<?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 Michael Chen on Medium]]></title>
        <description><![CDATA[Stories by Michael Chen on Medium]]></description>
        <link>https://medium.com/@profound_thot?source=rss-cefa94c15a32------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*nsfGDUeBY9dWtFtGBW8XEQ.jpeg</url>
            <title>Stories by Michael Chen on Medium</title>
            <link>https://medium.com/@profound_thot?source=rss-cefa94c15a32------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:46:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@profound_thot/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[ML Deep Dive: Multitask Learning]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://www.towardsdeeplearning.com/ml-deep-dive-multitask-learning-a344e7745d53?source=rss-cefa94c15a32------2"><img src="https://cdn-images-1.medium.com/max/2600/0*ehUrxqzWLNBGmuCs" width="5120"></a></p><p class="medium-feed-snippet">Multitask learning is all the rage these days, being especially useful and widely adopted in modern ranking systems like ads, search, and&#x2026;</p><p class="medium-feed-link"><a href="https://www.towardsdeeplearning.com/ml-deep-dive-multitask-learning-a344e7745d53?source=rss-cefa94c15a32------2">Continue reading on Towards Deep Learning »</a></p></div>]]></description>
            <link>https://www.towardsdeeplearning.com/ml-deep-dive-multitask-learning-a344e7745d53?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/a344e7745d53</guid>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[multitask-learning]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Thu, 02 Jul 2026 22:40:38 GMT</pubDate>
            <atom:updated>2026-07-16T05:12:07.302Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ML Bytes: Cross Entropy Loss]]></title>
            <link>https://medium.com/@profound_thot/ml-bytes-cross-entropy-loss-25348fda2802?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/25348fda2802</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Sun, 31 May 2026 20:52:49 GMT</pubDate>
            <atom:updated>2026-06-01T03:10:04.523Z</atom:updated>
            <content:encoded><![CDATA[<p>Today, I want to talk about one of the most common topics in machine learning, starting with first principles.</p><p>What is loss in machine learning? First off, usually people mean “loss function” when referring to “loss”. Loss is like feedback for models. It tells a model what worked and what didn’t. Given some inputs, a model predicts some outputs. A loss function then tells the model how correct or incorrect it was and the model can then decide what to do, usually by going through gradient descent to learn from their mistakes.</p><p>What is cross entropy? Cross entropy is a measure of the difference between two probability distributions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xnrSUeeLYyXKvbjr.png" /><figcaption>applied to machine learning, it is most commonly used as a “loss function” to quantify how closely a model’s predicted probability distribution aligns with the true distribution of the data</figcaption></figure><p>While this visualization is helpful for building intuition, most machine learning classification problems operate on discrete probability distributions rather than continuous ones.</p><p>Imagine we are training an image classifier with four possible labels:</p><ul><li>Apple 🍎</li><li>Orange 🍊</li><li>Banana 🍌</li><li>Grape 🍇</li></ul><p>Suppose we feed an image of an apple into our model. The true distribution can be represented as a one-hot vector:</p><blockquote>P = [1, 0, 0, 0]</blockquote><p>This simply means there is a 100% probability that the correct answer is apple and 0% probability for the other classes.</p><p>Now imagine our model predicts:</p><blockquote>Q = [0.7, 0.2, 0.07, 0.03]</blockquote><p>The model believes there is a 70% chance the image is an apple, a 20% chance it is an orange, and a 10% chance it is a banana.</p><p>Cross entropy loss measures how closely the predicted distribution (Q) aligns with the true distribution (P). In practice, the loss is computed for each individual training example and then averaged across the dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1SMxMXKOPr4zo4d4hA5lmw.png" /></figure><p>Given all of this, let’s take a look at what cross entropy loss looks like as a formula.</p><p>Extending our example, say we have a dataset of 3 images to classify.</p><ul><li>Image 1 is an apple and the model predicted 90% likelihood it was an apple</li><li>Image 2 is an orange and the model predicted 0.8</li><li>Image 3 is a banana and the model predicted 0.7</li></ul><p>Assuming each image is independent of one another, the probability the model thinks we have an apple and an orange and a banana would be 0.9 x 0.8 x 0.7. We are basically trying to answer:</p><blockquote>How much probability did the model assign to the collection of events that actually occurred?</blockquote><p>This quantity is what we would call the likelihood.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*7IvdrYT62gIUMjqNmUhPhA.png" /><figcaption>The big pi symbol means we are multiplying across each instance i. For our example, it would mean multiplying across our dummy dataset of 3 examples. The p(y|x) portion means the probability assigned to the correct true label y_i. So boiled down for our example, it would be 0.9 * 0.8 * 0.7</figcaption></figure><p>Likelihood is how much probability the model assigned to the reality that occurred. We would want our model to predict high likelihood for the reality that occurred since that would mean the model is in line with what happened, that the model has some predictive power.