SELF-ATTENTIVE NETWORKS FOR ONE-SHOT IMAGE RECOGNITION

Despite recent breakthroughs in applications of deep neural networks, learning from a limited number of examples remains a key challenge in academia and industry. A prototypical example is one-shot learning, in which we must make predictions when only given one sample of each unseen class. Most previous works either categorize instances by the pair similarities or use Long-Short Term Memory (LSTM) based method, however, they usually suffer from inter/intraclass variation problem or high computation cost. In this paper, we propose self-attentive networks to address the above issues in one-shot image recognition problem in which deep neural features and attention mechanism especially self-attention are employed. Experiments on Omniglot and Mini-Imagenet datasets show that our proposed self-attentive networks achieve the state-of-the-art accuracy on one-shot image recognition problem. In addition, owing to the low time complexity and parallelizable architecture, self-attentive networks require significantly less time to train and infer.