Probabilistic Trust Intervals for Out of Distribution Detection

2 Feb 2021  ·  Gagandeep Singh, Deepak Mishra ·

Building neural network classifiers with an ability to distinguish between in and out-of distribution inputs is an important step towards faithful deep learning systems. Some of the successful approaches for this, resort to architectural novelties, such as ensembles, with increased complexities in terms of the number of parameters and training procedures. Whereas some other approaches make use of surrogate samples, which are easy to create and work as proxies for actual out-of-distribution (OOD) samples, to train the networks for OOD detection. In this paper, we propose a very simple approach for enhancing the ability of a pretrained network to detect OOD inputs without even altering the original parameter values. We define a probabilistic trust interval for each weight parameter of the network and optimize its size according to the in-distribution (ID) inputs. It allows the network to sample additional weight values along with the original values at the time of inference and use the observed disagreement among the corresponding outputs for OOD detection. In order to capture the disagreement effectively, we also propose a measure and establish its suitability using empirical evidence. Our approach outperforms the existing state-of-the-art methods on various OOD datasets by considerable margins without using any real or surrogate OOD samples. We also analyze the performance of our approach on adversarial and corrupted inputs such as CIFAR-10-C and demonstrate its ability to clearly distinguish such inputs as well. By using fundamental theorem of calculus on neural networks, we explain why our technique doesn't need to observe OOD samples during training to achieve results better than the previous works.

PDF Abstract

Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods


No methods listed for this paper. Add relevant methods here