</p><p>We are not quite at the traditional form of cross entropy loss however. The issue here is that if we have many examples and not just three, our multiplication would create a really large decimal value. Enter this math trick: take the log of the likelihood. This turns our long product of probabilities into a sum of probabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D3Bp_wGsTYgDIsWX0Pu6Iw.png" /></figure><p>One last step. The way we have framed things, we would want to maximize the log likelihood, since we want the model to assign high likelihood to the real outcomes. However for many machine learning algorithms, we actually want to minimize as that allows for gradient descent to optimize properly. So instead of maximizing log L, we would minimize the negative of log L.</p><p>Ok actually one more step. We currently sum the log probabilities but we actually want the average so we would need to divide by N, the total number of examples in our training set. Thus, our final form of cross entropy loss looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*k3vB8y0McJc6nldob9HR7Q.png" /><figcaption>We often average the negative log likelihood across all N training examples. It makes the loss independent of dataset size and easier to compare across experiments. It also keeps gradient magnitudes more consistent across different batch sizes during training.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7dwCKOAZbxaZ9g4dSWvmpA.png" /></figure><p>Recall our current setup has four classes: apple, orange, etc. This loss function is elegant in that it applies to both multi-class and binary classification. We only care about the true label of each training example, so it doesn’t matter how many classes we have. If the model is assigning probabilities to any class other than the true class, then it’s wrong and the loss will show it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lZgaMnq2nWVlNPuo-ZJbbw.png" /><figcaption>the special name is “Binary Cross Entropy” or BCE but the idea is still the same</figcaption></figure><p>If you look closely, there are two main sides to this equation and for any given data point, only one side is active at a time. If y = 1, the equation simplifies to — y * log(p) or log(p) of the true label of 1. When y = 0, the equation turns into — (1 — y) * log(1 — p) or log(1 — p) where p is still the probability of the true label being 1. Since there’s only two possibilities, our actual probability is the probability of the true label being a 0. Like the multi-class scenario, we really only care about the probability assigned to the true label of the datapoint. I guess for Binary Cross Entropy, to make the equation look cool, someone a long time ago decided it should always be about the positive 1 class and have the log terms be log(p) and log(1 — p).</p><p>That’s all folks! Happy learning~</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25348fda2802" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ML Bytes: Normalization]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@profound_thot/ml-bytes-normalization-a12cdb3ac77f?source=rss-cefa94c15a32------2"><img src="https://cdn-images-1.medium.com/max/1536/1*yRIdqHazBae-4zkwQnNK4g.png" width="1536"></a></p><p class="medium-feed-snippet">Whenever you train a machine learning model, normalization plays a big role. Whatever model you choose to represent the underlying problem&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@profound_thot/ml-bytes-normalization-a12cdb3ac77f?source=rss-cefa94c15a32------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@profound_thot/ml-bytes-normalization-a12cdb3ac77f?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/a12cdb3ac77f</guid>
            <category><![CDATA[normalization]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Tue, 05 May 2026 15:43:55 GMT</pubDate>
            <atom:updated>2026-07-09T00:33:28.830Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ML Deep Dive: DCN v1 vs DCN v2 — Explicit Feature Crossing for Modern Deep Learning Models]]></title>
            <link>https://medium.com/@profound_thot/ml-deep-dive-dcn-v1-vs-dcn-v2-explicit-feature-crossing-for-modern-deep-learning-models-eedec1810792?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/eedec1810792</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Thu, 26 Mar 2026 23:06:15 GMT</pubDate>
            <atom:updated>2026-11-02T04:55:14.777Z</atom:updated>
            <content:encoded><![CDATA[<p>Deep learning architectures have evolved over the years to best capture and represent the underlying data. Traditional multilayer perceptron networks capture nonlinearity. Convolutional networks were created to capture spatial information and recurrent networks were created to capture sequential information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3H1H4Yq4H6BKTzhv" /><figcaption>Photo by <a href="https://unsplash.com/@hharritt?utm_source=medium&amp;utm_medium=referral">Hunter Harritt</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>What about sparse information coming from large web scale systems? Imagine we have to model high cardinality user profile information. Deep learning models are often forced to learn important feature interactions implicitly. ML engineers can also do manual feature engineering on top of these sparse features. These approaches work but are inefficient. In many industrial ML systems, the hardest part of the problem is not raw nonlinearity — it is learning the right <em>crosses</em>: user × item, query × document, merchant × issuer, device × context, and so on.</p><p>This is where the Deep &amp; Cross Network family is useful.</p><p><a href="https://www.brainjuice.club/papers/47378b3f-bcf1-4318-86fa-abd0340723be">DCN v1</a> [1] introduced a simple and elegant mechanism for learning explicit bounded-degree feature interactions without manual feature engineering. <a href="https://www.brainjuice.club/papers/cd00b1b2-0c10-4a1a-84bd-3465141ab460">DCN v2</a> [2] kept the same core idea, but upgraded the cross layer to be much more expressive and practical at scale.