Predicting Neural Network Accuracy from Weights

About

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 · 2020

Related benchmarks

Task	Dataset	Result
Performance Prediction	Small CNN Zoo ReLU subset (test)	Kendall’s Tau0.915	35
Weight-Space Classification	Model Jungle	Accuracy55.8	20
Predicting Transformer Generalization	MNIST-Transformers No threshold	Kendall's Tau0.881	8
Predicting Transformer Generalization	MNIST-Transformers 20% threshold	Kendall's tau0.872	8
Predicting Transformer Generalization	MNIST-Transformers 60% threshold	Kendall's tau0.86	8
Predicting Transformer Generalization	MNIST-Transformers 40% threshold	Kendall's tau0.868	8
Predicting Transformer Generalization	MNIST-Transformers (80% threshold)	Kendall's Tau0.856	8
Predicting Transformer Generalization	AGNews Transformers No threshold	Kendall's Tau0.841	8
Predicting Transformer Generalization	AGNews Transformers 20% threshold	Kendall's tau0.839	8
Predicting Transformer Generalization	AGNews Transformers (40% threshold)	Kendall's Tau0.812	8