Exploring Complementary Strengths of Invariant and Equivariant Representations for Few-Shot Learning

In many real-world problems, collecting a large number of labeled samples is\ninfeasible. Few-shot learning (FSL) is the dominant approach to address this\nissue, where the objective is to quickly adapt to novel categories in presence\nof a limited number of samples. FSL tasks have been predominantly solved by\nleveraging the ideas from gradient-based meta-learning and metric learning\napproaches. However, recent works have demonstrated the significance of\npowerful feature representations with a simple embedding network that can\noutperform existing sophisticated FSL algorithms. In this work, we build on\nthis insight and propose a novel training mechanism that simultaneously\nenforces equivariance and invariance to a general set of geometric\ntransformations. Equivariance or invariance has been employed standalone in the\nprevious works; however, to the best of our knowledge, they have not been used\njointly. Simultaneous optimization for both of these contrasting objectives\nallows the model to jointly learn features that are not only independent of the\ninput transformation but also the features that encode the structure of\ngeometric transformations. These complementary sets of features help generalize\nwell to novel classes with only a few data samples. We achieve additional\nimprovements by incorporating a novel self-supervised distillation objective.\nOur extensive experimentation shows that even without knowledge distillation\nour proposed method can outperform current state-of-the-art FSL methods on five\npopular benchmark datasets.\n