<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Siobhan Cronin on Medium]]></title>
        <description><![CDATA[Stories by Siobhan Cronin on Medium]]></description>
        <link>https://medium.com/@siobhankcronin?source=rss-edea7e1819ec------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*9zdwUa57s7DzGY0fq40q4A@2x.jpeg</url>
            <title>Stories by Siobhan Cronin on Medium</title>
            <link>https://medium.com/@siobhankcronin?source=rss-edea7e1819ec------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 01:11:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@siobhankcronin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Redshift from the command line]]></title>
            <link>https://medium.com/data-science/redshift-from-the-command-line-5d6b3233f649?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d6b3233f649</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[redshift]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Thu, 04 Apr 2019 23:32:11 GMT</pubDate>
            <atom:updated>2019-04-05T01:40:16.566Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4wcDxA22CNsbX0RlEAVq8w.jpeg" /></figure><p>I recently found myself writing and referencing Saved Queries in the AWS Redshift console, and knew there must be an easier way to keep track of my common sql statements (which I mostly use for bespoke COPY jobs or checking the logs, since we use <a href="https://modeanalytics.com/">Mode</a> for all of our BI).</p><p>Turns out there IS an easier way, and it’s called psql (Postgres’ terminal-based interactive tool)! Getting started with psql is super easy, and you’ll rejoice in the amount of AWS console clicking it cuts out.</p><h4>Step 1: Install</h4><p>Start by <a href="https://www.postgresql.org/download/">installing postgres.</a></p><pre>$ brew install postgres</pre><h4>Step 2: Establish a Redshift connection</h4><p>Next, <a href="https://docs.aws.amazon.com/redshift/latest/mgmt/connecting-from-psql.html">connect to your Redshift cluster</a>.</p><p>Since you will be using psql ALL the time, I recommend creating an alias in your ~/.bash_profile so you can easily establish your database connection with a single word. I use redshift as my alias. You can use the credentials you’ve defined in your .bash_profile:</p><pre>alias redshift=&#39;psql &quot;host=$REDSHIFT_HOST user=$REDSHIFT_USER dbname=$REDSHIFT_DB port=$REDSHIFT_PORT sslmode=verify-ca     sslrootcert=&lt;path to your postgres root .crt file&gt; password=$REDSHIFT_PASSWORD&quot;&#39;</pre><h4>Step 3: Add your dev IP address</h4><p>You will need to ensure your cluster security group is set to receive inbound traffic from whatever IP address you’ll be using in development. If your cluster is in a custom VPC, you can do this from the command line using the <a href="https://docs.aws.amazon.com/cli/latest/reference/ec2/authorize-security-group-ingress.html">CLI</a>’s authorize-security-group-ingress. Otherwise, if you’re using the default VPC, you can add your IP address to the Inbound rules for the Security Group manually in the console</p><h4>Step 4: Explore your warehouse</h4><p>Now that you’re connected, type redshift on the command line, and try out these handy commands:</p><p><strong>\dt — </strong>view your tables<br><strong>\df — </strong>view your functions<br><strong>\dg — </strong>list database roles<br><strong>\dn — </strong>list schemas<br><strong>\dy — </strong>list event triggers<br><strong>\dp — </strong>show access privileges for tables, views, and sequences</p><h4>Step 5: Take a query for a spin</h4><pre>$ SELECT * FROM your_schema.your_table LIMIT 10;</pre><h4>Step 6: Run a simple transaction</h4><p>Open your favorite text editor and write a simple transaction (a series of sql — statements to be run as a single body of work, doing nothing if any one of them fails), for example:</p><pre>BEGIN; <br>INSERT INTO your_table VALUES (1), (2), (3), (4);<br>INSERT INTO your_table VALUES (5), (6), (7), (8);<br>INSERT INTO your_table VALUES (9), (10), (11), (12);<br>COMMIT;</pre><p>Save this transaction and run in the psql repl with the \i command:</p><pre>dev=# \i your_transaction.sql</pre><p>Have fun and let me know what you discover!</p><p>Siobhán is the Data Engineering Lead at <a href="http://www.landed.com/">Landed</a>. Landed is on a mission to help essential professionals build financial security near the communities they serve, and we invest alongside educators when they’re ready to buy homes in expensive markets. Sound awesome? <a href="https://landed.com/jobs">Join us</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d6b3233f649" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/redshift-from-the-command-line-5d6b3233f649">Redshift from the command line</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using  Mode definitions (and dbt) to bootstrap your data model]]></title>
            <link>https://medium.com/data-science/using-mode-definitions-and-dbt-to-bootstrap-your-data-model-9ea6c9691131?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ea6c9691131</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Wed, 06 Feb 2019 23:21:32 GMT</pubDate>
            <atom:updated>2019-02-07T00:16:41.611Z</atom:updated>
