Core Concepts
Spurious features impact core feature learning dynamics in neural networks.
Abstract
Existing research focuses on spurious features' impact on neural network optimization.
Proposed theoretical framework and synthetic dataset for studying feature learning dynamics.
Stronger spurious correlations slow down core feature learning.
Spurious features are retained even after core features are learned.
Last Layer Retraining reduces reliance on spurious subnetwork.
Popular debiasing algorithms may fail in complex settings.
Dataset provides insights into learning dynamics under spurious correlations.
Stats
강한 가짜 상관관계 또는 간단한 가짜 기능은 핵심 기능의 학습 속도를 늦춘다.
가짜 기능과 핵심 기능의 학습 단계는 항상 분리되어 있지 않다.
가짜 기능은 유지된다.
마지막 레이어 재학습(LLR)은 가짜 서브네트워크에 대한 의존성을 줄인다.
인기있는 편향 제거 알고리즘은 복잡한 설정에서 실패할 수 있다.
Quotes
"Stronger spurious correlations or simpler spurious features slow down the rate of learning for the core features."
"Spurious features are not forgotten even after core features are fully learned."
"Last Layer Retraining decreases reliance on spurious subnetwork."