with a Nvidia GTX 1080ti mounted</figcaption></figure><p>We have already suffered it, let us save you a couple of days of desperation.</p><blockquote>Tim Cook and Nvidia are going to go to hell for this</blockquote><blockquote>— Myself after a lot of hours of fighting</blockquote><p>At <a href="https://medium.com/u/587193c3b950">xplore.ai</a> we are building Artificial Intelligence products using tailor-made algorithms. We create Deep Learning solutions to problems, mainly in the area of Computer Vision, for clients that hold large amounts of data and need a level of custom solutions further than what they can be provided from using already existing APIs. Our team’s main choice of local development hardware is Apple MacBooks, thus we mainly work on MacOS. Although we started training our Deep Learning models on the cloud (AWS), it became more and more expensive when our level of innovation increased. So we decided to build up our own high-end desktop machine, we call it <em>Franky</em>, where we would train large architectures or big Machine Learning models.</p><blockquote>Training a GAN (Generative Adversarial Network) with a training set of 100.000 images with a single GPU can take various weeks</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-nY8huClbx04emJTw7ETBw.jpeg" /><figcaption>Franky: i7-7700K, 64Gb RAM, RTX 2080Ti</figcaption></figure><p>The main problem appeared when we started growing as a company, only one model could be train at once so there is now a queue of teammates waiting to train their own models for their own tasks. So we decided to investigate the possibility of providing a GPU to each data team member so they could use it with their own laptop. A GPU is normally attached to a motherboard, which is capable of handling such a monster (PCIe), which is inside a big case, which is not a laptop… but… <strong>welcome to the external GPUs world</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*15Nw8sbG1SAYQpBtiwffwQ.png" /><figcaption>Thunderbolt 3: Taking USB-C ports to the next level</figcaption></figure><p>Taking into account that we already had available high-end GPUs our best option was <a href="https://www.razer.com/gaming-laptops/razer-core-x">Razer Core X</a>, but you can select yours with the help of the great <a href="https://egpu.io/">eGPU community</a>. I’m not going to lie, we chose Core X because its powerful specs and, of course, for its slickly design. So far so good, it had not disappointed us, it has demonstrated to be a high-end tool and amazingly silent.</p><p><strong><em>Note</em></strong><em>: the included Thunderbolt 3 cables with the Razer Core X are quite short (0.5m). Consider this and the fact that buying a 1m version can cost easily 40eur and we couldn’t find a 2m version suitable for our scenario.</em></p><p>So far everything looks fine: buy the case online, have it delivered in a few days (<em>how nice it is to live in this world of popularized e-commerce</em>), easily plug your graphics card and we are done 😊. Not so fast… this is only the beginning of our odyssey 😦.</p><p><strong><em>Note:</em></strong><em> We work with CUDA software so AMD wasn’t an option. If you’re a lucky AMD user, you can close the post here and enjoy your new GPU, lucky you!</em></p><p>I’m sure you have heard about the <a href="https://appleinsider.com/articles/19/01/18/apples-management-doesnt-want-nvidia-support-in-macos-and-thats-a-bad-sign-for-the-mac-pro">ridiculous war between Nvidia and Apple</a> which leaves us, the consumers, totally stranded. Nvidia no longer distributes official drivers to Apple MacOS is no longer officially compatible with Nvidia GPUs. So if you like to be up to date with the latest Apple software prepare to say good bye to <a href="https://egpu.io/state-of-egpu-for-macs-mojave-10-14-update/">Mojave</a> and its very high-end dark mode technology because, at least, by the time I’m writing this, we need to downgrade to High Sierra.</p><p>Once we have our clean MacOS 10.13 ready to be eGPU boosted we need to start tricking the system, again the <a href="https://egpu.io/purge-wrangler.sh">eGPU community</a> will save our backs. You can follow the official Purge-Wrangler.sh guide <a href="https://github.com/mayankk2308/purge-wrangler">here</a> or continue to read because I’m going go through all the process as friendly as I can.</p><h3><strong>Installing eGPU on MacOS</strong></h3><h4><strong>1. Disable SIP</strong></h4><p>Reboot the system into Recovery Mode (<strong>⌘+R</strong> during boot), then in the upper bar open <strong>Utilities &gt; Terminal</strong> and:</p><pre>csrutil disable</pre><p>Reboot again, this time normally.</p><h4><strong>2. Install Purge-Wrangler.sh</strong></h4><p>With the eGPU plugged-in before login, open Terminal App (iTerm not supported):</p><pre>curl -s &quot;https://api.github.com/repos/mayankk2308/purge-wrangler/releases/latest&quot; | grep &#39;&quot;browser_download_url&quot;:&#39; | sed -E &#39;s/.*&quot;([^&quot;]+)&quot;.*/\1/&#39; | xargs curl -L -s -0 &gt; purge-wrangler.sh &amp;&amp; chmod +x purge-wrangler.sh &amp;&amp; ./purge-wrangler.sh &amp;&amp; rm purge-wrangler.sh</pre><p>For future uses:</p><pre>purge-wrangler</pre><p><strong>NOTE: </strong>Script will need permissions, don’t panic, this thing is probed. It shouldn’t break anything, but shit happens, it’s not bad practice to do a backup.</p><p>Very simple GUI will display in the terminal, select the setup you need (Nvidia eGPU installation is 2) and reboot after the script had done its things.</p><h4><strong>eGPU Plug — Unplug recommendations</strong></h4><p>Until MacOS natively supports Hot-Unplug you should unplug your eGPU after shutdown or while Mac is off during reboot. Theoretically you can plug your eGPU at any moment but is recommended to do it before login.</p><p>First test you could do, go to <strong>About This Mac </strong>from the Apple menu () in the upper-left corner of your screen. You should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/582/1*EqZkrBhVfRbwOtY_8q291g.png" /><figcaption>See eGPU Detected Correctly</figcaption></figure><p>At this point very likely you are in one of this two situations:</p><ol><li>Everything went well and you’re already enjoying the power of your new eGPU boosted Mac. Continue.</li><li>Everything went well but you are experimenting a bit of lag on the MacOS UI when eGPU acceleration is in use, try using <a href="https://github.com/Eun/DisableMonitor">DisableMonitor</a> to disable the built-in retina display, not fancy but it worked for me. If you find a better solution please leave a comment with it.</li><li>Something went wrong. Prepare yourself to go on a wonderful journy across the <a href="https://egpu.io/forums/mac-setup/guide-troubleshooting-egpus-on-macos/">forums</a> to find your problem and a solution that solves your problem.</li></ol><p>Once you have it I personally recommend to try some eGPU boosted applications, you can test it with a game or a benchmark like <a href="https://benchmark.unigine.com/valley">Valley Benchmark</a>, so satisfying by the way. But surely you want to test your eGPU boost with your heavy Deep Learning models.</p><p>I’d bet that if you had problems installing the eGPU after the time spent you are rethinking about how much profitable was going trough all that hell. But think that most of the way is done, congratulations! Maybe.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*Ag5vOnaUNYG8ey2R" /></figure><p>Remember when I wrote about that war between Nvidia and Apple? Well its consequences didn’t end there. If you’re a Data Scientist who has worked a bit with Tensorflow, you surely know this but if it not the case I will remember it, TensorFlow GPU works with CUDA, a Nvidia software, so as Nvidia hardware is no longer officially supported by Apple, for the Tensorflow team it does not make sense to give an oficial MacOS Tensorflow-GPU support anymore.