munema / survey

サーベイ論文まとめ
0 stars 1 forks source link

A Benchmark for Interpretability Methods in Deep Neural Networks #12

Open munema opened 4 years ago

munema commented 4 years ago

Paper link : https://papers.nips.cc/paper/9167-a-benchmark-for-interpretability-methods-in-deep-neural-networks Author : Hooker, Sara and Erhan, Dumitru and Kindermans, Pieter-Jan and Kim, Been Conference : NeurIPS 2019

概要

深層学習モデルにおいて特徴の重要度を与えて解釈性を与える手法の重要度を評価する尺度 ( ROAR, Remove And Retrain)を提案した. これにより, 多くの解釈可能な機械学習モデルの特徴重要度がランダムに計算された特徴重要度よりも悪いことが示された. VarGradおよびSmoothGrad-Squaredがランダムに計算された特徴重要度よりも高い性能を示した.

ROAR: Remove And Retrain