Augmented Dirichlet Prior Networks for uncertainty estimation and out of distribution detection

Abstract

We propose a method to augment a classifier network into a dirichlet prior network. We also propose a new loss for dirichlet prior networks, in consideration of the resulting classifier as an energy based model and the factorization characteristics of data in view of the likelihood ratio test. Looking at the utility of incorporating image transformations in out of distribution detection, self supervised feature learning we propose a feature learning approach. We establish a setting to perform and evaluate experiments on uncertainty estimation and out of distribution detection with CIFAR-10 as in domain data and SVHN as out domain data.

Experiment

We have divided the analysis into two parts. In one, we use CIFAR10 as our in-distribution data and SVHN as our out-distribution data. In the next part, we use CIFAR10 as our in-distribution data and adverserial examples generated by fast gradient sign method (FGSM).

We use three models for feature extraction.

1. We use the features from the penultimate layer of our pre-trained ResNet.
2. We train a Simclr model on CIFAR10 for 1200 epochs and use it to extract the features.
3. We train an augmented Simclr which is trained on a combination of in-distribution and out-distribution data with respective temperature scaling.

We also tried to use temperature scaling of the predicted outputs using the targets, as an additional parameter to give more weightage to the in-distribution data as compared to out-distribution data during the training of our Augnet.