</p><p>Again we are going to trick the system. In the best case you will only need to use the last tensorflow-gpu-macos pre-build, you can try it by:</p><pre>pip3 install tensorflow-gpu-macosx</pre><p><strong><em>NOTE:</em></strong><em> If your specific configuration match with CUDA 10, cuDNN 7.4, Python 3.6.8, we have this gift for you, our own </em><a href="https://github.com/XPi2/tensorflow-gpu-macosx/releases/tag/1.8.0"><strong><em>Tensorflow1.8 GPU build wheel</em></strong></a><em>. Download and try:</em></p><pre>pip install *.whl</pre><p>If this is not the case welcome to the <strong>tutorial of how to build Tensorflow GPU to your own configuration</strong>, this tutorial is based on <a href="https://github.com/zylo117/tensorflow-gpu-macosx">this great repository</a> by <a href="https://github.com/zylo117">@zylo117</a>, to which we have contributed to. You can follow the process there or continue reading.</p><h4>How to build Tensorflow GPU</h4><p><strong>1.</strong> You need to have your <strong>CUDA and cuDNN environment working</strong>. If you come here as Deep Learning developer you should have done this many times. If CUDA is installed you can check version by:</p><pre>nvcc -V</pre><p>Else: <a href="https://developer.nvidia.com/cuda-toolkit">Cuda Toolkit</a> | <a href="https://developer.nvidia.com/cudnn">cuDNN</a> installations(NVIDIA Developer account needed, follow the instructions)</p><p><strong>2.</strong> <strong>Install XCode/Command Line Tools 9.3+</strong>(recommended, older versions could work)</p><p>You can easy check if installed version by:</p><pre>gcc -v</pre><p>Else: <a href="https://developer.apple.com/download/more/">XCode</a> | <a href="https://developer.apple.com/download/more/">Command Line Tools</a> installations (Apple Developer account needed)</p><p><strong>NOTE: </strong>If you want to downgrade your Command Line Tools, is recommended to remove the old files by:</p><pre>rm -rf /Library/Developer/CommandLineTools/</pre><p><strong>3. Install CoreUtils &amp; LLVM by Homebrew</strong></p><p>If you don’t know what Homebrew is you should <a href="https://brew.sh/">read</a> about it. Once you have it:</p><pre>brew install coreutils llvm</pre><p><strong>4. Install OpenMP</strong></p><pre>brew install cliutils/apple/libomp</pre><p>Easy steps.</p><p><strong>5. Install Bazel 0.16.1 from </strong><a href="https://github.com/bazelbuild/bazel/releases/tag/0.16.1"><strong>GitHub</strong></a><strong> </strong>(newer/older version may cause failure)</p><p>Follow <a href="https://docs.bazel.build/versions/master/install-os-x.html">Bazel instructions</a>.</p><p><strong>6. Clone Tensorflow-GPU-MacOS repo</strong></p><pre>git clone <a href="https://github.com/zylo117/tensorflow-gpu-macosx">https://github.com/zylo117/tensorflow-gpu-macosx</a><br>cd tensorflow-gpu-macosx</pre><p><strong>7. Config (to your custom CUDA and cuDNN build)</strong></p><pre>./configure</pre><pre>  #Please specify the location of python.: Accept the default option<br>    #Please input the desired Python library path to use.:  Accept the default option<br>    #Do you wish to build TensorFlow with Google Cloud Platform support? [Y/n]: n<br>    #Do you wish to build TensorFlow with Hadoop File System support? [Y/n]: n<br>    #Do you wish to build TensorFlow with Amazon S3 File System support? [Y/n]: n<br>    #Do you wish to build TensorFlow with XLA JIT support? [y/N]: n<br>    #Do you wish to build TensorFlow with GDR support? [y/N]: n<br>    #Do you wish to build TensorFlow with VERBS support? [y/N]: n<br>    #Do you wish to build TensorFlow with OpenCL SYCL support? [y/N]: n<br>    #Do you wish to build TensorFlow with CUDA support? [y/N]: y<br>    #Please specify the CUDA SDK version you want to use, e.g. 7.0.: Put yours<br>    #Please specify the location where CUDA 9.1 toolkit is installed.: Accept the default option<br>    #Please specify the cuDNN version you want to use.: Put yours<br>    #Please specify the location where cuDNN 7 library is installed.: Accept the default option<br>    ##Please specify a list of comma-separated Cuda compute capabilities you want to build with.<br>    ##You can find the compute capability of your device at: https://developer.nvidia.com/cuda-gpus. (GTX10X0: 6.1, GTX9X0: 5.2)<br>    #Please note that each additional compute capability significantly increases your build time and binary size.: 6.1,5.2,5.0,3.0<br>    #Do you want to use clang as CUDA compiler? [y/N]: n<br>    #Please specify which gcc should be used by nvcc as the host compiler.: Accept the default option<br>    #Do you wish to build TensorFlow with MPI support? [y/N]: n<br>    #Please specify optimization flags to use during compilation when bazel option &quot;--config=opt&quot; is specified:  Accept the default option<br>    #Would you like to interactively configure ./WORKSPACE for Android builds? [y/N]: n</pre><pre># If it haven&#39;t done before:<br>export CUDA_HOME=/usr/local/cuda<br>export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:/usr/local/cuda/extras/CUPTI/lib<br>export LD_LIBRARY_PATH=$DYLD_LIBRARY_PATH<br>export PATH=$DYLD_LIBRARY_PATH:$PATH<br><br># bazel clean --expunge # Use this if you failed to compile before.</pre><p><strong>8. Build</strong></p><p>This step will take a few hours, it’s time to run the following comment and watch one of the movies you’ve been procrastinating forever 🍿.</p><pre>bazel build --config=cuda --config=opt --cxxopt=&quot;-D_GLIBCXX_USE_CXX11_ABI=0&quot; --action_env PATH --action_env LD_LIBRARY_PATH --action_env DYLD_LIBRARY_PATH //tensorflow/tools/pip_package:build_pip_package</pre><p><strong>9. NCCL Patch</strong></p><p>You can compile NCCL_OPS:</p><pre>gcc -c -fPIC ./nccl_patched/nccl_ops.cc -o ./nccl_patched/_nccl_ops.o</pre><pre>gcc ./nccl_patched/_nccl_ops.o -shared -o ./nccl_patched/_nccl_ops.so</pre><p>Then replace the original nccl lib:</p><pre>mv ./bazel-bin/tensorflow/contrib/nccl/python/ops/_nccl_ops.so ./bazel-bin/tensorflow/contrib/nccl/python/ops/_nccl_ops.so.bk</pre><pre>cp ./nccl_patched/_nccl_ops.so ./bazel-bin/tensorflow/contrib/nccl/python/ops/</pre><p><strong>10. Build Python binding</strong></p><pre>./bazel-bin/tensorflow/tools/pip_package/build_pip_package/ /tmp/tensorflow_pkg</pre><p><strong>Finally. Install python wheel</strong></p><pre>pip3 install /tmp/tensorflow_pkg/*.whl</pre><p>And don’t forget to share your .whl with the world by Forking + Pull Requesting on <a href="https://github.com/zylo117/tensorflow-gpu-macosx">here</a>, you can save a lot of time to a lot of people, think about it this way.</p><p><strong>Verify your tensorflow-gpu installation</strong></p><pre>python<br>&gt;&gt;&gt; import tensorflow as tf<br>&gt;&gt;&gt; hello = tf.constant(&#39;Hello, TensorFlow!&#39;)<br>&gt;&gt;&gt; sess = tf.Session()<br>...<br>#Something like<br>...<br>name: GeForce GTX 1070 major: 6 minor: 1 memoryClockRate(GHz): 1.7845<br>pciBusID: 0000:c4:00.0<br>totalMemory: 8.00GiB freeMemory: 5.62GiB<br>...<br>2019-01-25 18:22:02.776549: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 5402 MB memory) -&gt; physical GPU (device: 0, name: GeForce GTX 1070, pci bus id: 0000:c4:00.0, compute capability: 6.1)<br>&gt;&gt;&gt; print(sess.run(hello))<br>b&#39;Hello, Tensorflow!&#39;</pre><p>*Out of the tensorflow-gpu-macos dir.</p><p>If you got here, congratulations, you’ve made it! 👏👏👏</p><p>I hope you have found this guide useful. In such case, don’t forget to clap it, share it and subscribe! Should you have any change requests on this guide, you can leave a comment below. I’ll try to answer / redirect everything as fast as I can.</p><p>Check our services and explorations with Deep Learning, Machine Learning, Computer Vision and GANs on our <a href="https://www.linkedin.com/company/xplore-ai/">LinkedIn page</a>, <a href="https://twitter.com/XploreAi">Twitter</a>, <a href="https://www.instagram.com/xplore.ai">@xplore.ai Instagram account</a> and don’t forget to follow us on <a href="https://medium.com/xplore-ai">Medium</a> to not miss any post like this in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f40e831f26ea" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/nvidia-egpu-macos-tensorflow-gpu-the-definitive-setup-guide-to-avoid-headaches-f40e831f26ea">Nvidia eGPU + MacOS+ TensorFlow-GPU? The Definitive Setup Guide to Avoid Headaches</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 10 curated AI reads for December 2018]]></title>
            <link>https://medium.com/xplore-ai/top-10-curated-ai-reads-for-december-2018-390fea8a1e40?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/390fea8a1e40</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Mon, 10 Dec 2018 18:59:17 GMT</pubDate>
            <atom:updated>2019-01-18T10:27:11.381Z</atom:updated>
