Initial Guessing Bias: How Untrained Networks Favor Some Classes

TL;DR

This paper reveals that untrained deep neural networks can inherently favor certain classes due to their structure, leading to initial biases that impact model fairness and performance.

Contribution

The paper provides a theoretical analysis of Initial Guessing Bias (IGB), showing how network architecture influences class prediction biases before training.

Findings

Untrained networks can assign all predictions to a single class due to IGB.

Model choices like activation functions and depth affect the severity of IGB.

Theoretical insights inform better architecture and initialization strategies.

Abstract

Understanding and controlling biasing effects in neural networks is crucial for ensuring accurate and fair model performance. In the context of classification problems, we provide a theoretical analysis demonstrating that the structure of a deep neural network (DNN) can condition the model to assign all predictions to the same class, even before the beginning of training, and in the absence of explicit biases. We prove that, besides dataset properties, the presence of this phenomenon, which we call \textit{Initial Guessing Bias} (IGB), is influenced by model choices including dataset preprocessing methods, and architectural decisions, such as activation functions, max-pooling layers, and network depth. Our analysis of IGB provides information for architecture selection and model initialization. We also highlight theoretical consequences, such as the breakdown of node-permutation symmetry, the violation of self-averaging and the non-trivial effects that depth has on the phenomenon.