An Investigation of the Weight Space to Monitor the Training Progress of Neural Networks

TL;DR

This paper explores the structure of neural network weight space to develop methods for monitoring training progress and detecting domain shifts, potentially reducing the need for expensive testing.

Contribution

It reveals that neural networks follow unique, smooth trajectories in weight space during training, which can be exploited to track progress and identify model versions.

Findings

Models follow smooth, unique trajectories in weight space.

Trajectory properties can indicate training progress and domain shifts.

Checkpoints can be ordered along trajectories for versioning.

Abstract

Safe use of Deep Neural Networks (DNNs) requires careful testing. However, deployed models are often trained further to improve in performance. As rigorous testing and evaluation is expensive, triggers are in need to determine the degree of change of a model. In this paper we investigate the weight space of DNN models for structure that can be exploited to that end. Our results show that DNN models evolve on unique, smooth trajectories in weight space which can be used to track DNN training progress. We hypothesize that curvature and smoothness of the trajectories as well as step length along it may contain information on the state of training as well as potential domain shifts. We show that the model trajectories can be separated and the order of checkpoints on the trajectories recovered, which may serve as a first step towards DNN model versioning.