            <content:encoded><![CDATA[<h4>December, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oYXnMcVe6JWDzQ-9bGa17w.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/photos/5fedGbqYwvM?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Aaron Burden</a> on <a href="https://unsplash.com/search/photos/winter?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Welcome to xplore.ai’s sixth post of the <em>10 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><h3>Computer Vision</h3><blockquote><strong>1. DeOldify</strong></blockquote><p>A Deep Learning based project for colorizing and restoring old images.</p><p><a href="https://github.com/jantic/DeOldify">GitHub - jantic/DeOldify: A Deep Learning based project for colorizing and restoring old images (and video!)</a></p><blockquote><strong>2. DeepCTR</strong></blockquote><p>DeepCTR is a Easy-to-use, modular and extendible package of deep-learning based Click-Through Rate models along with lots of core components layer which can be used to build your own custom model easily.</p><p><a href="https://github.com/shenweichen/DeepCTR">shenweichen/DeepCTR</a></p><blockquote><strong>3. Retina U-Net</strong></blockquote><p>All-in-one 2D and 3D medical images semantic segmentation + object detection framework. Main idea is to fuse Retina Net one-stage detector with the U-Net architecture widely used for semantic segmentation in medical images.</p><p><a href="https://arxiv.org/abs/1811.08661">[1811.08661] Retina U-Net: Embarrassingly Simple Exploitation of Segmentation Supervision for Medical Object Detection</a></p><blockquote><strong>4. Progressive Calibration Networks</strong></blockquote><p>Progressive Calibration Networks (PCN) is an accurate rotation-invariant face detector running at real-time speed on CPU, published in CVPR 2018.</p><p><a href="https://github.com/Jack-CV/PCN-FaceDetection">Jack-CV/PCN-FaceDetection</a></p><blockquote><strong>5. Progressive Growing of GANs</strong></blockquote><p>Ideas is to train Generative Adversarial Networks starting from a low resolution, we add new layers that model increasingly fine details as training progresses. At xplore.ai we are testing this idea in our lab to create <a href="https://www.instagram.com/p/Bq0pyPhHYuv/">artificial pictures of food dishes</a></p><p><a href="https://arxiv.org/pdf/1710.10196.pdf">https://arxiv.org/pdf/1710.10196.pdf</a></p><blockquote><strong>6. Applying Convolutional Neural Networks in the Google Pixel 3 camera software</strong></blockquote><p>Very cool explanation of how depth estimation powers the Portrait Mode on the Google Pixel 3. It uses a CNN written in TensorFlow, check it out!</p><p><a href="https://ai.googleblog.com/2018/11/learning-to-predict-depth-on-pixel-3.html?m=1">Learning to Predict Depth on the Pixel 3 Phones</a></p><h3>Machine Learning</h3><blockquote><strong>7. Glossary of Machine Learning terms</strong></blockquote><p>Glossary of Machine Learning terms. Pretty good idea, especially these days where new concepts keep growing like mushrooms. Still a lot to be added such as “euclidean distance”, “Region of Interest”, “upsampling”, “gaussian mixture model”, …</p><p><a href="https://semanti.ca/blog/?glossary-of-machine-learning-terms">https://semanti.ca/blog/?glossary-of-machine-learning-terms</a></p><blockquote><strong>8. AI Related Announcements From Amazon’s AWS Re:Invent 2018</strong></blockquote><p>At xplore.ai, we have been keeping up with the latest news by Amazon at the Amazon Web Services largest event. Such official event is of great interest because it is the moment we can infer what are the things AWS is betting on, and therefor how they foresee the next year’s areas with potential. For us, the most relevant things were:</p><p>1. Larger GPU instances such as the P3dn = Deep Learning is the future</p><p>2. SageMaker large improvements + Amazon Elastic Inference = ML/AI on its way to get commoditized</p><p>3. Amazon SageMaker RL (Reinforcement Learning) = RL is not only for machines to overperform humans in videogames</p><p>4. AWS Training and Certification for Machine Learning = demand for DS still too far from offer</p><p><a href="https://www.forbes.com/sites/cognitiveworld/2018/11/28/ai-related-announcements-from-amazons-aws-reinvent-2018-show/#2fde01cc6c3b">AI Related Announcements From Amazon&#39;s AWS Re:Invent 2018 Show</a></p><blockquote><strong>9. Awesome list of ML interpretability</strong></blockquote><p>One of those “awesome” lists. We love lists, you have probably noticed it, so here you have one with interesting resources to take into consideration when dealing with interpretability of ML models.</p><p><a href="https://github.com/jphall663/awesome-machine-learning-interpretability">GitHub - jphall663/awesome-machine-learning-interpretability: A curated list of awesome machine learning interpretability resources.</a></p><h3>Natural Language Processing</h3><blockquote><strong>10. Recent History of NLP</strong></blockquote><p>This first of 2 post series discusses major recent advances in NLP focusing on neural network-based methods. The second post will discuss open problems in NLP.</p><p><a href="http://blog.aylien.com/a-review-of-the-recent-history-of-natural-language-processing/">A Review of the Neural History of Natural Language Processing</a></p><p>And this is it for what we found out to be interesting during November. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em>,</em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em> and </em><a href="https://www.instagram.com/xplore.ai/"><em>now Instagram</em></a><em> or drop </em><a href="https://www.linkedin.com/in/eherreros/"><em>me</em></a><em> a message. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=390fea8a1e40" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/top-10-curated-ai-reads-for-december-2018-390fea8a1e40">Top 10 curated AI reads for December 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 10 curated AI reads for November 2018]]></title>
            <link>https://medium.com/xplore-ai/top-10-curated-ai-reads-for-november-2018-c7d702a40cf3?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/c7d702a40cf3</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Tue, 06 Nov 2018 13:08:12 GMT</pubDate>
            <atom:updated>2018-11-06T13:08:12.181Z</atom:updated>
            <content:encoded><![CDATA[<h4>November, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KSQQWghEAcnDZSSfB81X1g.jpeg" /></figure><p>Welcome to xplore.ai’s fifth post of the <em>10 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><h3>Computer Vision</h3><blockquote><strong>1. European Conference on Computer Vision (ECCV) 2018 Summary</strong></blockquote><p>2018 edition of European Conference on Computer Vision (ECCV) insights by Scortex. In summary, very interesting findings regarding Group Normalization (GN) which improves object detection algorithms, Hourglass Networks in which Corner Net leverage, Receptive Field Block Net (RFB Net) for more “human” receptive fields, BiseNet for lightweight segmentation, DeepLab v3+, etc</p><p><a href="https://scortex.io/eccv-2018-deep-learning-summary/">Our ECCV 2018 recap | Scortex</a></p><blockquote><strong>2. <em>🏀 </em>Interesting usage of 3D CNNs to label NBA videos</strong></blockquote><p>Labelling NBA actions using deep learning. Label NBA broadcast footage with play-by-play descriptions, using 3D ConvNet-based video classification.</p><p><a href="https://github.com/neeilan/DeepPlayByPlay">neeilan/DeepPlayByPlay</a></p><blockquote><strong>3. Deep Exemplar-based Colorization paper implementation</strong></blockquote><p>Deep Exemplar-based Colorization is the ﬁrst deep learning approach for exemplar-based local colorization. Given a reference color image, our convolutional neural network directly maps a grayscale image to an output colorized image.