Structured Sparse Back-propagation for Lightweight On-Device Continual Learning on Microcontroller Units

With many devices deployed at the extreme edge in dynamic environments, the ability to learn continually on-device is a fast emerging trend for ultra-low-power Microcontrollers (MCUs). The key challenge in enabling Continual Learning (CL) on highly constrained MCUs is to curtail memory and computational requirements. This paper proposes a novel CL strategy based on sparse weight updates coupled with Latent Replay. We reduce the latency and memory requirements of the backpropagation algorithm by computing structured sparse update tensors for the trainable parameters, retaining only partial activations during the forward pass and limiting the per-layer gradient computation to a subset of channels. When applied to lightweight Deep Neural Network (DNN) models for image classification, namely PhiNet and MobileNetV2, our method can reduce up to 1.3 × the memory and computation costs of the backpropagation algorithm, with a minor accuracy drop (2%). Furthermore, we evaluate the accuracy-latency-memory trade-off, targeting a class-incremental CL setup on a RISC-V multi-core MCU. The proposed approach allows to learn on-device a new class-incremental task, composed of two unseen classes, in 18min with 4.63 MB considering the most demanding configuration, i.e., a MobileNetV2 trained on the CORe50 dataset.