<?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 João on Medium]]></title>
        <description><![CDATA[Stories by João on Medium]]></description>
        <link>https://medium.com/@judasarchive27?source=rss-07ef35202b34------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*JIwUbP1bMuWo4iBoazoJTw@2x.jpeg</url>
            <title>Stories by João on Medium</title>
            <link>https://medium.com/@judasarchive27?source=rss-07ef35202b34------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 03:33:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@judasarchive27/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[The Multivariable Chain Rule: The Mathematical Heart of Backpropagation]]></title>
            <link>https://medium.com/@judasarchive27/the-multivariable-chain-rule-the-mathematical-heart-of-backpropagation-ffca19642931?source=rss-07ef35202b34------2</link>
            <guid isPermaLink="false">https://medium.com/p/ffca19642931</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[analysis]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[João]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 13:52:06 GMT</pubDate>
            <atom:updated>2026-10-15T13:52:06.676Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why neural networks “add the gradients,” and how a 19th-century calculus result powers modern deep learning.</em></p><h3>Introduction</h3><p>Every time a neural network updates its weights, it is applying a precise mathematical result from multivariable calculus: the <strong>chain rule</strong>. In one dimension the idea is simple. In many dimensions it becomes the reason back-propagation works at all — and specifically why gradients are <em>added</em> when they flow through shared weights or branching paths.</p><p>This article gives a clean statement of the multivariable chain rule, a short proof from the definition of differentiability, and a direct mapping onto the computational graphs used in neural networks. The goal is to make the “why” as clear as the “how.”</p><p>(This is the first piece in a series on the mathematical foundations of AI. Later articles will move into optimization, information theory, cryptography in Web3, and beyond.)</p><h3>A Quick Reminder: The Single-Variable Case</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2AHrJZUaUVOqoCtsOQwltA.png" /></figure><p>The derivative of a composition is the product of the derivatives. Back-propagation is exactly this idea applied repeatedly, layer by layer.</p><h3>The Multivariable Chain Rule</h3><p>Now let the outer function depend on several intermediate variables, and let those intermediates themselves depend on several inputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D5aUZN_t6wjY7wIBnR_wDw.png" /></figure><h3>Proof from the Definition of Differentiability</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rfBEz6pNREVVcxS7EeAUeg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WTXJOkK5wP_pbWi54zMPKg.png" /></figure><p>In coordinates the composition of linear maps is matrix multiplication of Jacobians, which expands into the familiar sum of products of partial derivatives. That sum is why, in back-propagation, we <em>add</em> gradient contributions that arrive along different paths.</p><h3>From Calculus to Neural Networks</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LKbiYBas1xHbOuqkW_YJIw.png" /></figure><p>Back-propagation is reverse-mode automatic differentiation: it evaluates exactly this expression by flowing the upstream gradient backward and <em>accumulating</em> (adding) every contribution that reaches a given parameter.</p><p><strong>Tiny concrete example.</strong><br> Suppose a weight is used twice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*klicRUkzN_2r4SnVsMu1JQ.png" /></figure><p>Omitting the addition would give the wrong gradient. The same logic scales to every weight matrix, every bias, and every residual connection in a modern network.</p><p>In matrix form the same idea appears as the outer product</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nVs8E44zOGgFPaaoiGwrMA.png" /></figure><h3>Why the Mathematical Detail Matters</h3><ul><li>It explains <em>why</em> frameworks must accumulate gradients rather than overwrite them.</li><li>It shows that vanishing/exploding gradients are consequences of repeated multiplication of Jacobians, not mysterious numerical accidents.</li><li>It makes clear that residual connections, batch normalization, and careful initialization are attempts to keep those intermediate derivatives well-behaved so the summed chain-rule products remain informative.</li><li>It is the same principle that later appears in higher-order derivatives, natural gradients, and second-order optimizers.</li></ul><p>Understanding the chain rule at this level turns back-propagation from a black-box algorithm into a transparent application of classical calculus.</p><h3>Closing</h3><p>The multivariable chain rule is the precise reason a neural network can credit (or blame) every parameter for its contribution to the final loss. Once you see the proof, the “add the gradients” step stops looking like an implementation trick and starts looking like the only mathematically correct choice.</p><p>In future articles I plan to share what I learn across several areas I’m interested in: the mathematical and systems side of AI, cryptography and Web3, cybersecurity, finance, and eventually quantum computing. This piece is the first on the AI side; the others will come as I go deeper into each topic.</p><p>If you are following the same path of deliberate learning, I hope this piece gives you a solid foothold.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ffca19642931" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>