</p><p><a href="https://github.com/msracver/Deep-Exemplar-based-Colorization">msracver/Deep-Exemplar-based-Colorization</a></p><p>Paper: <a href="https://arxiv.org/pdf/1807.06587.pdf">https://arxiv.org/pdf/1807.06587.pdf</a></p><blockquote><strong>4. Playing Mortal Kombat with your webcam in a the browser</strong></blockquote><p>Transfer learning, data augmentation, MobileNet, tensorflow.js, Mortal Kombat, webcam… It is a really fun project worth taking a look at</p><p><a href="https://blog.mgechev.com/2018/10/20/transfer-learning-tensorflow-js-data-augmentation-mobile-net/">Playing Mortal Kombat with TensorFlow.js. Transfer learning and data augmentation</a></p><blockquote><strong>5. Object Detection status summary</strong></blockquote><p>Recent Advances in Object Detection in the Age of Deep Convolutional Neural Networks</p><p><a href="https://arxiv.org/pdf/1809.03193.pdf">https://arxiv.org/pdf/1809.03193.pdf</a></p><h3>Machine Learning</h3><blockquote><strong>6. Deep Learning on Collaborative Filtering at Hulu</strong></blockquote><p>Applying Deep Learning to Collaborative Filtering: How Hulu builds its industry leading recommendation engine using Neural Autoregressive Distribution Estimator for Collaborative filtering (CF-NADE). Will be nice to see how they adapt CF-NADE to implicit feedback in the future.</p><p><a href="https://medium.com/hulu-tech-blog/applying-deep-learning-to-collaborative-filtering-how-hulu-builds-its-industry-leading-3b10a4ed7470">Applying Deep Learning to Collaborative Filtering: How Hulu builds its industry leading…</a></p><h3>General Tools</h3><blockquote><strong>7. When you git a jupyter notebook and then want to diff changes… <em>💣</em></strong></blockquote><p>Tool for diffing and merging of Jupyter Notebooks. Provides “content-aware” diffing and merging of Jupyter notebooks. It understands the structure of notebook documents. Therefore, it can make intelligent decisions when diffing and merging notebooks.</p><p><a href="https://nbdime.readthedocs.io/en/stable/">nbdime - diffing and merging of Jupyter Notebooks - nbdime 1.0.3 documentation</a></p><h3>Recommender systems</h3><blockquote><strong>8. Collaborative Embeddings for Lipstick Recommendations</strong></blockquote><p>Exploiting a nice idea that is normally over our heads: generating embeddings out of items in the inventory. Also, we appreciate that the author accounted for discoverability, factor that is typically ignored in recommender systems applications. Worth a couple of reads.</p><p><a href="https://towardsdatascience.com/collaborative-embeddings-for-lipstick-recommendations-98eccfa816bd">Collaborative Embeddings for Lipstick Recommendations</a></p><blockquote><strong>9. Applying Deep Learning To Airbnb Search</strong></blockquote><p>Remarkable paper from Airbnb making use of Deep Learning models in their business. Not only they tell the successful creations but also the unsuccessful atempts. Hashing tricks for high cardinality categoricals, rethinking the entire system surrounding the model when moving to deep learning, protobuf, Spark, multi-task learning, … joy.</p><p><a href="https://arxiv.org/pdf/1810.09591.pdf">https://arxiv.org/pdf/1810.09591.pdf</a></p><h3>Graph Networks</h3><blockquote><strong>10. Graph Nets library for TensorFlow</strong></blockquote><p>Build Graph Nets in Tensorflow. A graph network takes a graph as input and returns a graph as output. The input graph has edge- (<em>E</em> ), node- (<em>V</em> ), and global-level (u) attributes. The output graph has the same structure, but updated attributes.</p><p><a href="https://github.com/deepmind/graph_nets">GitHub - google-deepmind/graph_nets: Build Graph Nets in Tensorflow</a></p><p>And this is it for what we found out to be interesting during October. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em> and </em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em> or drop </em><a href="https://www.linkedin.com/in/eherreros/"><em>me</em></a><em> a message. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7d702a40cf3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/top-10-curated-ai-reads-for-november-2018-c7d702a40cf3">Top 10 curated AI reads for November 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 10 curated AI reads for October 2018]]></title>
            <link>https://medium.com/xplore-ai/top-10-curated-ai-reads-for-october-2018-b06887b75d67?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/b06887b75d67</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Fri, 05 Oct 2018 07:31:02 GMT</pubDate>
            <atom:updated>2018-10-05T07:31:02.455Z</atom:updated>
            <content:encoded><![CDATA[<h4>October, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JzrMer5voaSeYysYl8vSAA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/photos/Fxv7RFLD8kc?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Ethan Weil</a> on <a href="https://unsplash.com/search/photos/satellite?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Welcome to xplore.ai’s forth post of the <em>10 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><h3>Computer Vision</h3><blockquote><strong>1. What-If Tool for Tensorflow</strong></blockquote><p>A new feature of the open-source <a href="https://github.com/tensorflow/tensorboard">TensorBoard</a> web application, which let users analyze an ML model without writing code. Given pointers to a TensorFlow model and a dataset, the What-If Tool offers an interactive visual interface for exploring model results. Although it applies to modeling anything in TF, we think Computer Vision is the most interesting application</p><p><a href="https://ai.googleblog.com/2018/09/the-what-if-tool-code-free-probing-of.html?m=1">The What-If Tool: Code-Free Probing of Machine Learning Models</a></p><blockquote><strong>2. GAN Lab</strong></blockquote><p>An Interactive, Visual Experimentation Tool for Generative Adversarial Networks understanting. We also liked the textual explanation bellow.</p><p><a href="https://poloclub.github.io/ganlab/">GAN Lab: Play with Generative Adversarial Networks in Your Browser!</a></p><h3>Machine Learning</h3><blockquote><strong>3. New Pandas-like features in Scikit-Learn</strong></blockquote><p>We are excited to see that managing feature creation and transformation will be from now on simpler, more feature-rich, robust, and standardized. Scikit-Learn 0.20 looks so promising! Thanks to the community for such advances.</p><p><a href="https://medium.com/dunder-data/from-pandas-to-scikit-learn-a-new-exciting-workflow-e88e2271ef62">From Pandas to Scikit-Learn — A new exciting workflow</a></p><blockquote><strong>4. Forecasting at Uber (I)</strong></blockquote><p>“This article is the first in a series dedicated to explaining how Uber leverages forecasting to build better products and services”. At xplore.ai we are working on forecasting problems in the travel (hotel) industry and completely agree with the “Prediction intervals are just as important as the point forecast itself and should always be included in your forecasts” statement.</p><p><a href="https://eng.uber.com/forecasting-introduction/">Forecasting at Uber: An Introduction</a></p><blockquote><strong>5. Algorithmic Marketing Book</strong></blockquote><p>Very complete book that covers the main areas of marketing that require micro-decisioning — targeted promotions and advertisements, eCommerce search, recommendations, pricing, and assortment optimization.</p><p><a href="https://algorithmic-marketing.online/">Book</a></p><h3>Reinforcement Learning</h3><blockquote><strong>6. DQN RL Flappy Bird game</strong></blockquote><p>PyTorch reinforcement learning tutorial, the author showed how a computer can learn to play Flappy Bird without any previous knowledge about the game, using only a trial-and-error approach as a human would do when encountering the game for the first time.</p><p><a href="https://github.com/nevenp/dqn_flappy_bird">nevenp/dqn_flappy_bird</a></p><blockquote><strong>7. 60 days Reinforcement Learning Challenge</strong></blockquote><p>Learn Deep Reinforcement Learning in Depth in 60 days</p><p><a href="https://github.com/andri27-ts/60_Days_RL_Challenge">andri27-ts/60_Days_RL_Challenge</a></p><blockquote><strong>8. Usage of Embeddings at OpenAI five</strong></blockquote><p>OpenAI Five (Dota 2 AI player) aspect study of their network architecture — their inventive use of embeddings to handle a huge and variable number of policy inputs and outputs. While the use of embeddings and dot product for attention are standard techniques in natural language processing, they are not widely used in reinforcement learning.