            <content:encoded><![CDATA[<h3>Using Mode definitions (and dbt) to bootstrap your data model</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Yi7dwj3mjSkDOggZRY1mA.png" /><figcaption>Penny model. Oh, if our network graphs were so simple.</figcaption></figure><p>At Landed, we track a variety of customer journeys through a variety of lifecycle stages across a variety of products. To explore all this variety, our teams rely on a collection of analytics dashboards to help them anticipate the needs of our customers and partners. The questions our teams ask change over time, which requires a flexible and responsive practice for building and updating data models. We recently went through a process of upgrading our data infrastructure, and asked ourselves “how might we define an agile sprint that will take us from raw data to a BI Platform MVP in a couple of weeks?”</p><p>We knew we were looking for a data modeling solution that would provide version control, a staging environment with the ability to define tests, and a nice community of practice to support and learn from. After shopping around, we settled on two tools that work nicely together as a pair, <a href="https://modeanalytics.com/">Mode Analytics</a> and Fishtown Analytics’ data build tool (<a href="https://www.getdbt.com/">dbt</a>). Mode is a SQL-driven BI Platform that supports R and Python, which means you can augment their internal charting with all your favorite libraries (seaborn, matplotlib, etc.). Dbt is a tool (and perhaps a state of mind?) that allows you to clean, denormalize, test, and pre-aggregate your data.</p><p>The goal of this post is to share the broad strokes of how we used these two tools together to get our BI Platform on its feet during a 4-week sprint.</p><p>Here’s the general approach we took:</p><ul><li>STEP 1: Confirm definitions and calculations (prep)</li><li>STEP 2: Create base tables (dev)</li><li>STEP 3: Document and define tests in dbt .yaml files (staging)</li><li>STEP 4: Schedule builds using dbt cloud (prod)</li><li>STEP 5: Give back to the open-source community (😄)</li><li>Rinse and repeat! (iterate)</li></ul><h4>STEP 1: Confirm Definitions and Calculations</h4><p>Following the sage advice of my database elders, I created a catalog of all of our properties as well as a list of the queries that drive our existing reports and current OKR metrics. This documentation became my source of truth during the build. A simple shared document where teams could track progress in real-time.</p><p>I was lucky to join Landed after a period of rigorous property definition and data cleaning, so our starting point for this process was a place of clarity and consensus. If you’re bootstrapping a model with less clarity, I encourage you to scale the project back until you arrive at the handful of properties and calculations everyone can agree upon (even if, at first, that’s justfirst_name and last_name). As you build your system out and have some success with your early charting projects (“Look, a list of all our customer names!”), this will inspire others to stabilize their logic and clean up their data so they can jump in the game.</p><h4>STEP 2: Create Base Tables with Mode Definitions</h4><p>Working from our list of queries I then created base tables for each team. These tables included only the properties needed to drive the reporting; a curated collection of VIP features. I defined these base tables using Mode Definitions, which are select statements you can reference in other queries (up to 10 layers of nesting!). These base tables were defined iteratively — as I wrote each query in Mode, I added the necessary properties I needed from our Redshift cluster.</p><p>These base tables ensure our users never query our raw data directly. Even though I can fully ensure that our users don’t have the ability to mutate the raw data (by virtue of our read-only integration with Mode), we still want a clean record of the data types and calculations serving our analytics, which these base tables provide. They are also where we define our aliases for each column, as our ETL from HubSpot to Redshift (which we use Stitch to coordinate) pulls in the internal names of our data properties, which differ from the field names our teams engage with. While most of our team members will not be exploring our SQL queries directly, it’s always a good idea to norm on terminology.</p><h4>STEP 3: Document tables in dbt .yaml files</h4><p>This is the part of our journey where dbt comes in. To be honest, it wasn’t entirely clear to me at the onset when we should switch from defining our model in Mode definitions to setting up our tests in dbt and building our materialized views. This is partly because it’s quite easy to use Mode definitions in your Mode reports, and I noticed a bit of inertia when confronted with the hump of transferring things over to dbt. When I remembered dbt’s test coverage tooling, I rallied!</p><p>We arrived at an internal rule of thumb which was to start shifting things over from Mode once we’d covered 90% of the queries in our initial plan, which we hit about three weeks in. This part of the process involved defining a .yaml file for each base table that included pithy descriptions for all properties, as well as test definitions, of which there are four flavors in dbt: not_null, unique, relationships, and accepted_values.</p><p>Here’s an example:</p><pre>version: 2<br><br>models:<br>  - name: customer_id<br>    description: Unique customer identifier <br>    columns:<br>      - name: uid<br>        tests:<br>          - not_null<br>          - unique<br>          <br>      - name: best_friend<br>        description: Customer&#39;s BFF<br>        tests:<br>          - relationships:<br>              to: ref(&#39;friends&#39;)<br>              field: uid<br>            <br>      - name: favorite_ice_cream_flavors<br>        description: Customer&#39;s favorite ice cream flavors<br>        tests:<br>          - not_null<br>          - accepted_values:<br>              values: [&#39;cherry&#39;, &#39;chocolate&#39;]</pre><p>It’s worth noting that Mode supports persistent derived tables (pdt), which replicate the materialized views functionality provided by dbt. I was drawn to dbt’s active open source community and testing methods, but if you’re already using Mode Analytics for your BI, you may want to look into pdt.