Out-of-distribution Detection by Quantifying the Uncertainty with the Stochastic Weight Ensemble

초록

The performance of a deep neural network (DNN) decreases significantly when it encounters out-of- distribution (OOD) samples that deviate from the training data distribution. Current OOD detection methods tend to overlook model uncertainty because measurements of model uncertainty require inferring the posterior distribution of the DNN parameters, which is intractable for a single-point estimate DNN, resulting in suboptimal performance. This paper introduces a novel OOD detection method utilizing a stochastic weight ensemble (SWE) to quantify model uncertainty. The SWE was created using multiple checkpoints traversed by a stochastic gradient descent optimizer. Each checkpoint represents a distinct model state within weight space, leading to diverse predictions for the same input. The proposed method was evaluated using multiple in-distribution and OOD benchmark datasets. The approach considers data and model uncertainties, providing a more robust and reliable detection performance. On a CIFAR-100 pre-trained on ResNet-50, the proposed method reduces the average FPRat95 by 8.05% compared to the maximum softmax probability score. Extensive experimental results showed that the proposed method achieves state-of-the-art OOD detection performance on common benchmarks.

키워드