Increasing biases can be more efficient than increasing weights
Abstract
We introduce a novel computational unit for neural networks that features multiple biases, challenging the traditional perceptron structure. This unit emphasizes the importance of preserving uncorrupted information as it is passed from one unit to the next, applying activation functions later in the process with specialized biases for each unit. Through both empirical and theoretical analyses, we show that by focusing on increasing biases rather than weights, there is potential for significant enhancement in a neural network model's performance. This approach offers an alternative perspective on optimizing information flow within neural networks. See source code at https://github.com/CuriosAI/dac-dev.
Cite
@article{arxiv.2301.00924,
title = {Increasing biases can be more efficient than increasing weights},
author = {Carlo Metta and Marco Fantozzi and Andrea Papini and Gianluca Amato and Matteo Bergamaschi and Silvia Giulia Galfrè and Alessandro Marchetti and Michelangelo Vegliò and Maurizio Parton and Francesco Morandin},
journal= {arXiv preprint arXiv:2301.00924},
year = {2024}
}
Comments
Major rewriting. Supersedes v1 and v2. Focusing on the fact that not all parameters are born equal: biases can be more important than weights. Accordingly, new title and new abstract, and many more experiments on fully connected architectures. This is the extended version of the paper published at WACV 2024