</p><h4>STEP 4: Schedule builds using dbt cloud</h4><p>This is the stage we’re at now with Landed; scheduling our nightly dbt builds after our ETL integration. Since we integrate data from our CRM every night, we schedule our dbt cloud job to materializes our base table views directly after. This part is pretty straightforward, because by this point all the hard work of building and testing the model is done.</p><p>One thing to note, if you’re using Redshift you’ll probably want to use <a href="https://aws.amazon.com/about-aws/whats-new/2017/09/amazon-redshift-now-supports-late-binding-views-referencing-amazon-redshift-and-redshift-spectrum-external-tables/">late-binding views</a> so that you don’t hit any snags in your nightly build. Late-binding views bind tables at run-time, so you can drop a build a view that binds to another table without throwing any errors. I hit this when we set up our ETL with Stitch, and switched to late-binding and now everything runs fine.</p><h4>STEP 5: Give back to the open source community</h4><p>Whether picking off an issue for <a href="https://github.com/fishtown-analytics/dbt/issues">dbt</a> or writing up a blog post about what you built, it’s always nice to round out a build cycle by giving some love back to the community. Learning with others is one of the greatest joys of engineering, as is supporting projects that advance best practices in the field. Find a ticket, make some friends through the repo, and submit a PR!</p><h4>Rinse and Repeat!</h4><p>That’s it! Now comes endless cycles of iteration :)</p><p>Siobhán is the Data Engineering Lead at <a href="http://www.landed.com/">Landed</a>. Landed is on a mission to help essential professionals build financial security near the communities they serve, and we invest alongside educators when they’re ready to buy homes in expensive markets. Sound awesome? <a href="https://landed.com/jobs">Join us</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ea6c9691131" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/using-mode-definitions-and-dbt-to-bootstrap-your-data-model-9ea6c9691131">Using  Mode definitions (and dbt) to bootstrap your data model</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What's auto ML?]]></title>
            <link>https://medium.com/data-science/whats-auto-ml-b457d2710f9d?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/b457d2710f9d</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[product]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Fri, 11 May 2018 23:20:19 GMT</pubDate>
            <atom:updated>2018-05-13T14:18:10.470Z</atom:updated>
            <content:encoded><![CDATA[<h3>What’s auto ML?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*xsP9A6dV0TmxRIs-1xMJRw.jpeg" /><figcaption>The purpose of computing is insight, not numbers — Richard Hamming (photo by CoWin)</figcaption></figure><p>I recently told a friend I was looking into auto machine learning (ML) projects, and he asked me if that meant machine learning for cars, so … I thought I’d write up a brief post.</p><p>Auto ML services provide machine learning at the click of a button, or, at the very least, promise to keep algorithm implementation, data pipelines, and code, in general, hidden from view. If you mention this prospect to a crowd of engineers, some will say “hooray!” and some will say “oh no!”, and, I suppose until recently I found myself in the middle of those exclamations (“orayoh!”).</p><p>I first played with an auto ML platform last year during a demo at IBM. Coming from iTerm, Python, and vim, I found it strange to be so removed from the code. But I could see how processes that had already been vetted in R&amp;D could be streamlined with auto ML. Not every team is working on bleeding edge algorithm design, and so there is definitely room to optimize the process of making minor adjustments to already-verified models.</p><p>I could see how auto ML might situate itself in the life cycle of a growing ML/AI product, but could such a platform truly become “a data scientist’s best friend”? I wanted some other data points, so I reached out to my friend <a href="https://medium.com/u/1f7bd8723c4e">Pierre Forcioli-Conti</a> who was busy working on his startup <a href="http://workbenchdata.com/">Workbench</a>, and we sat together in his office on Market Street and played with sample data in his environment while chatting about workflow. His vision was more of a collaborative common workspace for teams (like Google Docs for data projects), and I remember being particularly excited about the data parsing and cleaning features. But, for whatever reason, I returned home to my tools with no clear plan to integrate auto ML into anything I was building (although I did sneak off and play with <a href="https://automl.github.io/auto-sklearn/stable/">auto-sklearn</a> and <a href="https://automl.info/tpot/">TPOT</a> a bit, just to see what they were up to).</p><p>Fast forward to this week, where I saw a post about <a href="https://www.h2o.ai/">H2O</a> from someone I trust in data science, <a href="https://github.com/ledell">Erin LeDell</a>. I followed her link to the <a href="http://h2o-release.s3.amazonaws.com/h2o/rel-wolpert/9/index.html">product</a>, and looking into the docs, I found a good amount of support for different languages and cloud integration, which was encouraging to see.