Learning Case Features with Proxy-Guided Deep Neural Networks

Abstract: The cost and difficulty of acquiring case features motivates interest in machine learning for feature acquisition. For computer vision domains, manual feature extraction has proven infeasible, but previous studies have shown the effectiveness of extracting features from deep neural models for case-based classification. Such approaches have generally been based on training the network for stand-alone classification accuracy, under the assumption that effective classification reflects high quality network features. However, it is not clear that the features best suited to network processing will be best for CBR. In response, this paper proposes refining previous network feature extraction approaches by adapting network training to reflect the goal of using network features for CBR. Specifically, it proposes augmenting conventional cross-entropy loss with a proxy term that reflects how the CBR system will use extracted features for similarity assessment. To this end, we investigate using Pairwise Distance, Cosine Similarity, and Sinkhorn Divergence as proxy functions within a triplet loss training framework. Evaluations on the benchmark image classification datasets MNIST, Animals with Attributes 2, and CIFAR-10 support the effectiveness of this method, with an integrated case-based classification system using the extracted features outperforming the feature extraction network applied end-to-end as well as integrated models developed in our previous research.