Low-Shot Learning from Imaginary Data

TL;DR

This paper introduces a low-shot learning method that uses a hallucinator to generate additional training data, significantly improving classification accuracy with few examples, inspired by human visualization abilities.

Contribution

It proposes a novel combination of meta-learning with a hallucinator to generate synthetic training data, enhancing low-shot learning performance.

Findings

Up to 6 point accuracy boost with one example

Achieves state-of-the-art results on ImageNet low-shot benchmark

Effective integration of hallucinator with various meta-learners

Abstract

Humans can quickly learn new visual concepts, perhaps because they can easily visualize or imagine what novel objects look like from different views. Incorporating this ability to hallucinate novel instances of new concepts might help machine vision systems perform better low-shot learning, i.e., learning concepts from few examples. We present a novel approach to low-shot learning that uses this idea. Our approach builds on recent progress in meta-learning ("learning to learn") by combining a meta-learner with a "hallucinator" that produces additional training examples, and optimizing both models jointly. Our hallucinator can be incorporated into a variety of meta-learners and provides significant gains: up to a 6 point boost in classification accuracy when only a single training example is available, yielding state-of-the-art performance on the challenging ImageNet low-shot classification benchmark.