</p><p>The install was painless, and the interface was approachable, situated at the intersection of data analysis packages like Pandas and modeling tools like <a href="http://scikit-learn.org/">scikit-learn</a>, <a href="https://www.tensorflow.org/">Tensorflow</a> and <a href="https://keras.io/">Keras</a>, with a Jupyter-like cascading stream of cells. Both working with the local host interface (putting together processing steps in a document called a “Flow”) and importing the software directly into the Python shell felt more intuitive and familiar than what I’d encountered with the ML tools on <a href="https://aws.amazon.com/machine-learning/">AWS</a> or <a href="https://datascience.ibm.com/">IBM’s Watson Studio</a>.</p><p>I did find myself running a clock in my head tracking “how long would this usually have taken me?”. I imagine one could learn the idiosyncrasies of H20 fairly quickly, and then from that vantage point gauge where the tool might best fit into one’s toolkit. I’m personally interested in hyperparameter optimization for neural networks, particularly for larger deep learning projects, and found myself imagining how that exploration might go. What kinds of bugs would I run into? What kind of support would I find? If H2O really could be one-stop shopping for data munging, EDA, modeling, and processing on GPUs, then I could totally see moving H2O more front and center on my desktop.</p><p>I worry that teams will use auto ML to hunt and peck for models that can deliver “results.”. Machine learning is an art as much as a statistical modeling discipline, and while I believe deeply in the creative exploration aspect of what we do, what kinds of guideposts and sanity checks should we stitch into auto ML processes to ensure we continue asking strong questions as our pipelines become more abstracted? Perhaps <a href="http://knowyourmeme.com/memes/clippy">Clippy</a> has a model-minded cousin who could pop up periodically and ask things like, “Does this data representation match your intuitions?”, or “Are we at risk of over-inflating the error here?”, or “How are you interpreting these performance results?”. Having access to the code alone doesn’t answers such questions, and perhaps abstraction will free up more capacity to ask such questions, but my concern is that over time such questions might start to generate responses such as, “I don’t know, that was button I was told to press”. This is not a caution for H2O alone, but for all auto ML enterprises, and all of us charged with ensuring we use mathematical instruments effectively.</p><p>I see H2O coming out ahead on cloud integration, which signals a prioritization of customers of a certain scale (or with plans of scaling), and they definitely have their ducks in a row by not assuming any particular stack (offering their service to Java, Python, and R communities alike). All in all, I feel more of a “hooray” than an “oh no” about auto ML, and look forward to seeing how teams will use these tools in their product cycle.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b457d2710f9d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/whats-auto-ml-b457d2710f9d">What&#39;s auto ML?</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cutting through the fog of AI hype]]></title>
            <link>https://medium.com/@siobhankcronin/cutting-through-the-fog-of-ai-hype-74bbd88d4cd5?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/74bbd88d4cd5</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[mit]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Sat, 21 Apr 2018 22:14:48 GMT</pubDate>
            <atom:updated>2018-04-23T19:15:32.479Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qgB669U93gzLiROpOYbDAg.jpeg" /><figcaption>MIT on a clear day when the trees were still toddlers</figcaption></figure><p>Yesterday I attended the MIT Club of Northern California’s AI conference, <strong>Industries of the Future | Future of Industries. </strong>I left feeling energized, and here’s why (a play by play of my journey through the multiple tracks of concurrent panel discussions):</p><p><strong>Zia Chishti </strong>(Chairman and CEO of<strong> </strong><a href="https://twitter.com/afiniti"><strong>@</strong>afiniti</a>) kicked things off with a grounded reminder that AI buzzwords are often simply algorithms from the canon wrapped in new packaging. What’s changed the game in recent years is computational speed, and the heightened visibility of breakout gains made by novel applications. Chishti encouraged us all (a crowd of investors, developers, and business execs) to be leery of smoke and mirrors, focus on sticky, necessary problems, and always seek to apply the right tools for the job, instead of implementing ML models just to tick some AI box.</p><p><strong>Sandra Carrico </strong>(<a href="https://twitter.com/wattzon"><strong>@</strong>wattzon</a>), who has definitely helped teams find novel applications of AI in industry, flagged places one might look for such opportunities. These included <strong>where things are changing a lot in the code</strong>, <strong>where heuristics are applied (and therefore where dials could perhaps be autotuned)</strong>, <strong>where there are high maintenance costs</strong>, and <strong>where staffing levels increase</strong>. Across the board she advised us to direct AI solutions to the “heart of operations and profitability”, as we can build trust internally for new approaches to solving problems when we directly address the bottom line. This theme continued throughout the day— building trust in AI solutions within the organization by starting with small, easy-to-explain problems and securing visible wins.