Noise Injection as a Probe of Deep Learning Dynamics

TL;DR

This paper introduces Noise Injection Nodes (NINs) as a novel method to analyze deep neural network training dynamics by injecting uncorrelated noise and observing phase transitions during learning.

Contribution

It presents a new probing technique using NINs that does not alter the optimization process, revealing distinct training phases influenced by noise scale.

Findings

System exhibits different phases during training based on noise scale

In some cases, noise node evolution mirrors unperturbed loss dynamics

NINs can potentially be used to understand full system behavior

Abstract

We propose a new method to probe the learning mechanism of Deep Neural Networks (DNN) by perturbing the system using Noise Injection Nodes (NINs). These nodes inject uncorrelated noise via additional optimizable weights to existing feed-forward network architectures, without changing the optimization algorithm. We find that the system displays distinct phases during training, dictated by the scale of injected noise. We first derive expressions for the dynamics of the network and utilize a simple linear model as a test case. We find that in some cases, the evolution of the noise nodes is similar to that of the unperturbed loss, thus indicating the possibility of using NINs to learn more about the full system in the future.