<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Sparsity-Pruning on Jaehun's Blog</title><link>https://jaehun.me/categories/sparsity-pruning/</link><description>Recent content in Sparsity-Pruning on Jaehun's Blog</description><generator>Hugo</generator><language>ko-kr</language><lastBuildDate>Sun, 13 Sep 2026 09:29:41 +0900</lastBuildDate><atom:link href="https://jaehun.me/categories/sparsity-pruning/index.xml" rel="self" type="application/rss+xml"/><item><title>SQS: Bayesian DNN Compression through Sparse Quantized Sub-distributions</title><link>https://jaehun.me/posts/sqs-bayesian-dnn-compression-through-sparse-quantized-sub-distributions/</link><pubDate>Sun, 13 Sep 2026 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/sqs-bayesian-dnn-compression-through-sparse-quantized-sub-distributions/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2510.08999"target="_blank"&#10; class="inline-flex items-center gap-1"&#10; &gt;논문 링크&lt;svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/&gt;&lt;/svg&gt;&#10; &lt;/a&gt;&lt;/p&gt;&#10;&lt;h2 id="sqs-가지치기와-양자화를-하나의-베이지안-학습으로--스파이크-앤-슬랩과-gmm의-만남"&gt;SQS: 가지치기와 양자화를 하나의 베이지안 학습으로 — 스파이크-앤-슬랩과 GMM의 만남&lt;a href="#sqs-%ea%b0%80%ec%a7%80%ec%b9%98%ea%b8%b0%ec%99%80-%ec%96%91%ec%9e%90%ed%99%94%eb%a5%bc-%ed%95%98%eb%82%98%ec%9d%98-%eb%b2%a0%ec%9d%b4%ec%a7%80%ec%95%88-%ed%95%99%ec%8a%b5%ec%9c%bc%eb%a1%9c--%ec%8a%a4%ed%8c%8c%ec%9d%b4%ed%81%ac-%ec%95%a4-%ec%8a%ac%eb%9e%a9%ea%b3%bc-gmm%ec%9d%98-%eb%a7%8c%eb%82%a8" class="heading-anchor" aria-label="이 섹션에 대한 링크"&gt;&lt;svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"&gt;&lt;path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/&gt;&lt;path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/&gt;&lt;/g&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt; — 프루닝(pruning)과 저비트 양자화(quantization)를 따로 하면 남는 잉여가 생긴다. SQS는 &lt;strong&gt;스파이크-앤-슬랩 사전분포&lt;/strong&gt;와 **가우시안 혼합 모델(GMM)**을 하나의 변분(variational) 학습으로 묶어, 동일한 비트폭에서 더 높은 압축률을, 동일한 압축률에서 더 작은 정확도 손실을 달성한다.&lt;/p&gt;</description></item><item><title>X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation</title><link>https://jaehun.me/posts/x-aut-progressive-audio-encoder-compression-for-speech-llms-with-cross-scale-distillation/</link><pubDate>Sat, 12 Sep 2026 00:00:00 +0900</pubDate><guid>https://jaehun.me/posts/x-aut-progressive-audio-encoder-compression-for-speech-llms-with-cross-scale-distillation/</guid><description>&lt;p&gt;&lt;a&#10; href="https://arxiv.org/abs/2609.11412"target="_blank"&#10; class="inline-flex items-center gap-1"&#10; &gt;논문 링크&lt;svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/&gt;&lt;/svg&gt;&#10; &lt;/a&gt;&lt;/p&gt;&#10;&lt;h2 id="x-aut-행동-프로브와-크로스스케일-증류로-음성-llm의-오디오-인코더를-깎아내는-방법"&gt;X-AuT: 행동 프로브와 크로스스케일 증류로 음성 LLM의 오디오 인코더를 깎아내는 방법&lt;a href="#x-aut-%ed%96%89%eb%8f%99-%ed%94%84%eb%a1%9c%eb%b8%8c%ec%99%80-%ed%81%ac%eb%a1%9c%ec%8a%a4%ec%8a%a4%ec%bc%80%ec%9d%bc-%ec%a6%9d%eb%a5%98%eb%a1%9c-%ec%9d%8c%ec%84%b1-llm%ec%9d%98-%ec%98%a4%eb%94%94%ec%98%a4-%ec%9d%b8%ec%bd%94%eb%8d%94%eb%a5%bc-%ea%b9%8e%ec%95%84%eb%82%b4%eb%8a%94-%eb%b0%a9%eb%b2%95" class="heading-anchor" aria-label="이 섹션에 대한 링크"&gt;&lt;svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"&gt;&lt;path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/&gt;&lt;path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/&gt;&lt;/g&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h2&gt;&lt;h2 id="한-줄-요약-tldr"&gt;한 줄 요약 (TL;DR)&lt;a href="#%ed%95%9c-%ec%a4%84-%ec%9a%94%ec%95%bd-tldr" class="heading-anchor" aria-label="이 섹션에 대한 링크"&gt;&lt;svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"&gt;&lt;path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/&gt;&lt;path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/&gt;&lt;/g&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h2&gt;&lt;p&gt;음성 LLM의 오디오 인코더(오디오 Transformer)를 &lt;strong&gt;18→16→14 레이어로 점진적으로 프루닝&lt;/strong&gt; 하면서, 짧은 &lt;strong&gt;행동 프로브(behavioral probe)&lt;/strong&gt; 로 회복 가능한 레이어 조합을 고르고 &lt;strong&gt;크로스스케일 증류(1.7B teacher → 0.6B student)&lt;/strong&gt; 와 3단계 회복으로 손실을 복구하는 프레임워크 &lt;strong&gt;X-AuT&lt;/strong&gt; 를 제안한다. 14레이어 모델은 오디오 타워 파라미터를 &lt;strong&gt;20.7%&lt;/strong&gt; 줄이면서 매크로 오류율을 &lt;strong&gt;5.75%&lt;/strong&gt; (baseline 5.61%)로 거의 유지했고, 16레이어 모델은 오히려 &lt;strong&gt;5.27%&lt;/strong&gt; 로 기준선을 능가했다 (근거: Abstract, §1).&lt;/p&gt;</description></item></channel></rss>