HAct: Out-of-Distribution Detection with Neural Net Activation Histograms

TL;DR

This paper introduces HAct, a novel activation histogram-based method for out-of-distribution detection in neural networks, demonstrating superior accuracy and efficiency over existing techniques on image classification benchmarks.

Contribution

The paper presents a new descriptor, HAct, that effectively captures neural network layer activations for improved OOD detection, with a simple and scalable implementation.

Findings

HAct achieves 95% TPR with only 0.03% false positives on Resnet-50.

HAct outperforms state-of-the-art OOD detection methods by 20.67% in false positive rate.

The method is computationally efficient and suitable for real-time deployment.

Abstract

We propose a simple, efficient, and accurate method for detecting out-of-distribution (OOD) data for trained neural networks. We propose a novel descriptor, HAct - activation histograms, for OOD detection, that is, probability distributions (approximated by histograms) of output values of neural network layers under the influence of incoming data. We formulate an OOD detector based on HAct descriptors. We demonstrate that HAct is significantly more accurate than state-of-the-art in OOD detection on multiple image classification benchmarks. For instance, our approach achieves a true positive rate (TPR) of 95% with only 0.03% false-positives using Resnet-50 on standard OOD benchmarks, outperforming previous state-of-the-art by 20.67% in the false positive rate (at the same TPR of 95%). The computational efficiency and the ease of implementation makes HAct suitable for online implementation in monitoring deployed neural networks in practice at scale.