</p><p>Deep &amp; Cross Network (DCN) layers are not exclusively used for ranking models, although they are primarily renowned for their success in recommendation systems and Click-Through Rate (CTR) prediction. While DCN was designed to efficiently learn explicit feature interactions — a common need in ranking — the architecture is label-agnostic and can be applied to any task requiring the modeling of complex feature crosses.</p><p>This article breaks down the difference, why it matters, and how to implement both cleanly.</p><h3>Why explicit feature crossing matters</h3><p>Suppose your model input is a concatenated dense vector of length d:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/198/1*s4UMDsPFvWyjybQ1VqeD-w.png" /></figure><p>This may include:</p><ul><li>dense numerical features</li><li>one-hot or multi-hot features mapped into embeddings</li><li>context features</li><li>learned user or item representations</li></ul><p>A plain MLP can in principle learn interactions among these features, but it does so implicitly. In practice, many tabular and ranking problems benefit from a more structured inductive bias: <em>explicitly construct cross terms</em> while keeping the model efficient.</p><p>That is the role of the cross network.</p><h3>DCN v1: elegant, cheap, but constrained</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*VGsLvWp7XwSer3d7NrIrGA.png" /><figcaption>image credit to the original paper authors [1]</figcaption></figure><p>I believe <strong><em>x’</em></strong> and <strong><em>x</em></strong> here are the same thing so I’m not sure why the paper authors decided to use this notation.</p><p>The central implementation idea is for the layer to keep track of both the previous layer’s output <strong><em>x’</em></strong> as well as the original input we started with <strong><em>x_0</em></strong>. We compute the output of a v1 cross layer by first applying a set of weights against the previous layer’s output, resulting in a scalar value to modulate the original input vector with. We then go ahead and scale the original input by that scalar and add that to the previous layer input. This becomes the input for the next cross layer.</p><p>This builds explicit polynomial-style feature interactions layer by layer.</p><h4>PyTorch implementation</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FuRfy1PgNNkjHmzuUebfDg.png" /><figcaption>basic DCNv1 cross layer</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4rmwD44vwcXkZBC4MOwT5g.png" /><figcaption>basic stacked cross tower</figcaption></figure><h4>What v1 gets right</h4><p>DCN v1 is attractive because each layer is extremely cheap:</p><ul><li><strong><em>O(d)</em></strong> parameters per layer for the cross term</li><li>explicit crossing without manual feature engineering</li><li>bounded interaction degree that grows with depth</li></ul><p>For many production systems, this is already a strong baseline.</p><h4>The limitation</h4><p>The weakness is in the structure of the cross transformation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/314/1*ttDRSK6XJAw_FFV93US2gw.png" /><figcaption>our cross function outputs a single scalar value</figcaption></figure><p>Each subsequent cross layer update is just some scaled version of the original input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/428/1*3GET62EG5KFZ5u4GNE3tqA.png" /><figcaption>the entire cross contribution is controlled by <strong>one scalar</strong>.</figcaption></figure><p>This is expressive enough to create useful cross terms, but it is also pretty restrictive. The layer cannot independently modulate different coordinates based on richer mixtures of the current representation.</p><p>This is why people often describe the v1 cross transformation as effectively <strong>rank-1</strong>.</p><h4>What “rank-1” means in practice</h4><p>The v1 cross update can be rewritten as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/424/1*NlqTllbHewOii0r6SdO5PA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1BtGTZ7HFrmyph9ghOW7mw.png" /></figure><h3>DCN v2: same idea, much stronger cross layer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/674/1*x62iRsg6r3iFobZ3PCMyqg.png" /><figcaption>image credits to original paper authors [2]</figcaption></figure><p>DCN v2 upgrades the cross layer by applying a matrix of weights rather than a vector of weights on the previous layer vector. As a result, the middle term now produces a <strong>full vector</strong> of interaction strengths instead of a single scalar. This is multiplied element-wise against the original input, allowing for more granular interaction.</p><h4>PyTorch implementation</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SI4vW1xHnEeM3Xs8r0ejUA.png" /></figure><p>And the stacked cross tower:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NaH3-2nPGtqIUAiq1na-Ig.png" /></figure><p>This is the cleanest basic implementation of a DCN v2 cross tower.</p><h3>Why DCNv2 is materially better</h3><p>The improvement changes the representational power of the cross layer.</p><p>In DCNv1:</p><ul><li>one scalar controls the whole cross update</li></ul><p>In DCNv2:</p><ul><li>each coordinate gets its own learned interaction strength</li><li>those strengths are derived from the full current representation</li></ul><p>So instead of saying:</p><blockquote>summarize the current state into one number, then rescale the base input</blockquote><p>v2 says:</p><blockquote>transform the current state into a full vector, then modulate the base input coordinate-wise</blockquote><p>That is a much richer inductive bias for ranking and tabular systems.