Low-Shot Learning from Imaginary Data

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