</p><p><a href="https://neuro.cs.ut.ee/the-use-of-embeddings-in-openai-five/">The use of embeddings in OpenAI Five</a></p><h3>Bayesian methods</h3><blockquote><strong>9. Summer School Deep|Bayes</strong></blockquote><p>The Bayesian Methods Research group is one of the leading machine learning research groups in Russia. Members of the group have developed a range of university courses in Bayesian Methods, Deep Learning, Optimization and Probabilistic Graphical Models. You can find all the videos, slides and assignments open sourced in their website</p><p><a href="http://deepbayes.ru/#materials">Summer school Deep|Bayes</a></p><h3>Natural Language Processing</h3><blockquote><strong>10. Key topics extraction and contextual sentiment of users’ reviews</strong></blockquote><p>Key topics extraction and contextual sentiment of users’ reviews using Spacy + NLTK (for VaderSentiment) + Sklearn + difflib. Pretty simple yet cool pipeline for such task</p><p><a href="https://tech.goibibo.com/key-topics-extraction-and-contextual-sentiment-of-users-reviews-20e63c0fd7ca">Key topics extraction and contextual sentiment of users’ reviews</a></p><p>And this is it for what we found out to be interesting during September. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check the other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em> and </em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em> or drop </em><a href="https://www.linkedin.com/in/eherreros/"><em>me</em></a><em> a message. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b06887b75d67" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/top-10-curated-ai-reads-for-october-2018-b06887b75d67">Top 10 curated AI reads for October 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[15 curated AI reads for September 2018]]></title>
            <link>https://medium.com/xplore-ai/15-curated-ai-reads-for-september-2018-59886d45fb73?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/59886d45fb73</guid>
            <category><![CDATA[curation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Tue, 04 Sep 2018 07:31:01 GMT</pubDate>
            <atom:updated>2018-09-04T17:10:55.941Z</atom:updated>
            <content:encoded><![CDATA[<h4>September, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JgEJ4vlnaPNQjsOoKy05Dw.jpeg" /><figcaption>Credits to our team. Photo taken while testing iPhone’s X waterproofness</figcaption></figure><p>Welcome to xplore.ai’s third post of the <em>15 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><blockquote><strong>15. </strong>✏️ Academic Torrents</blockquote><p>A scalable, secure, and fault-tolerant repository for data, with blazing fast download speeds. We are thankful for such work, kudos to Lo, Henry Z. and Cohen, Joseph P., for such a big effort.</p><p><a href="http://academictorrents.com/">Academic Torrents</a></p><blockquote><strong>14. </strong>💃 DanceNet</blockquote><p>Dance generator using Autoencoders, LSTM and Mixture Density Network. Build in Keras. <a href="https://arxiv.org/abs/1605.06921v1">Paper</a> and <a href="https://github.com/jsn5/dancenet">code</a> clicking.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FSc7RiNgHHaE%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DSc7RiNgHHaE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FSc7RiNgHHaE%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/7e0023374563a4ad0765749b3323071a/href">https://medium.com/media/7e0023374563a4ad0765749b3323071a/href</a></iframe><p><a href="https://raw.githubusercontent.com/ft-interactive/chart-doctor/master/visual-vocabulary/poster.png">https://raw.githubusercontent.com/ft-interactive/chart-doctor/master/visual-vocabulary/poster.png</a></p><blockquote><strong>13. </strong><em>💣</em>AutoKeras</blockquote><p>A novel framework enabling Bayesian optimization to guide the network morphism for efficient neural architecture search by introducing a neural network kernel and a tree-structured acquisition function optimization algorithm. With Bayesian optimization to select the network morphism operations, the exploration of the search space is more efficient</p><p><a href="https://github.com/jhfjhfj1/autokeras">Code</a> and <a href="https://arxiv.org/abs/1806.10282v1">paper</a></p><p><a href="https://autokeras.com/">AutoKeras</a></p><blockquote><strong>12. </strong><em>🏠</em> Creating 3D renders out of a single image of a space</blockquote><p>Holistic 3D Scene Parsing and Reconstruction from a Single RGB Image using a stochastic grammar model integrated with latent human context, geometry and physics. Paper <a href="https://arxiv.org/pdf/1808.02201.pdf">here</a>. Pretty cool job.</p><p><a href="https://github.com/thusiyuan/holistic_scene_parsing">thusiyuan/holistic_scene_parsing</a></p><blockquote><strong>11. </strong>🖇️️ Adversarial Vision Challenge at NIPS 2018</blockquote><p>“The overall goal of this challenge is to facilitate measurable progress towards robust machine vision models and more generally applicable adversarial attacks. As of right now, modern machine vision algorithms are extremely susceptible to small and almost imperceptible perturbations of their inputs (so-called <em>adversarial examples</em>). This property reveals an astonishing difference in the information processing of humans and machines and raises security concerns.”.</p><p><a href="https://github.com/bethgelab/adversarial-vision-challenge">bethgelab/adversarial-vision-challenge</a></p><blockquote><strong>10. </strong><em>📙</em>Model-based ML book — Early access</blockquote><p>Applied ML book, early access. “In this book we look at machine learning from a fresh perspective which we call model-based machine learning. This viewpoint helps to address all of these challenges, and makes the process of creating effective machine learning solutions much more systematic.”</p><p><a href="http://www.mbmlbook.com/toc.html">Model-Based Machine Learning (Early Access): Table of Contents</a></p><blockquote><strong>9. </strong><em>💧</em>Attentive Generative Adversarial Network for Raindrop Removal from A Single Image (CVPR’2018)</blockquote><p>Very interesting Deep Learning project to remove raindrops from images. They used 1119 pairs of images, with various background scenes and raindrops, simulated with sprayed glasses. The other side of the pair is the same scene but without the glass with the drops. The method utilizes a generative adversarial network, where the generative network produces the attention map via an attentive-recurrent network and applies this map along with the input image to generate a raindrop-free image through a contextual autoencoder. The discriminative network then assesses the validity of the generated output globally and locally. To be able to validate locally, they inject the attention map into the network. The authors novelty lies on the use of the attention map in both generative and discriminative network..</p><p><a href="https://github.com/rui1996/DeRaindrop">rui1996/DeRaindrop</a></p><blockquote><strong>8. </strong><em>🔖</em> Learning Market Dynamics for Optimal Pricing (not only ML)</blockquote><p>Strongly recommended (pure art) post from Engineering at AirBnB where they model the lead time probability of a booking. “This model is currently used in product, primarily to power their Smart Pricing. The tool uses the predicted lead time distributions for each check-in to help hosts keep prices up to date. We also use it to inform hosts about booking lead times statistics to help them make informed decisions around calendar availability”.</p><p>They combine 2 optimization functions: 1 time to event can be approximated by a gamma distribution because the bookings in a unit of time are Poisson distributed. The other function is the oscillating waveform adjuster that will multiply the 1st function. Click to keep reading:</p><p><a href="https://medium.com/airbnb-engineering/learning-market-dynamics-for-optimal-pricing-97cffbcc53e3">Learning Market Dynamics for Optimal Pricing</a></p><blockquote><strong>7. </strong><em>📘 S</em>tate of the art text classification with universal language models</blockquote><p>Fine tuning NLP models with ULMFiT. Full neural networks in practice contain many layers, so only using transfer learning for a single layer (word2vec) was clearly just scratching the surface of what’s possible. J Howard et al reached the performance of training from scratch on 100x more data.