A Too-Good-to-be-True Prior to Reduce Shortcut Reliance

TL;DR

This paper proposes a two-stage training method using low-capacity networks to detect and reduce reliance on superficial shortcuts, thereby improving out-of-distribution generalization in deep networks.

Contribution

It introduces a novel two-stage approach where low-capacity networks identify shortcuts to guide high-capacity networks towards invariant features, enhancing generalization.

Findings

Two-stage LCN-HCN approach reduces shortcut reliance

Method improves out-of-distribution generalization on modified CIFAR-10

LCNs effectively serve as shortcut detectors

Abstract

Despite their impressive performance in object recognition and other tasks under standard testing conditions, deep networks often fail to generalize to out-of-distribution (o.o.d.) samples. One cause for this shortcoming is that modern architectures tend to rely on "shortcuts" - superficial features that correlate with categories without capturing deeper invariants that hold across contexts. Real-world concepts often possess a complex structure that can vary superficially across contexts, which can make the most intuitive and promising solutions in one context not generalize to others. One potential way to improve o.o.d. generalization is to assume simple solutions are unlikely to be valid across contexts and avoid them, which we refer to as the too-good-to-be-true prior. A low-capacity network (LCN) with a shallow architecture should only be able to learn surface relationships, including shortcuts. We find that LCNs can serve as shortcut detectors. Furthermore, an LCN's predictions can be used in a two-stage approach to encourage a high-capacity network (HCN) to rely on deeper invariant features that should generalize broadly. In particular, items that the LCN can master are downweighted when training the HCN. Using a modified version of the CIFAR-10 dataset in which we introduced shortcuts, we found that the two-stage LCN-HCN approach reduced reliance on shortcuts and facilitated o.o.d. generalization.