ManiNet: Manifold Network for Few-Shot Learning

Rui-Qi Wang, Hengcan Shi, Yi Chen, Yaonan Wang

Abstract: Few-shot Learning (FSL) aims to learn a model that can be seamlessly adapted to unknown classes with only a few labeled data. A concise but successful way is to learn a robust feature encoder to describe novel classes relying on given supervised data. Under the guidance of such insight, most methods define classes as standard Gaussian distributions with different means in feature spaces, where classification can be performed based on the distances between embedding and class centroids. In spite of considerable achievements, these methods always miss the structural information within classes, resulting in degraded performance. To tackle this problem, we develop a novel yet concise approach named Manifold Network (ManiNet) to perform few-shot classification based on manifolds. Technically, in the ManiNet, each class is represented as a tree rather than an isolated centroids to reserve the structural information. And a simple correction term is introduced to elevate the usage of data by representing each manifold with a graph. Benefiting from such modeling, the probability of unknown data belonging to a class is derived based on the relative energy change before and after adding this data into the class manifolds. Experimental results on popular benchmarks strongly demonstrate that our ManiNet suffices to achieve competitive performance with simpler modeling and higher robustness, compared to baselines.