</p><p><a href="https://blog.feedly.com/transfer-learning-in-nlp/">Transfer Learning in NLP</a></p><blockquote><strong>6. </strong>↩️ Hyperparameter tuning visual replay in Keras</blockquote><p>Deep Replay — Generate visualizations as in my “Hyper-parameters in Action!” series! Replay in a visual fashion the training process of a Deep Learning model in Keras</p><p><a href="https://github.com/dvgodoy/deepreplay">dvgodoy/deepreplay</a></p><blockquote><strong>5. </strong>🇹 TensorFlow 2.0</blockquote><p>We are glad to here that, among various changes, there’s a focus on the Eager execution mode, which will be a central feature.</p><p><a href="https://groups.google.com/a/tensorflow.org/forum/?utm_medium=email&amp;utm_source=footer#!topic/announce/qXfsxr2sF-0">Google Groups</a></p><blockquote><strong>4. </strong>🕺 Vid2Vid</blockquote><p>Pytorch implementation of our method for high-resolution (e.g. 2048x1024) photorealistic video-to-video translation. You can read the paper <a href="https://arxiv.org/pdf/1808.06601.pdf">here</a>. Goal is to map sequence of semantic segmentation masks to an output photorealistic video that precisely depicts the content of the source video.</p><p><a href="https://github.com/NVIDIA/vid2vid">NVIDIA/vid2vid</a></p><blockquote><strong>3. </strong><em>😀</em>Time Series Forecasting Cheat Sheet</blockquote><p>We have been waiting for something like this for a long time now. It was about time to have a comprehensible short document explaining the well-known Forecasting methods out there.</p><p><a href="https://machinelearningmastery.com/time-series-forecasting-methods-in-python-cheat-sheet">11 Classical Time Series Forecasting Methods in Python (Cheat Sheet)</a></p><blockquote><strong>2. </strong><em>📈</em> PolyRNN+</blockquote><p>Tool for Efficient Annotation of Segmentation Datasets.</p><p><a href="http://www.cs.toronto.edu/polyrnn/">polyrnn++</a></p><blockquote><strong>1. </strong>🌱 Automatic answers to Q about an image</blockquote><p>Visual Question Answering (VQA). Data: 82,783 training images from COCO (common objects in context) dataset + 443,757 question-answer pairs for training images + 40,504 validation images to perform own testing + 214,354 question-answer pairs for validation images. Model: 3 stacked 512-unit LSTM layers to process questions + VGG16 for the images.</p><p><a href="https://blog.floydhub.com/asking-questions-to-images-with-deep-learning/">Asking questions to images with deep learning: a visual-question-answering tutorial</a></p><p>And this is it for what we found out to be interesting during sunny August. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check the other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em> and </em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em> or drop </em><a href="https://www.linkedin.com/in/eherreros/"><em>me</em></a><em> a message. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=59886d45fb73" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/15-curated-ai-reads-for-september-2018-59886d45fb73">15 curated AI reads for September 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[15 curated AI reads for August 2018]]></title>
            <link>https://medium.com/xplore-ai/15-curated-ai-reads-for-august-2018-e7a1fa6c7be5?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/e7a1fa6c7be5</guid>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Wed, 01 Aug 2018 19:46:01 GMT</pubDate>
            <atom:updated>2018-08-01T19:46:01.263Z</atom:updated>
            <content:encoded><![CDATA[<h4>August, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9LMQb73mekDMLSSwzgH9ZA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/photos/ZZLiZmBtA8I?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">J. Kelly Brito</a> on <a href="https://unsplash.com/search/photos/summer-work?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Welcome to xplore.ai’s second post of the <em>15 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><blockquote><strong>15. </strong>✏️ Agile methodologies in Data Analytics</blockquote><p>The good, the bad and the adjustments to make Agile methodologies work in the Data industry:</p><p><a href="https://www.locallyoptimistic.com/post/agile-analytics-p1/">https://www.locallyoptimistic.com/post/agile-analytics-p1/</a></p><p><a href="https://www.locallyoptimistic.com/post/agile-analytics-p2">https://www.locallyoptimistic.com/post/agile-analytics-p2</a></p><p><a href="https://www.locallyoptimistic.com/post/agile-analytics-p3">https://www.locallyoptimistic.com/post/agile-analytics-p3</a></p><blockquote><strong>14. </strong>✏️ Types of charts, visual vocabulary cheatsheet</blockquote><p>Simple and to the point complete cheatsheet of the common types of charts</p><p><a href="https://raw.githubusercontent.com/ft-interactive/chart-doctor/master/visual-vocabulary/poster.png">https://raw.githubusercontent.com/ft-interactive/chart-doctor/master/visual-vocabulary/poster.png</a></p><blockquote><strong>13. </strong><em>💣</em>Robust ML, adversarial attacks</blockquote><p>A community-run hub for learning about robust ML, a rapidly growing field that spans diverse communities across academia and industry. Robust models and precisely specifying the threat models under which they claim to be secure. Also includes interfaces for specifying attacks and evaluating attacks against models.</p><p><a href="https://www.robust-ml.org/">Robust ML</a></p><blockquote><strong>12. </strong><em>📄</em>Repo where tracking progress of NLP happens</blockquote><p>Repository to track the progress in Natural Language Processing (NLP), including the datasets and the current state-of-the-art for the most common NLP tasks. Really nice to keep an eye on, if you want to be up to date in NLP.</p><p><a href="https://github.com/sebastianruder/NLP-progress">sebastianruder/NLP-progress</a></p><blockquote><strong>11. </strong>✏️ Cancer Metastasis detection using Neural Conditional Random Field (NCRF)</blockquote><p>Breast cancer diagnosis often requires accurate detection of metastasis in lymph nodes through Whole-slide Images (WSIs). Recent advances in deep convolutional neural networks (CNNs) have shown significant successes in medical image analysis and particularly in computational histopathology. Good results are achieved when using neural conditional random field (NCRF). The corresponding paper can be found <a href="https://openreview.net/pdf?id=S1aY66iiM">here</a>.</p><p><a href="https://github.com/baidu-research/NCRF">baidu-research/NCRF</a></p><blockquote><strong>10. </strong>✏️ Organized Resources for Deep Learning in Natural Language Processing (NLP)</blockquote><p>Yet another set of organized resources for Deep Learning in Natural Language Processing (NLP). Pretty well segmented. Though, more than one month without updates, let’s see if they can keep up to date the info.</p><p><a href="https://github.com/astorfi/Deep-Learning-NLP">astorfi/Deep-Learning-NLP</a></p><blockquote><strong>9. </strong>✏️ 10 coolest papers from Computer Vision and Pattern Recognition 2018 conference</blockquote><p>During 18 jun. 2018–22 jun. 2018, CVPR (Computer Vision and Pattern Recognition) conference took place in Salt Lake City, US. The 10 coolest papers are presented and commented. Seems that the trending topics at the moment are spinning around image segmentation (efficient annotation tools, R-CNN, etc), GAN continues, training with synthetic data, extra slow-mo using interpolation.</p><p><a href="https://towardsdatascience.com/the-10-coolest-papers-from-cvpr-2018-11cb48585a49">The 10 coolest papers from CVPR 2018</a></p><blockquote><strong>8. </strong>✏️ Stockholm’s new promising startup called Peltarion, from ex-engineers at Spotify and King</blockquote><p>Senior engineers from a raft of other Swedish tech unicorns — Spotify, Klarna and Truecaller — have joined to create a startup called Peltarion, which has raised $16 million to simplify the usage of off-the-shelf machine-learning tools like Google’s TensorFlow, Amazon’s Sage Maker and Microsoft’s Azure Machine Learning.</p><p><a href="https://www.forbes.com/sites/parmyolson/2018/07/03/spotify-klarna-candy-crush-king-ai-sweden-peltarion/">Coders From Spotify, Klarna And Candy Crush&#39;s King Have Flocked To This AI Startup</a></p><blockquote><strong>7. </strong>✏️ What’s new in YOLO v3?