</p><h4>Practical intuition</h4><p>If your input contains:</p><ul><li>user embeddings</li><li>item embeddings</li><li>context features</li><li>device or geography features</li></ul><p>then DCNv2 can learn that:</p><ul><li>coordinate 17 should be amplified when a certain user-item-context pattern appears</li><li>coordinate 42 should be suppressed for a different pattern</li><li>different parts of the representation can respond differently to the same example</li></ul><p>DCNv1 cannot do this nearly as flexibly because all coordinates share the same scalar gate.</p><h3>Low-rank DCN v2: the production-friendly version</h3><p>A full <strong><em>d × d</em></strong> matrix per layer can get expensive when <strong><em>d</em></strong> is large. DCN v2 therefore becomes especially practical when paired with <strong>low-rank factorization</strong>.</p><p>Instead of learning the weight matrix directly, approximate it as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/1*zeaYlNxk7eIZSJtoZQUiYA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/942/1*j9b-nPPz6mAvV-c0l0gIyg.png" /></figure><h3>Example</h3><p>If:</p><ul><li><strong><em>d = 1024</em></strong></li><li><strong><em>r = 64</em></strong></li></ul><p>then:</p><ul><li>full-rank layer: about 1,048,576 weights</li><li>low-rank layer: about 131,072 weights</li></ul><p>That is a massive reduction!</p><h3>PyTorch implementation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*17bUsr4AZYKg5CuYMde8vg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rJeH2bEkp33YK-tlMw2l6Q.png" /></figure><h3>Cross tower only vs full DCN model</h3><p>A common mistake is to implement only the stacked cross network and call that “DCN”. Strictly speaking, the cross tower is only one part of the architecture family.</p><p>In practice, DCN models usually combine:</p><ul><li>a <strong>cross tower</strong> for explicit feature interactions</li><li>a <strong>deep tower</strong> for nonlinear implicit interactions</li></ul><p>There are two main ways to do this.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/684/1*JZXLlKaQJJ0TPYMoHNL6Ig.png" /><figcaption>image credits to original paper authors [2]</figcaption></figure><h3>Parallel architecture</h3><p>Both towers consume the same raw input <strong><em>x_0</em></strong>:</p><ul><li>cross tower: learns explicit multiplicative interactions</li><li>deep tower: learns general nonlinear representations</li></ul><p>Then their outputs are concatenated and fed to a final prediction layer.</p><h4>PyTorch sketch</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7qF5Qw9ac5lfC-uYYhsXBQ.png" /></figure><h4>Why parallel is strong</h4><p>This is often the clean default:</p><ul><li>both branches see the raw signal directly</li><li>the cross branch specializes in explicit feature interactions</li><li>the deep branch specializes in implicit nonlinear structure</li><li>easier to reason about, debug, and ablate</li></ul><p>In most production ranking systems, this is the version I would start with.</p><h3>Stacked architecture</h3><p>The alternative is to run the deep tower on top of the cross tower:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/792/1*pk96VhN0LgxzHI-gL6g8pQ.png" /></figure><h4>PyTorch sketch</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ue6CCbQWaUURmu5s1xajEw.png" /></figure><h4>When stacked helps</h4><p>This architecture makes sense when you want the MLP to operate on already cross-enriched features.</p><p>The tradeoff is that the deep branch no longer sees the untouched raw input directly. That can be fine, but it couples the roles of the two towers more tightly.</p><p>My bias is:</p><ul><li><strong>parallel</strong> as the default</li><li><strong>stacked</strong> as an empirical alternative worth testing</li></ul><h3>What kinds of problems benefit most from DCN-style models</h3><p>DCN is not a universal architecture. It shines in domains where:</p><ul><li>the input is tabular or embedding-heavy</li><li>feature interactions matter a lot</li><li>manual crosses would be painful or incomplete</li><li>you care about parameter efficiency relative to widening an MLP</li></ul><p>The best fits are usually:</p><ul><li>CTR prediction</li><li>CVR prediction</li><li>recommendation ranking</li><li>search ranking</li><li>marketplace ranking</li><li>ads systems</li><li>risk/fraud models with lots of heterogeneous categorical features</li></ul><p>These are exactly the settings where “feature A matters differently depending on feature B and feature C” is the norm, not the exception.</p><h3>What DCN is not trying to be</h3><p>It is worth stating clearly: DCN is not a replacement for every deep model.</p><p>If your input is primarily:</p><ul><li>raw text</li><li>images</li><li>audio</li><li>sequences where a Transformer is the main bottleneck</li></ul><p>then DCN is usually not the primary lever.</p><p>It is best thought of as a <em>structured interaction layer for tabular / sparse-feature systems</em>, especially when embeddings are already part of the representation.</p><h3>Practical guidance from a production perspective</h3><p>If I were building a ranking model today, my decision flow would be:</p><p>Start with:</p><ul><li>strong embedding setup</li><li>plain MLP baseline</li><li>DCN v2 parallel variant</li><li>low-rank cross layers if dimension is large</li></ul><p>Tune:</p><ul><li>number of cross layers</li><li>bottleneck rank</li><li>deep tower width/depth</li><li>feature normalization and embedding regularization</li></ul><p>Compare against:</p><ul><li>deep-only baseline</li><li>cross-only baseline</li><li>parallel vs stacked</li><li>full-rank vs low-rank cross layers</li></ul><p>The most common failure mode is not that DCN is wrong — it is that the surrounding feature pipeline or embedding design is weak, and the architecture gets blamed for it.