Slowing Learning by Erasing Simple Features

TL;DR

This paper introduces QLEACE, a novel method for surgically removing quadratic information about concepts from neural network representations, and explores its effects on learning speed across different architectures and datasets.

Contribution

The paper presents QLEACE, a closed-form quadratic erasure technique, and investigates its impact on neural network learning, revealing complex effects depending on architecture and data.

Findings

QLEACE consistently slows learning in feedforward networks.

Quadratic erasure can both hinder and enhance learning depending on context.

Approximate variants of quadratic erasure sometimes act as data augmentation, improving learning speed.

Abstract

Prior work suggests that neural networks tend to learn low-order moments of the data distribution first, before moving on to higher-order correlations. In this work, we derive a novel closed-form concept erasure method, QLEACE, which surgically removes all quadratically available information about a concept from a representation. Through comparisons with linear erasure (LEACE) and two approximate forms of quadratic erasure, we explore whether networks can still learn when low-order statistics are removed from image classification datasets. We find that while LEACE consistently slows learning, quadratic erasure can exhibit both positive and negative effects on learning speed depending on the choice of dataset, model architecture, and erasure method. Use of QLEACE consistently slows learning in feedforward architectures, but more sophisticated architectures learn to use injected higher order Shannon information about class labels. Its approximate variants avoid injecting information, but surprisingly act as data augmentation techniques on some datasets, enhancing learning speed compared to LEACE.