</blockquote><p>Improvements from this popular object detection Neural Network. We are still amazed about the performance and smart ideas behind.</p><p><a href="https://towardsdatascience.com/yolo-v3-object-detection-53fb7d3bfe6b">What’s new in YOLO v3?</a></p><blockquote><strong>6. </strong>✏️ Hierarchical Attention Networks</blockquote><p>How to train and visualize Hierarchical Attention Networks?</p><p><a href="https://github.com/minqi/hnatt">minqi/hnatt</a></p><blockquote><strong>5. </strong>✏️ World Models</blockquote><p>Using <a href="https://chainer.org/">Chainer</a>, a powerful, flexible, and intuitive framework for Neural Networks, <a href="https://arxiv.org/abs/1803.10122">World Models </a>paper are implemented.</p><p><a href="https://github.com/AdeelMufti/WorldModels">AdeelMufti/WorldModels</a></p><blockquote><strong>4. </strong>🕺 GAIN (Guided Attention Inference Network)</blockquote><p>Contains implementation of Guided Attention Inference Network (GAIN) presented in Tell Me Where to Look (CVPR 2018). This repository aims to apply GAIN on fcn8 architecture used for segmentation.</p><p><a href="https://github.com/alokwhitewolf/Guided-Attention-Inference-Network">alokwhitewolf/Guided-Attention-Inference-Network</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3NjogA91dIXzbsbcOh7-VA.png" /><figcaption>Credits to <a href="https://github.com/alokwhitewolf">alokwhitewolf</a> repo</figcaption></figure><blockquote><strong>3. </strong><em>😀</em>Text to Face</blockquote><p>This study aims at transforming an English description of a face to an image, automatically. It combines two of the recent architectures <a href="https://arxiv.org/abs/1710.10916">StackGAN </a>and <a href="https://arxiv.org/abs/1710.10196">ProGAN</a>for synthesizing faces from textual descriptions.</p><p><a href="https://github.com/akanimax/T2F">akanimax/T2F</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/412/1*ySAGPRjlIA-XQiW72T_BXQ.png" /></figure><blockquote><strong>2. </strong><em>📈</em> M4 Forecasting Competition: Introducing a New Hybrid ES-RNN Model</blockquote><p>A nice discussion on how Slawek Smyl (Uber) hybrid Exponential Smoothing-Recurrent Neural Networks (ES-RNN) method work. It out-performed many other complex and simple but smart approaches. It mixes hand-coded parts like ES formulas with a black-box <a href="https://en.wikipedia.org/wiki/Recurrent_neural_network">recurrent neural network</a> (RNN) forecasting engine. Versions of the ES-RNN model are under development to tackle some of the most challenging problems in time-series forecasting here at Uber, across various use cases.</p><p><a href="https://eng.uber.com/m4-forecasting-competition/">M4 Forecasting Competition: Introducing a New Hybrid ES-RNN Model</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GZfp1_gGdS9A9V5DYQHvvA.png" /><figcaption>Credits to Slawek Smyl, Jai Ranganathan and Andrea Pasqua from Uber</figcaption></figure><blockquote><strong>1. </strong>🌱 Seedbank</blockquote><p>Yet another Google technical idea that will help a lot of people find classified cutting-edge examples and snippets on how to apply the latest AI advances with already created Google Colab notebooks.</p><p><a href="http://tools.google.com/seedbank/seeds">Seedbank</a></p><p>And this is it for what we found out to be interesting in June. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check the other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em> and </em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em> or drop </em><a href="https://www.linkedin.com/in/eherreros/"><em>me</em></a><em> a message. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e7a1fa6c7be5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/15-curated-ai-reads-for-august-2018-e7a1fa6c7be5">15 curated AI reads for August 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[15 curated AI reads for July 2018]]></title>
            <link>https://medium.com/xplore-ai/15-curated-ai-reads-for-july-2018-3dbf3ba71ba5?source=rss----8b2abc8b59a9---4</link>
            <guid isPermaLink="false">https://medium.com/p/3dbf3ba71ba5</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Enrique Herreros]]></dc:creator>
            <pubDate>Mon, 02 Jul 2018 18:58:28 GMT</pubDate>
            <atom:updated>2018-07-02T19:56:06.966Z</atom:updated>
            <content:encoded><![CDATA[<h4>July 3rd, 2018</h4><p><em>News about Machine Learning (ML), Artificial Intelligence (AI), Data Science (DS) and related advanced analytics areas.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HlX5coxH9Yl5db1S3-HMEw.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/photos/IXHNBGTKJfw?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Crew</a> on <a href="https://unsplash.com/?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Welcome to xplore.ai’s first post of the <em>15 curated AI reads</em> monthly series. The objective of these series is to provide the audience with a curated list of the most interesting news, publishings and tools that our team have ran into during the previous month.</p><blockquote><strong>15. </strong>✏️ TensorEditor</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1xkSkDAwW53y-HNa6qFAvw.png" /><figcaption>TensorEditor Beta screenshot</figcaption></figure><p>TensorEditor is a visual tool that allows even non-expert users to create their own TensorFlow models with just a few clicks. The user can tweak any parameter from the Neural Network. Once the model is ready, the tool will generate the respective Python code, so you can integrate anywhere and run predictions on it.</p><p><a href="https://www.tensoreditor.com/">TensorEditor. Build your deep learning code in seconds.</a></p><blockquote><strong>14. </strong><em>📈 </em>Probability Paradox Notebook</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7nKtsgIz74WYa6YG_N-oXA.png" /><figcaption>Monty Hall and its paradox</figcaption></figure><p>Although this notebook is already more than two years old, we ran into it this month and felt we have to share it with you. “Probability, Paradox, and the Reasonable Person Principle” holds a collection of 10 Python neatly-explained solutions.</p><p><a href="http://nbviewer.jupyter.org/url/norvig.com/ipython/ProbabilityParadox.ipynb">http://nbviewer.jupyter.org/url/norvig.com/ipython/ProbabilityParadox.ipynb</a></p><blockquote><strong>13. </strong><em>🔎</em> Python unified approach to explain the output of any machine learning model. Renders in JS</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dpb_lyepFgVBlHTzlAA4Xg.png" /></figure><p>Lately, more and more we see a trend in putting more effort in being able to explain ML and DL models. This is a good market sign, since it means these models made it to production and a lot of stakeholders need to understand the model’s output to take further decisions. We understand that grabbing insights at every step in the Data Science pipeline is vital. Hence, at <a href="http://xplore.ai/">xplore.ai</a>, we are trying out this module in our current projects.</p><p><a href="https://github.com/slundberg/shap">GitHub - slundberg/shap: A game theoretic approach to explain the output of any machine learning model.</a></p><blockquote><strong>12. </strong>🤼<em>‍ </em>Dense Human Pose Estimation In The Wild</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/1*OJv_lS3p0NwXHrPT25i2Vg.gif" /></figure><p>Facebook’s Research department has open-sources a real-time approach for mapping all human pixels of 2D RGB images to a 3D surface-based model of the body. They have open sourced the training data used: manually annotated 50K COCO images. The <a href="https://arxiv.org/pdf/1802.00434.pdf">paper</a> also discusses how region-based models outperform “classic” Convolutional Neural Networks. FB has used their own Deep Learning framework, Caffe2.</p><p><a href="https://github.com/facebookresearch/DensePose">GitHub - facebookresearch/DensePose: A real-time approach for mapping all human pixels of 2D RGB images to a 3D surface-based model of the body</a></p><blockquote><strong>11. </strong>🗣<em>️</em> Deep Learning for Conversational AI</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*KHeuYEOlEppa2C8UdTV5wA.png" /></figure><p>PolyAI has shared an impressive 187 pages slide deck that expose the current status of Conversational AI. They give an analysis of recent research trends in deep learning for conversational AI. They also present an industry-based perspective on current deep conversational AI. It seems to be a good moment to explore new scalable ways to generate impact in this huge industry.