</p><p>After a brief introduction to the overall structure of the <a href="http://mitibmwatsonailab.mit.edu/">MIT-AI Watson Lab</a>, helmed by <a href="https://medium.com/u/412dfd9a98f8"><strong>David Cox</strong></a>, one could feel a collective pondering in the room as we fathomed the competing interests colliding in this lab situated between academia and industry. A vision of shared IP and collegial checks and balances seems to be propelling these early days of the lab (which launched last September), and I suppose time will tell how everything will play out. Tangentially (or perhaps relatedly, with MIT being, as Cox pointed out, a relatively conservative player in a game dominated by “move fast and break things” agents), this was one of the few speakers I heard who addressed security outright, referencing <a href="https://blog.openai.com/adversarial-example-research/">adversarial examples</a> in particular which have been used to disrupt machine vision.</p><p>One company’s business model, <a href="https://manifold.ai/">Manifold</a>, represented by <strong>Vinay Mohta,</strong> seemed particularly interesting in these times of change. As a firm, Manifold offers intermediate ML staffing with their team of partners, as well as a catalyst for articulating ML staffing needs more broadly. Like Carrico, Mohta encouraged us to justify AI research projects in terms of future returns, and to build trust internally by “creating pull with smaller problems”. Their model is surely replicable by others, especially those able to translate ML solutions to specific domains, and speaks to a wider need for engineering firms dedicated to serving the field during times of systemic change (which applied ML could be argued to be heralding).</p><p>I was excited to hear <a href="https://medium.com/u/8824cc4462ab"><strong>Regina Clewlow</strong></a> speak about her work with <a href="http://www.populus.ai/">Populus AI</a>, and I’ve spent this morning reading through her meta-analysis of ride-sharing programs, <a href="http://www.reginaclewlow.com/pubs/2017_UCD-ITS-RR-17-07.pdf">Disruptive Transportation: The Adoption, Utilization, and Impacts of Ride-Hailing in the United States</a>. I’ll not got into that work here, but it is chuck full of analysis that enriches and complicates the narratives spun about the benefits of ride-sharing, including such nuggets as the following, “we find that 49% to 61% of ride-hailing trips would have not been made at all, or by walking, biking, or transit.” Populus AI is positioning itself as a data and intelligence provider for the fleets of of autonomous vehicles that are on their way, and I believe getting this right will play a critical role in smart city programs.</p><p><a href="https://medium.com/u/773870e259f6"><strong>Cynthia Yeung</strong></a> of <a href="https://cafexapp.com/">Cafe X </a>offered some solid insights into how value is generated and transferred to customers within any human robot interaction (HRI). This includes value stemming from the agency we ascribe to robotic motion, and how this can activate our genetically hardwired and socially conditioned sentiments. When the robot barista waves, we wave back. And when they hand us our cup of coffee, we feel something special as we stand there holding our caffeine-fix that, just moments ago, was meticulously prepared for us by a robot. But it’s not just Disneyland dazzle. The brew precision (specified by the rosters) and speed (you can order ahead through the company’s app) will undoubtedly keep customers coming back after the novelty dies down. I personally wouldn’t mind waiting a bit longer if there were some dance moves (maybe even from the Pepper repertoire?**).</p><p><a href="https://medium.com/u/2ca1f00f51c7"><strong>Ayesha Khanna</strong></a> of <a href="https://addo.ai/">Addo</a>, an AI advisory and incubator, weighed in on AI from a global perspective, with an emphasis on an axis of developing cities in Asia and the Middle East seeded by Chinese investment (“the <em>new</em> silk road”). Many nations currently in the shadow of the two AI monoliths, USA and China, are hoping to leapfrog by positioning themselves in this growing market. In particular, Khanna has been working Singapore (which aims to be the “Switzerland of AI”), where she is encouraging accountability and transparency. She also shared a compelling story of AI being used to accelerate the incarceration of a man who had raped and killed a young girl in Pakistan, and I was grateful to hear such storysharing, as it provided a concrete example of AI being leveraged to address real community concerns.</p><p>And finally, <a href="https://medium.com/u/429198c4c506"><strong>Greg R.</strong></a>of <a href="https://xbrain.io/">xBrain</a> (whose teams “build the technology and design the solutions that are making cars smarter and the road safer”), cut across the grain of nationalistic score tallying and invited us to consider civilization at large, the global humanity, as we work together to create more abundance. Not replacing people, but augmenting people to create a more capable populace. Renard’s last words of the session summed up an idea that had been surfacing in me the entire day, “we’re clearly in the era of the augmented human”. Yes indeed.</p><p>** When Pepper dances, and, specifically when Pepper looks at Pepper’s hands while dancing, I feel I’m living in the future. All the future feels.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74bbd88d4cd5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Preparing for Autonomous Fleets]]></title>
