On-Device Learning with Binary Neural Networks

Existing Continual Learning (CL) solutions only partially address the constraints on power, memory and computation of the deep learning models when deployed on low-power embedded CPUs. In this paper, we propose a CL solution that embraces the recent advancements in CL eld and the e ciency of the Binary Neural Networks (BNN), that use 1-bit for weights and activations to e ciently execute deep learning models. We propose a hybrid quantization of CWR* (an e ective CL approach) that considers di erently forward and backward pass in order to retain more precision during gradient update step and at the same time minimizing the latency overhead. The choice of a binary network as backbone is essential to meet the constraints of low power devices and, to the best of authors' knowledge, this is the rst attempt to prove on-device learning with BNN. The experimental validation carried out con rms the validity and the suitability of the proposed method.