</p><h3>That’s all folks</h3><p><a href="https://brainjuice.club/papers/47378b3f-bcf1-4318-86fa-abd0340723be">DCN v1</a> introduced explicit feature crossing through a cheap rank-1 update:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/554/1*wpNx7WBxeVqtePRMQWozJQ.png" /></figure><p><a href="https://www.brainjuice.club/papers/cd00b1b2-0c10-4a1a-84bd-3465141ab460">DCN v2</a> kept the same recursive crossing idea but replaced the scalar gate with a vector-valued transform:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/570/1*BJITCWzsQ0YADI4cU9WTwQ.png" /></figure><p>This change matters because it lets the model learn much richer coordinate-wise interaction structure. In practice, this makes v2 a better fit for industrial ranking and recommendation systems, especially when paired with low-rank factorization and a parallel deep tower.</p><p>DCN v1 is still a good conceptual baseline. DCN v2 is the version I would actually reach for. Good luck training your new models!</p><p><strong><em>BONUS</em></strong>: Google also made a nice open source tutorial on the subject, if you are interested.</p><p><a href="https://colab.research.google.com/github/tensorflow/recommenders/blob/main/docs/examples/dcn.ipynb">https://colab.research.google.com/github/tensorflow/recommenders/blob/main/docs/examples/dcn.ipynb</a></p><p>Thanks for reading. All thoughts are originally my own, now edited and assisted with ChatGPT.</p><h4>References</h4><p>[1] DCN v1 paper: <a href="https://brainjuice.club/papers/47378b3f-bcf1-4318-86fa-abd0340723be">https://brainjuice.club/papers/47378b3f-bcf1-4318-86fa-abd0340723be</a></p><p>[2] DCN v2 paper: <a href="https://www.brainjuice.club/papers/cd00b1b2-0c10-4a1a-84bd-3465141ab460">https://www.brainjuice.club/papers/cd00b1b2-0c10-4a1a-84bd-3465141ab460</a></p><p>Read these papers and more on <a href="http://brainjuice.club/">brainjuice.club</a>! Track what you’ve read, ask questions about papers, tackle daily challenges, and more! 🧠🧃</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eedec1810792" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning Adventures: Teaching a Recommender System to Understand Outfits]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@profound_thot/machine-learning-adventures-teaching-a-recommender-system-to-understand-outfits-273dc27f35b0?source=rss-cefa94c15a32------2"><img src="https://cdn-images-1.medium.com/max/2600/0*Awvgd5kF1E9_qhBC" width="6000"></a></p><p class="medium-feed-snippet">Let&#x2019;s say we&#x2019;re working at yibay, a clothing marketplace that showcases standalone clothing items. Our fine product manager comes over one&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@profound_thot/machine-learning-adventures-teaching-a-recommender-system-to-understand-outfits-273dc27f35b0?source=rss-cefa94c15a32------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@profound_thot/machine-learning-adventures-teaching-a-recommender-system-to-understand-outfits-273dc27f35b0?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/273dc27f35b0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[product-development]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Tue, 10 Mar 2026 22:21:11 GMT</pubDate>
            <atom:updated>2026-03-10T22:21:11.430Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Future of Transportation]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@profound_thot/the-future-of-transportation-717bb80c554d?source=rss-cefa94c15a32------2"><img src="https://cdn-images-1.medium.com/max/945/0*MGbzRyvyXRZxc0Xu" width="945"></a></p><p class="medium-feed-link"><a href="https://medium.com/@profound_thot/the-future-of-transportation-717bb80c554d?source=rss-cefa94c15a32------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@profound_thot/the-future-of-transportation-717bb80c554d?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/717bb80c554d</guid>
            <category><![CDATA[transportation]]></category>
            <category><![CDATA[urban-planning]]></category>
            <category><![CDATA[public-transit]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Mon, 27 Oct 2025 05:01:14 GMT</pubDate>
            <atom:updated>2025-11-02T20:57:08.778Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Guide to the Alternative Fuel Vehicle Refueling Property Credit]]></title>
            <link>https://medium.com/@profound_thot/guide-to-the-alternative-fuel-vehicle-refueling-property-credit-3bfee553dcb9?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/3bfee553dcb9</guid>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[taxes]]></category>
            <category><![CDATA[ev]]></category>
            <category><![CDATA[tax-returns]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Sun, 31 Mar 2024 19:19:19 GMT</pubDate>
            <atom:updated>2026-06-08T21:43:18.486Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m sharing my experience with this new credit that was made available recently in 2023. If you’re like me and you think EVs are the future, then you probably eventually want to install an at-home charger for convenience&#39;s sake. Well thanks to this new credit, you can now get money back during tax season, <strong>up to 30% of the cost with a maximum of $1,000 per item/charger</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tP_a8s4ItL6MB5-Gh1pc5Q.png" /><figcaption>I made a diagram</figcaption></figure><p>The official IRS page for this credit can be <a href="https://www.irs.gov/credits-deductions/alternative-fuel-vehicle-refueling-property-credit">found here</a>. However, they do not make it clear as to who exactly qualifies and don’t link anything that would help a person figure out if where they install the charger is an “eligible census tract”. I want to share what I used and show people ahead of time if they are eligible.