Initial Guessing Bias: How Untrained Networks Favor Some Classes
This addresses fairness and accuracy issues in classification for ML practitioners, but it is incremental as it builds on existing bias analysis.
The paper tackles the problem of neural networks being biased to predict the same class before training due to architectural choices, proving that factors like activation functions and depth influence this initial guessing bias.
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.