            <link>https://medium.com/@siobhankcronin/preparing-for-autonomous-fleets-b0bf1dabe734?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0bf1dabe734</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[autonomous-cars]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[ux]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Fri, 10 Mar 2017 01:05:56 GMT</pubDate>
            <atom:updated>2017-03-10T23:59:55.139Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*T6XlJBwNhMJxZexT1a7HdQ.png" /></figure><p>Recently I have been diving deep into researching the possible futures of autonomous fleet movement dynamics. When the time comes, how are AVs going to flock? This research had led me to uncover many of the necessary advancements in machine perception, mesh networking, and deep learning that will provide for the ground-breaking AV opportunities we have been reading about in the headlines. Yet, I have also been turning over stones of possible modeling traditions that might help us prepare for the AV future.</p><p>As a movement nerd, what excites me most about self-driving cars is the possibility that their spatial navigation will be safer and more energy efficient than what has historically been achieved by humans. How might this look? A great example comes from the field of evacuation dynamics. Scientists who model crowd dynamics in emergency settings go to great lengths to factor in all the psychosocially-driven, yet spatially-inefficient, ways people move their bodies when they’re freaked out. For visuals, <a href="https://youtu.be/pqBSNAOsMDc?t=1m17s">watch</a> what kind of modeling is required to deal with “inter-agent avoidance”, and check out this great talk by <a href="https://www.youtube.com/watch?v=-A_1_vT1W24">Dirk Helbing</a>.</p><p>I know there are people out there who will forever want to put their hands on a steering wheel, but when I think about the population at large, it’s hard for me not to get excited about iterating on an existing system that is so riddled with human error. How will our autonomous fleets move when their navigation is unfettered by daydreaming, inebriation, and texting?</p><p>I have been seeking out companies who are beginning to unpack this question, and was excited to read Lyft’s recent Medium <a href="https://eng.lyft.com/">series</a> discussing their approach to solving data science problems. I believe their market simulator presents a great opportunity for developing optimal models of autonomous fleet navigation, although I am sure charting this horizon is just one of many interests for them.</p><p>Perhaps we can get a head start on the questions.</p><ul><li>How will we model autonomous vehicle dynamics? Specifically, how quickly will we scale the shift to machine-optimized motion planning in our models of autonomous fleet simulations?</li><li>How will autonomous vehicle dynamics affect urban design, which in turn will impact autonomous vehicle dynamics? How will we include such evolutions in our models?</li><li>How will increased navigation efficiency impact revenue models that rely on dynamic pricing?</li><li>How will we measure the degree to which passengers trust the safety of our fleet as it becomes more and more autonomous?</li></ul><p>Those that anticipate these changing dynamics will be better equipped to leverage their shiny new toys when they finally arrive. Which, with many in the industry elbowing towards 2021, may be sooner than some anticipated (and, of course, later than <a href="http://blog.caranddriver.com/future-past-self-driving-cars-have-actually-been-around-for-a-while/">others</a>).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0bf1dabe734" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Frontiers of Computational Psychiatry]]></title>
            <link>https://medium.com/@siobhankcronin/frontiers-of-computational-psychiatry-b238494265a3?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/b238494265a3</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[parkinson-disease]]></category>
            <category><![CDATA[health]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Mon, 31 Oct 2016 01:56:32 GMT</pubDate>
            <atom:updated>2016-11-18T01:14:07.420Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iFt3QTGwhJP0RWUkRukS_g.jpeg" /><figcaption>Image by author</figcaption></figure><p>When was the last time you sat down and read someone’s PhD thesis?</p><p>Given how much blood, sweat, and tears go into these things, I’m ashamed to admit I’ve only read a few.</p><p>However, I recently came across a thesis that was so well written and so well aligned with my interests that it felt as though I had stumbled upon a New York Times popular science bestseller, of the likes of Oliver Sachs or António Demásio. The thesis is by Thomas Viktor Wiecki, and is titled <a href="http://ski.cog.brown.edu/papers/wiecki_phd_thesis.pdf"><strong>Computational Psychiatry: Combining multiple levels of analysis to understand brain disorders</strong></a>.</p><p>Unicorns like these come along only so often, so in honor of Wiecki’s stellar work, I offer this brief intro to his research, lingering on both methods and the broader implications his work has for the fields of psychiatry and computational neuroscience.</p><p>Let’s begin!