[Paper Review] Predicting Neural Network Accuracy from Weights

Thomas Unterthiner, Daniel Keysers+3 more|arXiv (Cornell University)|Feb 26, 2020

Adversarial Robustness in Machine Learning33 references38 citations

TL;DR

The paper demonstrates that a CNN’s expected accuracy can be predicted from its trained weights alone, using simple statistics and machine learning models, and shows transfer of this predictive signal across datasets and architectures.

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.

Motivation & Objective

• Motivate and formalize the problem of predicting CNN accuracy from weights alone.
• Create and publish a large dataset of CNNs with varied hyperparameters and datasets to enable the study.
• Demonstrate that weight-based predictors can rank networks by accuracy with high fidelity (R^2).
• Explore transfer of predictive signals across datasets and architectures (domain shift).

Proposed method

• Formal setting mapping trained weights to expected accuracy without accessing input data during prediction.
• Construct the Small CNN Zoo dataset (4 datasets, 30k configurations per dataset) with a fixed small CNN architecture.
• Train several predictive models (GBM, DNN, L-Linear) to map weight-based features to test accuracy, optimizing via cross-validation.
• Explore input feature representations from weights (full flattened weights, per-layer statistics, weight norms).
• Evaluate predictive performance with R^2, MSE, and Kendall’s tau for transfer across datasets/architectures.

Experimental results

Research questions

• RQ1 Can trained weights alone predict a CNN’s expected accuracy on unseen data?
• RQ2 How well do simple weight-based statistics predict accuracy across different datasets and architectures (domain shift)?
• RQ3 Which weight-derived features best predict accuracy, and how do different predictors compare?
• RQ4 Does transfer of predictive power hold when moving from small to large (over-parameterized) architectures?

Key findings

• GBM and DNN predictors outperform a logit-linear baseline across all datasets; GBM with per-layer statistics W_L achieves the strongest performance.
• Using only the entire flattened weight vector W or the last dense layer W^4 already yields strong prediction performance, while per-layer statistics W_L often performs best (e.g., R^2 up to ~0.993 on some datasets).
• Per-layer statistics W_L provide the strongest and most data-efficient features for prediction, outperforming raw weights or layer-wise norms in several cases.
• Predictors trained on one dataset can reasonably rank networks trained on another dataset, with Kendall’s tau values indicating meaningful rank correlation (e.g., up to 0.93 in some transfers).
• Predictions show some invariances (e.g., scaling of weights and convolutional layer permutations) emerge, though the model is most sensitive to permutations in the final dense layer.