Memristive Cosine-Similarity-Based Few-Shot Learning with Lifelong Memory Adaptation

External-memory-based few-shot learning has emerged as a promising solution to deal with the shortage of labeled samples in traditional deep neural networks (DNNs). The external memories here are used to perform explicit similarity calculations. Despite several encouraging schemes were proposed to merge data storage and similarity calculation using non-volatile memories, two major challenges, including the long encoding bits and the degraded accuracy of hardware implementations, remain to be tackled. Herein, a cosine-similarity-based few-shot learning method with analog memristors is designed. A regulation block in the embedding net is used as a hardware-software co-optimization unit to reduce the encoding lengths of the embedding vectors. Then, a lifelong memory adaptation method, termed the online prototype augmentation, is proposed for the first time to promote the accuracy of extremely low-shot learning problems. Results show that the encoding bits are 4-32 times smaller than popular schemes. Using the Omniglot dataset, record accuracy (98.3% and 97.3% for the 5-way 5-shot and 5-way 1-shot problems, respectively) can be obtained for the few-shot learning with memristive cosine similarities. In addition, the online prototype augmentation method leads to at least approximate to 2% improvement in accuracy for 1-shot learning, which is outstanding compared to prior studies.