</p><h3>Step 1: Figure out your Census Tract GEOID</h3><p>If you head on over to <a href="https://www.cdfifund.gov/cims">this web page created and run by the US Department of Treasury</a>, you’ll see six (at the time of this writing) options for the mapping tool. You want to click on <strong>CDFI</strong>. You’ll come up on a map view with some empty space on the left-hand side. There are also some buttons for navigation at the top of this empty space. You’ll want to press <strong>LAYERS</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IjOt3BqrF98YUoejMt3jCA.png" /></figure><p>Under this dropdown, you can select either <strong>2015 CDFI Tract</strong> or <strong>2020 CDFI Tract</strong>. As of this writing, both work for calculating if your home is within an “eligible census tract”. Once selected, you can then proceed to search for your address in the top left addressbar. That should pull up your home and census tract. I believe it’s called the <strong>Census Tract FIPS</strong> number and it can also be known as the <strong>Census Tract GEOID</strong>.</p><h3>Step 2: Figure out if your census is eligible</h3><p>If you use the 2015 option under LAYERS from step 1 above, you must use this list to see if your neighborhood qualifies. This list contains the census tracts by eligibility according to the 2011–2015 NMTC tracts.</p><h4>2015 census tract eligibility list</h4><p><a href="https://www.irs.gov/pub/irs-drop/appendix-a-list-of-2015-census-tract-boundary-30c-eligible-tracts-v2-1-4-2024.pdf">https://www.irs.gov/pub/irs-drop/appendix-a-list-of-2015-census-tract-boundary-30c-eligible-tracts-v2-1-4-2024.pdf</a></p><p>You can keep using this for the tax years of 2023 and 2024.</p><p>If you use the 2020 option under LAYERS from step 1 above, you must use this list to see if your neighborhood qualifies. This list contains the census tracts by eligibility according to the 2016–2020 NMTC tracts.</p><h4>2020 census tract eligibility list</h4><p><a href="https://www.irs.gov/pub/irs-drop/appendix-b-list-of-2020-census-tract-boundary-30c-eligible-tracts-v2-1-4-2024.pdf">https://www.irs.gov/pub/irs-drop/appendix-b-list-of-2020-census-tract-boundary-30c-eligible-tracts-v2-1-4-2024.pdf</a></p><p>You can use this for all tax years from now through 2029.</p><h3>Step 3: Fill out the form in your tax software</h3><p>The official form is Form 8911. I used Turbotax this year and I pulled this information out from their system. Most decent tax software will have the option for you to claim this credit.</p><p>Good luck with your taxes~</p><blockquote><strong>DISCLAIMER</strong>: not tax advice, just trying to share my experience!</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3bfee553dcb9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hands-on with Transformers]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/attenchen-to-detail/hands-on-with-transformers-d514720741d7?source=rss-cefa94c15a32------2"><img src="https://cdn-images-1.medium.com/max/2600/0*L8lbiEzUAb9MCoBY" width="3999"></a></p><p class="medium-feed-snippet">Deep dive into a not-so-large language model use case</p><p class="medium-feed-link"><a href="https://medium.com/attenchen-to-detail/hands-on-with-transformers-d514720741d7?source=rss-cefa94c15a32------2">Continue reading on Attenchen to Detail »</a></p></div>]]></description>
            <link>https://medium.com/attenchen-to-detail/hands-on-with-transformers-d514720741d7?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/d514720741d7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Sat, 05 Aug 2023 20:00:38 GMT</pubDate>
            <atom:updated>2023-08-05T20:00:38.384Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[My First Week @ Roblox]]></title>
            <link>https://medium.com/@profound_thot/my-first-week-roblox-fd86d1484ddc?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd86d1484ddc</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[metaverse]]></category>
            <category><![CDATA[roblox]]></category>
            <category><![CDATA[machine-learning-engineer]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Sat, 12 Nov 2022 22:06:25 GMT</pubDate>
            <atom:updated>2026-06-08T21:43:36.106Z</atom:updated>
            <content:encoded><![CDATA[<h4>Sharing my experience as a new hire. If you’re interested, <a href="https://careers.roblox.com/">Roblox is hiring</a>! DM for intro and referral.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vM9gADY20G5z44TL" /><figcaption>Photo by <a href="https://unsplash.com/@ninjason?utm_source=medium&amp;utm_medium=referral">Jason Leung</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Q: What’s your take on what Roblox is doing? Tell me about the company</h3><p><em>Roblox is reimagining the way people come together.</em> The company goal is to create a unique metaverse experience that intertwines all of the games that can be played on the platform.</p><p>I think the approach is much more sensical than Meta’s approach to VR and the metaverse. It would be so cool if one day Roblox could build out a world akin to that of Ready Player One, except hopefully less dystopian.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FwBYjsROAy5w&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DwBYjsROAy5w&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5172dea7d90f988c96a7e342aefc1e81/href">https://medium.com/media/5172dea7d90f988c96a7e342aefc1e81/href</a></iframe><h3>Q: What team are you on?</h3><p>I am on the Economy ML team. We serve all the organizational needs related to machine learning for the overarching Economy organization. Our responsibility is to foster healthy Robux spend and monitor this “virtual GDP”. Some aspects of this include building out better recommendation systems for avatar items and reducing fraud in our marketplace.