Deciphering Shortcut Learning from an Evolutionary Game Theory Perspective

TL;DR

This paper uses evolutionary game theory to analyze how shortcut learning bias forms in neural networks, revealing the roles of gradient descent methods and noise in this process.

Contribution

It provides a formal definition of core and shortcut features and models their competition using evolutionary game theory, offering new insights into the origins of shortcut bias.

Findings

GD and SGD lead to different stable states favoring shortcut or core features

Data noise and optimization noise influence the formation of shortcut bias

Theoretical framework characterizes shortcut bias dynamics and mitigation strategies

Abstract

Shortcut learning causes deep learning models to rely on non-essential features within the data. However, its formation in deep neural network training still lacks theoretical understanding. In this paper, we provide a formal definition of core and shortcut features and employ evolutionary game theory to analyze the origins of shortcut bias by modeling data samples as players and their corresponding neural tangent features as strategies, assuming the existence of core and shortcut subnetworks. We find that gradient descent (GD) and stochastic gradient descent (SGD) lead to two distinct stochastically stable states, each corresponding to a different strategy. The former primarily optimizes the shortcut subnetwork, while the latter primarily optimizes the core subnetwork. We investigate the influence of these strategies on shortcut bias through a continuous stochastic differential equation, and reveal the impact of data noise and optimization noise on the formation of shortcut bias. In brief, our work employs evolutionary game theory to characterize the dynamics of shortcut bias formation and provides a theoretical view on its mitigation.