Hypersphere-Based Weight Imprinting for Few-Shot Learning on Embedded Devices

Weight imprinting (WI) was recently introduced as a way to perform gradient descent-free few-shot learning. Due this, WI almost immediately adapted for performing learning on embedded neural network accelerators that do not support back-propagation, e.g., edge tensor processing units. However, suffers from many limitations, it cannot handle novel categories with multimodal distributions and special care should be given avoid overfitting the learned embeddings training classes since this can have devastating effect classification accuracy (for categories). In article, we propose hypersphere-based approach is capable of networks in regularized, imprinting-aware effectively overcoming aforementioned limitations. The effectiveness proposed method demonstrated using extensive experiments three image data sets.