Similarity-Induced Embeddings for Classification

Abstract

We apply similarity learning, a technique commonly used in the ranking setting, to obtain a task-independent embedding. Such an embedding space is more expressive than the embedding space induced by standard classifiers, in that it can express more detailed relationships between and within classes. In doing so, we can significantly reduce the dimensionality of the representation of the data in a way that does not depend directly on the number of classes. We apply this embedding to the task of image classification. This approach also has the advantage that similarity learning is a weaker form of supervision than classification. Using category-level similarity data, we obtain comparable performance to classifiers trained specifically for this task.