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fopen.spotify.com%2Fembed%2Fepisode%2F4k4SC1O5N3TnnkiPWZLw72%3Futm_source%3Doembed&amp;display_name=Spotify&amp;url=https%3A%2F%2Fopen.spotify.com%2Fepisode%2F4k4SC1O5N3TnnkiPWZLw72&amp;image=https%3A%2F%2Fi.scdn.co%2Fimage%2Fab67656300005f1f72d74e824fc359540e55cde2&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=spotify" width="456" height="152" frameborder="0" scrolling="no"><a href="https://medium.com/media/7bb1d0d519a2db9ee92165ca00b7923c/href">https://medium.com/media/7bb1d0d519a2db9ee92165ca00b7923c/href</a></iframe><h3>Q: How are your teammates?</h3><p>They are awesome! They come from all sorts of backgrounds and companies and it’s a pleasure to work with the breadth and depth of experience that the team has. My immediate project teammate that I see often in office is also super easy to work with.</p><h3>Q: What are some highlights during your first week?</h3><p>I found a group of guys to play basketball with and there’s a court that we reserve every Tuesday and Thursday! Parking is free and so is lunch. There’s also a gym located in one of the campus buildings. Yay convenience!</p><p>My team also organized an outing to the local mall. We had a nice lunch and then went to play some VR games at Sandbox VR.</p><h3>Q: What’s a thing you’ve noticed that’s odd that you’d want to improve?</h3><p>I’m not sure why we have to attest to getting a negative COVID test before going in every day. I also don’t know how to improve this situation except complain 😅</p><h3>Q: What are some of the unique perks of working at Roblox?</h3><p>You get lots of Robux! You also get ramp up time dedicated to play and even develop games on the platform. Some of the games blew my expectations away and the game editor felt relatively easy to work with.</p><h3>Q: What is something you are looking forward to?</h3><p>I’m looking forward to making a real impact on my team and on Roblox’s bottom line. As someone with an economics background, I’m excited to brainstorm and influence decision making on how we can improve Roblox’s in-game economy.</p><p>I am also excited to build out all things recommendations and machine learning for our org. I feel empowered to make a difference and can see the potential in what our team needs to build, from candidate generators to ranking algorithms.</p><p>Oh and there’s also the next Roblox week where Roblox flies all employees to the San Mateo HQ to celebrate and drink some company Koolaid!</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F9H11vNafpLw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D9H11vNafpLw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F9H11vNafpLw%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/5f38e684cbb405d246dd6469c46f6ec0/href">https://medium.com/media/5f38e684cbb405d246dd6469c46f6ec0/href</a></iframe><h3>Q: For folks who are interested in applying, what are some tips that you can provide on succeeding at the interview?</h3><p>Practice programming and Leetcode if you have to! The interview process for an ML eng is super similar to a traditional software eng. You’ll typically not encounter Leetcode hards as we aren’t trying to purposely trick you. For the system interviews, try to listen for what your interview wants to discuss more of and really dive deep into areas that you are an expert in. For senior folks, there is also a bar raiser interview where you’ll be asked to think creatively about improvements from a product and user perspective.</p><p>Also, I would recommend embodying some of the values Roblox upholds like taking the long view. It’s something the company and people at Roblox really care about.</p><h3>Q: How do I start the interview process with Roblox?</h3><p>If you’re super excited, please feel free to visit <a href="https://careers.roblox.com/">https://careers.roblox.com</a> and apply! Alternatively DM me with a position you are interested in and questions you may have and I can refer you.</p><p><em>All opinions are my own.</em></p><h4>Thanks for reading!</h4><h4>References:</h4><p><a href="https://open.spotify.com/show/6MEbDMxt2IXoTRvQegwmVf">Roblox Tech Talks podcast</a></p><p><a href="https://www.youtube.com/c/roblox/videos">Roblox official YouTube channel</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd86d1484ddc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ML Engineer Industry Knowledge]]></title>
            <link>https://medium.com/@profound_thot/ml-engineer-industry-knowledge-26b429adc5a0?source=rss-cefa94c15a32------2</link>
            <guid isPermaLink="false">https://medium.com/p/26b429adc5a0</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[industry]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[ml-so-good]]></category>
            <dc:creator><![CDATA[Michael Chen]]></dc:creator>
            <pubDate>Wed, 10 Aug 2022 18:04:13 GMT</pubDate>
            <atom:updated>2025-12-20T03:34:17.004Z</atom:updated>
            <content:encoded><![CDATA[<h3>Some ML Engineer Industry Knowledge</h3><p>Hello! I’m in between jobs currently and want to take some time to reflect on some of the interesting things I’ve picked up on the job.</p><h4>#1 — Monitoring your model performance entails many things!</h4><p>We want to monitor the model from a correctness standpoint and from a performance standpoint. From a performance standpoint, is the model returning in a reasonable amount of time? Are there any errors being thrown? From a correctness standpoint, is the output distribution in production the same as the output distribution during training? Have the distributions changed over time? Now let’s say that model output <em>has</em> changed over time. Let’s focus on monitoring features as a way to debug.