</p><h4>PART 1: CLUSTERING AND CLASSIFICATION</h4><p>To best understand why this research is such a breath of fresh air, one must hold in context the impact of DSM (the diagnostic and statistical manual of mental disorders) on focussing field-wide research targets, the shortcomings of which are best summed up in the author’s words:</p><p><em>Although primarily intended to be of value to clinicians, the DSM has also played a substantial role as a classification system for scientific research with the goals of validating the diagnostic categories and translating research results directly into clinical practice. Although these research goals are commendable, decisions regarding systematic classification are often more based on perceptions of clinical utility rather than scientific merit. As a consequence, DSM-based research programs have failed to deliver consistent, replicable, and specific results, and it has been widely observed that the validation of DSM categories has been limited, that DSM categories do not provide well-defined phenotypes, and that they have limited research utility.</em></p><p>One possible solution to this DSM pitfall is to create multidimensional profiles (MPs) from a series of various neuropsychological tasks. Rather than clumping patients together under crude delineations of symptoms, researchers can define neurocognitive phenotypes that provide a more nuanced lens for studying the origins, developmental stages, and response to specific treatments within meaningful subject clusters.</p><p>Wiecki gives a great example of this in relationship to ADHD (attention deficit hyperactivity disorder), citing a <a href="http://www.pnas.org/content/109/17/6769.abstract">paper</a> that used MPs to clusters patients, and found that these clusters did not cleave cleanly into groups of ADHD patients and Healthy Controls (HCs), but rather clustered in groups that included both. When a classifier was applied to these clusters, their analysis was better at predicting ADHD subjects than an analysis on the data as a whole. This finding underscores what we lose when we analyze data at the level of symptoms.</p><h4>DECISION MAKING: CASE STUDIES</h4><p>Wiecki goes on to provide concrete examples of how computational modeling can help us better understand psychopathology. As several brain disorders affect decision making, Wiecki provides examples from literature on Parkinson’s Disease (PD) and Schizophrenia (SZ) to illustrate how computational models have helped advance our understanding.</p><p>Current thinking on PD identifies the disorder as the “cell death of midbrain dopaminergic neurons in the substantia nigra pars compacta”. There are several visible symptoms that result from this degeneration, as it so fundamentally disrupts the motor system. However, the effects of PD on cognition can lead one to classify PD as an “action selection disorder” in which the development of the disorder can be tracked by performance on learning tasks. This is an example of how a model that takes into account behavioral markers can supplement existing models that track physical action; in this case, disruptions to the motor system.</p><h4>LEVELS OF COMPUTATIONAL PSYCHIATRY</h4><p>Cognitive tasks can help researchers zero in on what is happening in our brains, yet they are subject to the “task-impurity problem” — the problem of tasks producing results, yet not measuring what we think they are measuring. The limitations of traditional practices in fitting computational models in psychiatric contexts extend from fitting models to individual subjects as well as estimating parameters with an aggregated dataset from multiple subjects. The challenge with the former lies in masking any potential similarities of parameters, while the latter masks individual differences that are the actual research targets. That’s where <a href="https://en.wikipedia.org/wiki/Bayesian_hierarchical_modeling">hierarchical Bayesian</a> estimation can come to the rescue, with its ability to quantify uncertainty and increase sensitivity in parameter estimation.</p><p>“Computational models try to deconstruct behavior into its individual components and identify separable cognitive processes”, and this is where the richness lies. Even if null results are produced, I believe the sheer act itself of breaking phenomena down into more discreet cognitive processes is an advance for neuroscience. Such clustering techniques are being demonstrated to be better predictors of subject differences than aggregate performance scores. This is fertile terrain for additional research.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b238494265a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[25 years of inflated fMRI false-positive rates]]></title>
            <link>https://medium.com/@siobhankcronin/25-years-of-inflated-fmri-false-positive-rates-c60d026302df?source=rss-edea7e1819ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/c60d026302df</guid>
            <category><![CDATA[science]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[open-data]]></category>
            <dc:creator><![CDATA[Siobhan Cronin]]></dc:creator>
            <pubDate>Wed, 13 Jul 2016 21:39:13 GMT</pubDate>
