Levene & Harris new robustness measure bounds neural noise errors

In a new preprint on arXiv, researchers Mark Levene and Martyn Harris tackle the critical issue of neural network robustness to random input noise. They propose a measure that estimates an upper bound on the network's mean squared error (MSE) with high probability, given a specific level of input perturbation. The key innovation is that their method treats the neural network as a black box, requiring no knowledge of internal architecture or gradients. This makes it both simple and computationally efficient to apply across different models. The authors validated their approach on several real-world datasets, demonstrating its efficacy in predicting how much a network's output error will degrade under noise.

The paper also introduces the concept of 'robustness curves,' which allow researchers to visualize and compare noise tolerance both within a single dataset (e.g., different classes) and across different datasets. This provides deeper insight into where models are most vulnerable. While the paper focuses on mean squared error as the accuracy metric, the framework could extend to other loss functions. For practitioners, this work offers a practical, low-cost way to certify a model's stability before deployment in noisy environments, such as autonomous vehicles or medical imaging, where input perturbations are common.