多層個体群を用いた遺伝的プログラミングによるCNNアーキテクチャの進化的獲得

カテゴリ: 部門大会

論文No: TC2-2-1

グループ名: 【C】2024年電気学会電子・情報・システム部門大会

発行日: 2024/08/28

タイトル(英語): Evolutionary Acquisition of CNN Architectures Using Genetic Programming with Multi-Layered Population Structure

著者名: 藪内 一貴（大阪公立大学）,森 直樹（大阪公立大学）

著者名(英語): Kazuki Yabuuchi (Osaka Metropolitan University),Naoki Mori (Osaka Metropolitan University)

キーワード: CNN|AutoML|Neural Architecture Search|遺伝的プログラミング|進化的アルゴリズム|CNN|AutoML|Neural Architecture Search|Genetic Programming|Evolutionary Algorithm

要約(日本語): Recently, Convolutional Neural Networks (CNNs) have demonstrated exceptional performance across image recognition tasks. However, the diversification of problems and the increasing complexity of CNN architectures have complicated the search for optimal architectures. Consequently, manually searching for the optimal structure has become increasingly difficult. Genetic Programming (GP) has gained much attention as an automatic optimization method for CNN architectures. One extension of GP is Genetic Programming with Multi-Layered Population Structure (MLPS-GP). We propose a method inspired by MLPS-GP that uses a hierarchical population structure and evolutionarily acquires the optimal architecture through local search and crossover. We verify its utility through experiments.