<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0">
    <channel>
      <title>Bowen Lee</title>
      <link>https://bowen0701.github.io</link>
      <description>Last 10 notes on Bowen Lee</description>
      <generator>Quartz -- quartz.jzhao.xyz</generator>
      <item>
    <title>RE Log</title>
    <link>https://bowen0701.github.io/re-log/</link>
    <guid>https://bowen0701.github.io/re-log/</guid>
    <description><![CDATA[ Bowen’s RE Log Notes at the intersection of Research and Engineering for Deep Learning, Generative Modeling, Reinforcement Learning and AI4Science. ]]></description>
    <pubDate>Sat, 08 Aug 2026 00:06:00 GMT</pubDate>
  </item><item>
    <title>Bowen Lee</title>
    <link>https://bowen0701.github.io/</link>
    <guid>https://bowen0701.github.io/</guid>
    <description><![CDATA[  Bowen Lee I like AI research and engineering. ]]></description>
    <pubDate>Sat, 08 Aug 2026 00:06:00 GMT</pubDate>
  </item><item>
    <title>Conv2d - PyTorch 2D Convolutional Layer</title>
    <link>https://bowen0701.github.io/re-log/Conv2d---PyTorch-2D-Convolutional-Layer</link>
    <guid>https://bowen0701.github.io/re-log/Conv2d---PyTorch-2D-Convolutional-Layer</guid>
    <description><![CDATA[ Concept: Conv2d - PyTorch 2D Convolutional Layer torch.nn.Conv2d applies a 2D convolution over an input signal (e.g., an image). ]]></description>
    <pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>VAE with ELBO</title>
    <link>https://bowen0701.github.io/re-log/VAE-with-ELBO</link>
    <guid>https://bowen0701.github.io/re-log/VAE-with-ELBO</guid>
    <description><![CDATA[ Concept: VAE with ELBO We want to maximize log-evidence \log p(x) but integrating over all z is intractable. ]]></description>
    <pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>LLM Post-Training Iterative SFT, RLHF Branching</title>
    <link>https://bowen0701.github.io/re-log/LLM-Post-Training-Iterative-SFT,-RLHF-Branching</link>
    <guid>https://bowen0701.github.io/re-log/LLM-Post-Training-Iterative-SFT,-RLHF-Branching</guid>
    <description><![CDATA[ Concept: LLM Post-Training Iterative SFT, RLHF Branching Core Intuition In multi-round LLM post-training, the standard pattern is: Round i: \text{SFT}_i \rightarrow \text{RL}_i Round i+1: start from \text{SFT}_i to train \text{SFT}_{i+1}, then \text{SFT}_{i+1} \rightarrow \text{RL}_{i+1} Why not sta... ]]></description>
    <pubDate>Wed, 24 Jun 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>Why not model.forward() but model()</title>
    <link>https://bowen0701.github.io/re-log/Why-not-model.forward()-but-model()</link>
    <guid>https://bowen0701.github.io/re-log/Why-not-model.forward()-but-model()</guid>
    <description><![CDATA[ Concept: Why not model.forward() but model() Core Intuition Calling model(x) invokes nn.Module.__call__, a wrapper around forward() that runs hooks, manages state, sets up autograd, and handles distributed sync. ]]></description>
    <pubDate>Wed, 20 May 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>Monte Carlo Better than Quadrature</title>
    <link>https://bowen0701.github.io/re-log/Monte-Carlo-Better-than-Quadrature</link>
    <guid>https://bowen0701.github.io/re-log/Monte-Carlo-Better-than-Quadrature</guid>
    <description><![CDATA[ Concept: Monte Carlo Better than Quadrature Core Intuition Q: Why Monte Carlo method is often the default method for high-dimensional integral approximation, instead of Quadrature methods? I got this question while reading Fearnhead et al. ]]></description>
    <pubDate>Sun, 17 May 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>Manual Layer</title>
    <link>https://bowen0701.github.io/re-log/Manual-Layers</link>
    <guid>https://bowen0701.github.io/re-log/Manual-Layers</guid>
    <description><![CDATA[ Concept: Manual Layer Core Intuition A Linear Layer (or Fully Connected/Dense layer) performs an affine transformation on input data. ]]></description>
    <pubDate>Fri, 15 May 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>Xavier Initialization vs He Initialization</title>
    <link>https://bowen0701.github.io/re-log/Xavier-Initialization-vs-He-Initialization</link>
    <guid>https://bowen0701.github.io/re-log/Xavier-Initialization-vs-He-Initialization</guid>
    <description><![CDATA[ Concept: Xavier Initialization vs He Initialization [Updated on 2026-08-07: Extend to Xavier Initialization vs He Initialization.] Core Intuition Both methods solve the same problem: keep the variance of activations stable across layers so gradients neither vanish nor explode. ]]></description>
    <pubDate>Wed, 13 May 2026 00:00:00 GMT</pubDate>
  </item><item>
    <title>Correspondence Analysis</title>
    <link>https://bowen0701.github.io/re-log/Correspondence-Analysis</link>
    <guid>https://bowen0701.github.io/re-log/Correspondence-Analysis</guid>
    <description><![CDATA[ Concept: Correspondence Analysis Core Intuition Correspondence Analysis (CA) is a dimensionality reduction and data visualization technique for contingency tables of counts. ]]></description>
    <pubDate>Thu, 06 Oct 2016 00:00:00 GMT</pubDate>
  </item>
    </channel>
  </rss>