</p><p><strong>1a:</strong> Sometimes a feature can change over time, often known as feature drift. One way we can try to detect this is to log the features at a particular point in time (eg. at the first of every month or at the first model evaluation every month). We can then compute the distance using a metric like the <a href="https://stackoverflow.com/questions/45086712/how-to-kullback-leibler-divergence-of-two-datasets">KL divergence</a> between the two dataset distributions. Depending on the sample size, we’ll be able to <a href="https://stats.stackexchange.com/questions/338016/what-constitutes-a-large-kl-divergence">gauge how significant the distance is</a> and whether or not there was feature drift.</p><p><strong>1b:</strong> Alternatively, we can train a classifier to differentiate between the two datasets. Each data row from dataset A will get a 0 and each data row from dataset B will get a 1. If we can train a classifier to effectively classify with a decently high AUROC value, then we can say there has been significant feature drift.</p><p><strong>1c:</strong> Lastly we can rely on a third party service like <a href="https://whylabs.ai/">WhyLabs</a> that integrates with your model hosting solution to allow for data drift monitoring. The idea behind this is that whatever data plugs into your model also gets fed into the third party service and then the third party service allows for computing the distribution difference through either KL divergence or an alternate metric like the <a href="https://en.wikipedia.org/wiki/Hellinger_distance">Hellinger distance</a>. (note: these third party services likely implements 1a or 1b under the hood)</p><h4>#2 — Despite what others may claim, XGBoost is still one of the most used models in industry.</h4><p>So… you should definitely spend some time to learn the inner workings of this algorithm and play around with it for yourself!</p><p>For the longest time, Square relied on primarily <a href="https://xgboost.readthedocs.io/en/stable/">XGBoost</a> models with some support around naive bayes and regression based models. In fact, many companies that I have interviewed at also heavily leverage XGBoost. It is <a href="https://stackoverflow.com/questions/34151051/how-does-xgboost-do-parallel-computation">super optimized</a> and able to leverage the many cores on one machine to help learn the branches of the tree model.</p><p><strong>2a:</strong> While XGBoost innately uses all available cores on a machine to independently train branches in a tree model, we can also leverage parallelization in order to tune the XGBoost model in search of the best hyperparameters. I wanted to clarify that these two are separate things that happen to both use the concept of parallelization. If you’d like to read more, here’s <a href="https://medium.com/@romain.serres/e64644e75804">a good article I found</a> that goes further in depth on this topic.</p><p><strong>2b:</strong> Here’s <a href="https://machinelearningmastery.com/gentle-introduction-xgboost-applied-machine-learning/">a more practical hands on tutorial</a> on using XGBoost. Try also playing with some of the hyperparameters!</p><p><strong>2c:</strong> To clarify and give credit to where credit is due, recently the central Square ML Platform team has done great work to support AWS SageMaker. This in turn supports many other model types, including deep learning models using both PyTorch and Tensorflow!</p><p>If you don’t have a central ML Platform team to rely on, feel free to sign up for the free tier of SageMaker for 2 months and <a href="https://docs.aws.amazon.com/sagemaker/latest/dg/xgboost.html">try it out yourself</a>!</p><h4>#3 — There are many cool distance measures that can be leveraged to accurately model the underlying problem.</h4><p>Take for example CashApp.</p><p>CashApp provides a Boost program for debit card holders. When you activate a Boost, you’ll get a reward for shopping at a particular merchant such as 10% back in Bitcoin at Starbucks.</p><p>Imagine now that we want to build a very basic recommendation engine. We want to recommend Boosts to users that they have not used. We can employ a basic collaborative filtering approach where we find other customers most similar to our customer.</p><p>First we need a way to represent Cash customers. Then we can compare our customer with all the other customers using some distance metric and get the top most similar customers. From there, we can then get a list of Boosts that the customer has not used and recommend to them those.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tmpq8N99C2sitZphSPhiBg.png" /><figcaption>To read more on these distances, be sure to check out <a href="https://towardsdatascience.com/9-distance-measures-in-data-science-918109d069fa">this medium article</a>!</figcaption></figure><p>So how should we represent our Cash customer? We can have a vector of ones and zeroes that represents whether or not a customer has used a particular Boost. The vector will be N Boosts large where N is the total number of available Boosts.</p><p>Once we have a vector representation of two Cash customers, we can decide on which distance metric to use. One classic approach is to use the Cosine similarity metric. Another approach would be to use the Jaccard distance between two customers’ sets of Boosts. I’ll leave it as an exercise for the reader to consider the various distance metrics above and compare and contrast the tradeoffs!</p><p>Thanks for reading~ I hope you learned something new! Please “drop a clap” and leave a comment if you’d like to read more about a specific topic and I’ll do my best to accommodate.</p><p><a href="https://medium.com/mlearning-ai/mlearning-ai-submission-suggestions-b51e2b130bfb">Mlearning.ai Submission Suggestions</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26b429adc5a0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>