Is Feature Diversity Necessary in Neural Network Initialization?

TL;DR

This paper investigates the role of feature diversity at initialization in neural networks, showing that lack of diversity hampers training but can be mitigated by noise, and that identical features can still lead to successful training.

Contribution

It demonstrates that feature diversity is not strictly necessary for training neural networks, challenging prior assumptions, and introduces methods to compensate for low diversity at initialization.

Findings

Lack of feature diversity harms training performance.

Adding small noise can counteract low diversity effects.

Identical features with near-zero weights can still train effectively.

Abstract

Standard practice in training neural networks involves initializing the weights in an independent fashion. The results of recent work suggest that feature "diversity" at initialization plays an important role in training the network. However, other initialization schemes with reduced feature diversity have also been shown to be viable. In this work, we conduct a series of experiments aimed at elucidating the importance of feature diversity at initialization. We show that a complete lack of diversity is harmful to training, but its effects can be counteracted by a relatively small addition of noise - even the noise in standard non-deterministic GPU computations is sufficient. Furthermore, we construct a deep convolutional network with identical features at initialization and almost all of the weights initialized at 0 that can be trained to reach accuracy matching its standard-initialized counterpart.