In plain words: A tool guesses how accurate a trained neural network will be just by reading its internal weights, never running it on any data. Using simple weight statistics, it ranks networks by performance with a score above 0.98, even across different datasets and architectures.

Abstract · Predicting Neural Network Accuracy from Weights

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.

Thomas Unterthiner, Daniel Keysers, Sylvain Gelly, Olivier Bousquet, Ilya Tolstikhin
arXiv:2002.11448 · stat.ML, cs.LG · submitted Feb 26, 2020 · updated Apr 9, 2021
abstract · pdf · html · Updated the Small CNN Zoo dataset: reduced the maximal learning rate and got rid of multiple bad runs. Replaced all the experiments with the new numbers. Added MLP. Fixed typo in the abstract (R2 score instead of Kendall's tau). Added several earlier related works to the literature overview