To avoid parameter-space complexity measures such as KL divergences, weight norms and compression, the authors certify the deployed network through its last hidden layer representation.
Generalization certified via last-layer representation quality, bypassing parameter-space complexity
Using a k-nearest-neighbor decoder on last-hidden-layer features, the authors derive tight, non-vacuous certificates for deterministic networks on CIFAR-10, CIFAR-100 and 1,000-class ImageNet.
Top University
Niclas Göring · Shuofeng Zhang · Branton DeMoss · Ard A. Louis
University of Oxford
Research Digest··2 min read
Göring et al.
§