Predicting Neural Network Accuracy from Weights
This work addresses the challenge of understanding and ranking neural network performance efficiently, with potential applications in model selection and training analysis, though it is incremental in nature.
The paper tackles the problem of predicting neural network accuracy directly from weight statistics without data evaluation, achieving high prediction accuracy (R2 > 0.98) and generalization across datasets and architectures.
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.