Reviews: Out-of-Distribution Detection using Multiple Semantic Label Representations

The problem of the study is out-of-distribution detection; we would like to detect samples that do not belong to the train distribution. To solve this problem, the authors propose reframing the prediction problem as a regression in which the sparse one-hot encoding of the labels in the prediction space is replaced by word embeddings. The presented method also has a flavour of ensemble methods in the sense that a base feature representation (Resnet for image prediction, Lenet for speech recognition) is used and then the feature branches off to multiple fully connected networks each of which generates an embedding vector as the prediction. Each fully connected network is responsible for generating vectors in a specific embedding space. A total of five embedding spaces are used (Skip-gram, GloVe, and FastText trained on independent text datasets).