[논문 리뷰] ReAct: Out-of-distribution Detection With Rectified Activations

Yiyou Sun, Chuan Guo외 1명|arXiv (Cornell University)|2021. 11. 24.

Adversarial Robustness in Machine Learning인용 수 46

한 줄 요약

tldr: ReAct은 신경망의 활성화를 사후 수정하는 방식으로 OOD 탐지를 개선하기 위해 높은 활성화를 완화하고, retraining 없이 ImageNet 및 CIFAR 벤치마크에서 최첨단 성능을 달성합니다.

ABSTRACT

Out-of-distribution (OOD) detection has received much attention lately due to its practical importance in enhancing the safe deployment of neural networks. One of the primary challenges is that models often produce highly confident predictions on OOD data, which undermines the driving principle in OOD detection that the model should only be confident about in-distribution samples. In this work, we propose ReAct--a simple and effective technique for reducing model overconfidence on OOD data. Our method is motivated by novel analysis on internal activations of neural networks, which displays highly distinctive signature patterns for OOD distributions. Our method can generalize effectively to different network architectures and different OOD detection scores. We empirically demonstrate that ReAct achieves competitive detection performance on a comprehensive suite of benchmark datasets, and give theoretical explication for our method's efficacy. On the ImageNet benchmark, ReAct reduces the false positive rate (FPR95) by 25.05% compared to the previous best method.

연구 동기 및 목표

• Motivate the problem of neural network overconfidence on out-of-distribution (OOD) data and its impact on OOD detection.
• Introduce a simple post hoc method (ReAct) that rectifies penultimate-layer activations to mitigate overconfidence.
• Demonstrate empirical improvements across architectures and OOD scoring functions on large-scale and CIFAR benchmarks.
• Provide theoretical analysis explaining why activation truncation helps distinguish ID from OOD data.

제안 방법

• Apply a ReAct operation: bar{h}(x) = min(h(x), c) to the penultimate layer activations, where c is a percentile-based threshold chosen on in-distribution data.
• Leverage downstream OOD scoring functions (e.g., energy score, MSP, ODIN) with the rectified activations f^{ReAct}(x) = W^T bar{h}(x) + b.
• Choose c using a p-th percentile of in-distribution activations (e.g., p=90).
• Demonstrate compatibility with multiple architectures (ResNet-50, MobileNet-v2) and normalization schemes (BatchNorm, WeightNorm, GroupNorm).
• Provide theoretical analysis showing ReAct reduces mean OOD activations more than ID activations, and that this improves the separation between ID and OOD in output scores.

실험 결과

연구 질문

• RQ1 How does rectifying activations through a simple truncation affect OOD detection performance across architectures?
• RQ2 Does ReAct generalize across different OOD scoring functions and normalization schemes?
• RQ3 Why do OOD inputs trigger anomalous activations, and can activation truncation mitigate this effect?
• RQ4 What is the theoretical explanation for the improved separation between ID and OOD after applying ReAct?

주요 결과

• ReAct achieves state-of-the-art post hoc OOD detection performance on large-scale ImageNet and CIFAR benchmarks, reducing false positives significantly. For example, on ImageNet with ResNet, ReAct yields an FPR95 of 20.38 on iNaturalist and up to 47.30 on Textures with AUROC of 92.95, outperforming MSP, ODIN, and energy-based baselines.
• On CIFAR-10/100, ReAct consistently improves OOD detection when combined with various scoring functions, including energy and ODIN; energy+ReAct shows strong gains across datasets.
• A large reported improvement on ImageNet benchmark is a 25.05 percentage-point reduction in FPR95 compared to the previous best method.
• Theoretical analysis shows ReAct reduces OOD activations more than ID activations, especially when OOD activations exhibit positive skewness and higher chaotic-ness, which is typical for many OOD distributions; this leads to larger output gaps for OOD vs ID scores (e.g., energy score).
• Experiments indicate ReAct remains effective across architectures (ResNet-50, MobileNet-v2) and normalization schemes (BatchNorm, WeightNorm, GroupNorm), and is particularly beneficial when applied to the penultimate layer.