</p><p><a href="https://www.poly-ai.com/docs/naacl18.pdf">https://www.poly-ai.com/docs/naacl18.pdf</a></p><blockquote><strong>10.<em> </em></strong><em>💢</em> SNIPER: Efficient Multi-Scale Training</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*Mb_4AhZLzQreI4pwLkzJUw.gif" /><figcaption>Animation from SNIPER’s repository</figcaption></figure><p>SNIPER is an efficient multi-scale training approach for instance-level recognition tasks like object detection and instance-level segmentation. Instead of processing all pixels in an image pyramid, SNIPER selectively processes context regions around the ground-truth objects</p><p><a href="https://github.com/mahyarnajibi/SNIPER">mahyarnajibi/SNIPER</a></p><blockquote><strong>9. </strong><em>📑</em> Papers with Code</blockquote><p>This magnifique webpage shares links to research papers and their code implementations in GitHub. It even tells you the number of GitHub stars and what ML framework was used for the implementation. Don’t forget to subscribe to their weekly digest.</p><p><a href="https://paperswithcode.com/">Trending Papers - Hugging Face</a></p><blockquote><strong>8. </strong><em>📜</em> Learned in translation: contextualized word vectors</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZsAxIhWpQUEwjR9u-KAItQ.png" /><figcaption>Plot by Salesforce research team</figcaption></figure><p>Very accurate and to-the-point explanation by the research team at Salesforce. A very good intro on word vectors is given and then they explain how different techniques such as attention, encoders, hidden vectors, attention, generation, etc can help in NLP. Then, different NLP tasks are introduced: Machine Translation, Question Answering, Classification, Inference, Entailment, etc. They have researched that, by combining the information from GloVe, CoVe, and character vectors, a significant boost of the performance of our baseline models on a variety of NLP tasks occur.</p><p><a href="https://einstein.ai/research/learned-in-translation-contextualized-word-vectors">Salesforce research</a></p><blockquote><strong>7. </strong>👨<em>‍⚖️</em> Alibaba’s AI Lawyer Assitant</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cQLqA1EDdd4C5L_yb55y4w.png" /></figure><p>This study clearly shows that AI can successfully be implemented in the legal domain and that it can even improve accuracy when dealing with large volumes of legal documents when compared to professional lawyers. 8 lawyers compared 600 online legal agreements alongside the AI. The AI reached higher levels of accuracy in a millionth of the time.</p><p><a href="https://medium.com/@alitech_2017/artificial-intelligence-vs-lawyer-e5080225c718">Artificial Intelligence vs. Lawyers</a></p><blockquote><strong>6. </strong><em>🍎🍏</em> Detecting image similarity at Pinterest</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z-FNc5df1ke2OjNQM0vA8w.png" /><figcaption>Image from Pinterest Engineering Medium blog</figcaption></figure><p>In this post, Pinterest Engineering team explains how the made use of batched Local-Sensitive Hashing on top of Spark and TensorFlow to detect near-duplicate images. A wide range of smart tricks are applied to avoid having to do 10*10¹⁶ possible comparisons.</p><p><a href="https://medium.com/@Pinterest_Engineering/detecting-image-similarity-using-spark-lsh-and-tensorflow-618636afc939">Detecting image similarity using Spark, LSH and TensorFlow</a></p><blockquote><strong>5. </strong><em>🏎</em> ️100x faster NLP in Python</blockquote><p>By mixing Cython and Scipy’s internal data structures, Thomas Wolf exposes different techniques to create high-speed NLP algorithms out of useful profiling and C graciousness.</p><p><a href="https://medium.com/huggingface/100-times-faster-natural-language-processing-in-python-ee32033bdced">🚀 100 Times Faster Natural Language Processing in Python</a></p><blockquote><strong>4. </strong><em>💅</em> Understanding Latent Style</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KzqYbei1_GhJCX1gzgAW-g.png" /><figcaption>Stitch Fix visual tool</figcaption></figure><p>Stitch Fix is known, among other things, for its <a href="https://algorithms-tour.stitchfix.com/">attention to detail, care for design and love for data</a>. In this short post, Erin and Jana give an introduction of the tools they used to understand styles: PyTorch, Matrix Factorization, <a href="https://en.wikipedia.org/wiki/Principal_component_analysis">PCA</a> and <a href="https://en.wikipedia.org/wiki/Orthogonal_Procrustes_problem">Orthogonal Procrustes Matrix Approximation</a>. With those tools, they were able to build a latent style space where each item can be represented by a numerical vector. This reminded me of a <a href="https://kddfashion2016.mybluemix.net/kddfashion_finalSubmissions/Fashion%20DNA%20Merging%20Content%20and%20Sales%20Data%20for%20Recommendation%20and%20Article%20Mapping.pdf">Zalando’s Research Fashion DNA paper</a>.</p><p><a href="https://multithreaded.stitchfix.com/blog/2018/06/28/latent-style/">Understanding Latent Style | Stitch Fix Technology - Multithreaded</a></p><blockquote><strong>3. </strong><em>🦅🐊🐅🦍 Model Zoo</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*vBSUIo2RtgaN74GJCGtePA.png" /><figcaption>CycleGAN</figcaption></figure><p>This web allows their users to discover open source deep learning code and pre-trained models. Creating and training models is hard and expensive, especially because it depends on the <em>21st century oil, data</em>. In an open source era it makes sense to first pay attention at what others have done and learned regarding the problem you are facing, good initiative!</p><p><a href="https://modelzoo.co/">https://modelzoo.co/</a></p><blockquote><strong>2. </strong>Speech Recognition with TensorFlow</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SIdUYSKkmyppC6c7EkSiTg.jpeg" /><figcaption>MathWorks Spectogram plot</figcaption></figure><p>Implementation of a seq2seq model for Speech Recognition using TF. Architecture similar to “<a href="https://arxiv.org/pdf/1508.01211.pdf">Listen, Attend and Spell</a>”. The model is a pyramidal bidirectional LSTMs in the encoder, which reduces the time resolution and enhances the performance on longer sequences.</p><p><a href="https://github.com/thomasschmied/Speech_Recognition_with_Tensorflow">thomasschmied/Speech_Recognition_with_Tensorflow</a></p><blockquote><strong>1. </strong>TensorFlow: The Confusing Parts (1)</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*jvfFzOk33n5AB6XluC5gvQ.png" /></figure><p>Part 1 from a series of posts that will help reader get a better intuition for what TensorFlow is, how it works, and how to use it. Although the concepts presented in this tutorial are fundamental to all TensorFlow programs, it is pretty easy to end up disliking TF due to having to learn to use it the <em>hard way</em>. Thus, I would recommend this read to any novice to intermediate user of TensorFlow.</p><p><a href="https://jacobbuckman.com/post/tensorflow-the-confusing-parts-1/">Tensorflow: The Confusing Parts (1) | Buckman&#39;s Homepage</a></p><p>And this is it for what we found out to be interesting in June. At xplore.ai, we are always trying out the latest tools, experimenting with cutting edge algorithms and reading about the latest trends in every industry where data is generating unprecedented value.</p><p><em>If you liked the article please clap and subscribe. You can also check the other articles in our </em><a href="https://medium.com/xplore-ai"><em>xplore.ai blog publication</em></a><em>. You can also follow us in </em><a href="https://www.linkedin.com/company/xplore-ai/"><em>LinkedIn</em></a><em> and </em><a href="https://twitter.com/xploreai"><em>Twitter</em></a><em>. We hope you have a great month ahead!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3dbf3ba71ba5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/xplore-ai/15-curated-ai-reads-for-july-2018-3dbf3ba71ba5">15 curated AI reads for July 2018</a> was originally published in <a href="https://medium.com/xplore-ai">xplore.ai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>