            <atom:updated>2016-07-14T01:41:18.870Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*13Mc8GETG8sxoGN55QCEcg.png" /><figcaption>Toshiba 3T scanner drawing by author</figcaption></figure><p>While statistical analysis forms the backbone of experimental research, it is not often that statisticians step forward as the heroes in our innovation narratives. Yet here we have a study that boldly questions the statistical premises that underscore the validity of over two decades of fMRI research.</p><p><strong>Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates. </strong>Proceedings of the National Academy of Science. May 17, 2016 [<a href="http://www.pnas.org/content/early/2016/06/27/1602413113.full">Full Text</a>]</p><p>This study calls attention to false-positive rates embedded in the leading neuroimaging software platforms (SPM, FSL, AFNI). In particular, the results call into question the validity of spatial extent reported in results from research studies using these platforms, which the authors estimate at a staggering 40,000 studies. In fact, the authors assert that in the 25 year history of fMRI research, this is the first study to validate its core statistical methods using real data.</p><p>As a quick refresher, false-positive rate is the expectancy of the false positive ratio, and is arrived at by dividing the number of false positives (FP) by the sum of false positives (FP) and true negatives (TN). The shorthand, we typically say the false-positive rate is the probability of falsely rejecting the null hypothesis. For instance, if our null hypothesis states there are is no difference between the taste of apples and apples, the false positive data would boost the probability we would reject the null, leading us to believe there <em>are</em> indeed differences between the taste of apples and apples.</p><p>The authors of this study performed 3 million task group analyses in a study of 499 healthy subjects, and where they expected to find a false positive rate of 5%, they found that the most commonly software packages (SPM, FSL, AFNI) resulted in false-positive rates of up to 70%.</p><p>So, what are these statistical methods anyway, and why was this overlooked?</p><p>Well, they have to do with the spatial autocorrection of these software platforms, and the assumptions required to justify the use of Gaussian random-field theory (RFT) for FWE-corrected voxelwise and cluterwise inferences. Specifically, these assumption are that the 1) “spatial smoothness of the fMRI signal is constant over the brain”, and 2) “the spatial autocorrelation function has a specific shape (a squared exponential)”. For further reading on these assumptions, the authors cited Hayasaka &amp; Nichols’ 2003 article in Neuroimage, <a href="http://www.ncbi.nlm.nih.gov/pubmed/14683734">Validating cluster size inference: Random field and permutation methods.</a></p><p>The authors’ observed “heavy tails” of false clusters in resting-state fMRI data that appeared to become more pronounced after spatial smoothing. They analyzed these clusters to see if they appeared “randomly in the brain”, which would confirm the spatial smoothness assumption. However, when they generated their maps of voxelwise cluster frequency they showed the posterior cingulate to be most clustered and white matter to be least covered. The identification of this posterior cingulate “hot spot” led them to state “that violation of the stationary smoothness assumption may also be contributing to the excess of false positives”. The study goes on to test if false clusters are equally inflated in task-based fMRI data, which they indeed found to be the case using four task datasets downloaded from <a href="https://openfmri.org/">OpenfMRI</a>.</p><p><strong>What do we make of this?</strong></p><p>The research community will undoubtedly debate this article’s methods in the months to come as it works to fully digest its implications, and my hope is that additional statisticians will weigh in as well as the software companies themselves.</p><p>In the meantime, here are few talking points we can chew on:</p><ul><li><strong>Open data culture</strong> continues to provide researchers with the data access they need to conduct the meta-analyses necessary to validate research paradigms and such field-wise statistical analyses. Such work strengthens science as a whole, yet the commodification and siloing of knowledge by institutions continues to thwart this progress. What will it take to shift the economics of research in favor of open data?</li><li>This article begs the question — how can we in the the research community <strong>take responsibility to validate the code and address bugs in the proprietary software we use</strong>? Is the information we need too sewn up by the companies designing and selling this software? Could further developing open source research tools provide us with greater agency and accuracy? In the meantime, how common is it for researchers to publish the platform version they used in their analysis (which could be useful when bugs are identified that could affect previously published data)?</li><li>If we are doing our science right, our work will invariably complicate the research we conducted ten years ago, and should definitely complicate the research of our predecessors. Tools advance in their precision. Knowledge advances in its nuanced treatment of complex phenomena. Statistical analysis and computational strength grow up alongside our growing disciplines. How might we build toward a more robust<strong> read-write-<em>revise</em> global science culture? </strong>One that is comfortable with the historical erosion of certainty that the advance of knowledge demands? How can this been done in a climate where funding and job security are so intimately connected to research outcomes?</li></ul><p>These are horizons for continued discussion and collaboration. In the meantime, the authors point to some great projects I encourage you to check out, including the OpenMRI project already stated, as well as the <a href="http://fcon_1000.projects.nitrc.org/">1,000 Functional Connectomes Project</a> and <a href="http://neurovault.org/">NeuroVault.org</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c60d026302df" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>