| 内田 滋穂里 | D, 中間発表 | 数理情報学 | 池田 和司 | 諏訪 正樹 | 久保 孝富 | 山川 俊貴 |
|
title: Feature-Centric Pruning: Robust and Lightweight CNNs via Channel Selection Based on Activation Sparsity
abstract: Deploying convolutional neural networks (CNNs) on resource-constrained edge devices requires model compression, yet structured pruning often degrades robustness to real-world input changes such as noise, blur, weather effects, and style shifts. Representative pruning criteria select channels by parameter statistics such as weight magnitude or gradient-based importance, and do not consider which feature representations remain after pruning. In this talk, I present Feature-Centric Pruning, a structured pruning method that retains channels according to the properties of their feature representations. The network is first trained under a Top-K activation constraint, which encourages channels to specialize in local structures such as contours. Channels are then ranked by activation sparsity (zerorate), computed from forward passes alone, and the most spatially selective channels are retained before fine-tuning. Experiments with ResNet-18 show that the proposed method preserves robustness significantly better than L1-Norm pruning on Imagenette-C and outperforms existing criteria on CIFAR-10-C across multiple random seeds. Further analyses show that the Top-K constraint and the selection criterion act synergistically, and that the effect depends on how strongly channels specialize in a given architecture. Finally, I describe the remaining work, which focuses on clarifying why and under what conditions the method is effective, and the plan toward completing the doctoral dissertation. language of the presentation: Japanese 発表題目: Feature-Centric Pruning：活性スパース性にもとづくチャネル選択によるロバストかつ軽量なCNN 発表概要: エッジ機器でCNNを利用するにはモデルの軽量化が欠かせないが，構造化プルーニングでチャネルを削減すると，ノイズ，ぼけ，天候，スタイル変化といった実環境の入力変化に対するロバスト性が低下しやすい．代表的な選択基準は重みの大きさや勾配にもとづくもので，枝刈り後にどのような特徴表現が残るかを考慮していない．本発表では，特徴表現の性質にもとづいて残すチャネルを選ぶ構造化プルーニング手法Feature-Centric Pruningを報告する．本手法では，まずTop-K活性制約のもとで学習し，輪郭などの局所構造に特化したチャネルの形成を促す．次に，順伝播のみで計算できる活性スパース性（zerorate）によってチャネルを順位づけし，空間選択的に応答するチャネルを残してファインチューニングする．ResNet-18を用いた実験では，Imagenette-CにおいてL1-Normによる枝刈りよりも有意に高いロバスト性を保ち，CIFAR-10-Cでも複数の乱数シードで既存の基準を上回った．また，Top-K学習と選択基準が協調して働くこと，その効果がアーキテクチャにおけるチャネルの機能特化の度合いに依存することを示す．最後に，手法がなぜ・どのような条件で有効かを明らかにする今後の分析と，博士論文の完成に向けた計画について述べる． | ||||||