A Self-Supervised Learning Method to Mahalanobis Distance for Improving OOD Detection

In out-of-distribution detection, methods based on the Mahalanobis distance have excellent performance for detecting out-of-distribution data. The Mahalanobis distance measures the distance between samples through the mean and covariance matrix of the data, however, when the data do not conform to the Gaussian distribution, the mean and covariance matrices cannot effectively characterise the distribution of data points within the category. In reality, the category distribution of many datasets does not fully conform to the Gaussian distribution, especially when the data have multimodal distribution, skewness or outliers, and this assumption leads to errors in the computation of the Mahalanobis distance. To address this problem, this paper proposes a method for self-supervised fitting of Gaussian distributions. The method uses a strategy that combines an aggregation algorithm and a Triplet loss to optimise the data through an aggregation algorithm to make its distribution closer to a Gaussian distribution, while using a Triplet loss to enhance intra-class aggregation and inter-class separation.