Robust Hypersphere-based Weight Imprinting for Few-Shot Learning

Performing fast few-shot learning is increasingly important in a number of embedded applications. Among them, a form of gradient-descent free learning known as Weight Imprinting was recently established as an efficient way to perform few-shot learning on Deep Learning (DL) accelerators that do no support back-propagation, such as Edge Tensor Processing Units (Edge TPUs). Despite its efficiency, WI comes with a number of critical limitations. For example, WI cannot effectively handle multimodal novel categories, while it is especially prone to overfitting that can have devastating effects on the accuracy of the models on novel categorizes. To overcome these limitations, in this paper we propose a robust hypersphere-based WI approach that allows for regularizing the training process in an imprinting-aware way. At the same time, the proposed formulation provides a natural way to handle multimodal novel categories. Indeed, as demonstrated through the conducted experiments, the proposed method leads to significant improvements over the baseline WI approach.