<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Machine-Learning on lgcode</title>
    <link>https://blog.lgcode.me/tags/machine-learning/</link>
    <description>Recent content in Machine-Learning on lgcode</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Wed, 26 Aug 2026 09:00:00 -0300</lastBuildDate>
    <atom:link href="https://blog.lgcode.me/tags/machine-learning/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>A Classifier That Can Say &#34;Not Mine&#34;</title>
      <link>https://blog.lgcode.me/posts/a-classifier-that-can-say-not-mine/</link>
      <pubDate>Wed, 26 Aug 2026 09:00:00 -0300</pubDate>
      <guid>https://blog.lgcode.me/posts/a-classifier-that-can-say-not-mine/</guid>
      <description>Almost all classification writing assumes a closed set. In production you need three answers, not two — and &amp;#34;wrong family entirely&amp;#34; is a different failure from &amp;#34;unsure which one&amp;#34;.</description>
    </item>
    <item>
      <title>Don&#39;t Resize Your Textures Away</title>
      <link>https://blog.lgcode.me/posts/dont-resize-your-textures-away/</link>
      <pubDate>Wed, 12 Aug 2026 09:00:00 -0300</pubDate>
      <guid>https://blog.lgcode.me/posts/dont-resize-your-textures-away/</guid>
      <description>One class was hopeless and the rest were fine. The model was never the problem — a Resize() at the top of the transform destroyed the only signal that separated them. Right diagnosis, over-applied fix.</description>
    </item>
    <item>
      <title>The Validation Split That Flattered Every Number</title>
      <link>https://blog.lgcode.me/posts/the-validation-split-that-flattered-every-number/</link>
      <pubDate>Sun, 09 Aug 2026 09:00:00 -0300</pubDate>
      <guid>https://blog.lgcode.me/posts/the-validation-split-that-flattered-every-number/</guid>
      <description>94% of my validation set was in training under a different filename. Fixing it dropped the numbers, the field benchmark said the leaked model was still better — and the field benchmark was broken too.</description>
    </item>
  </channel>
</rss>
