Laplacian Regularized Variational Few-Shot Learning for Image Classification

We propose a two-stage meta-learning approach for few-shot image classification. The first (training) stage is realised by exploiting stochastic variational approximation of true posterior distributions, via both query and support samples in a few-shot learning paradigm. During the second (inference) phase, transductive clustering is applied to the query points implementing Laplacian regularisation to encourage smooth labelling, in an effort to enhance classification performance. We conduct empirical evaluations on both Omniglot and miniImagenet datasets, comparing our approach against the state-of-the-art techniques. We also report on the run time performance of our proposed method to demonstrate its efficiency.