Predicting Neural Network Accuracy from Weights

TL;DR

This paper demonstrates that neural network accuracy can be accurately predicted solely from weights, enabling performance ranking without data evaluation, which advances understanding of network training and generalization.

Contribution

It introduces a formal setting for predicting neural network accuracy from weights and shows high prediction accuracy across different datasets and architectures.

Findings

Weight-based predictors achieve R2 > 0.98 in ranking networks

Predictors generalize across datasets and architectures

A large dataset of 120k trained CNNs is released for further research

Abstract

We show experimentally that the accuracy of a trained neural network can be predicted surprisingly well by looking only at its weights, without evaluating it on input data. We motivate this task and introduce a formal setting for it. Even when using simple statistics of the weights, the predictors are able to rank neural networks by their performance with very high accuracy (R2 score more than 0.98). Furthermore, the predictors are able to rank networks trained on different, unobserved datasets and with different architectures. We release a collection of 120k convolutional neural networks trained on four different datasets to encourage further research in this area, with the goal of understanding network training and performance better.