On-Device-Learning for MicroControllers

Ultra-low power devices are now powered with on-board processing resources to extract meaninful information from sensor data using Deep Learning models. These models are trained using powerfull high-performance machines and then frozen to be deployed on scale on low-power Micro-Controller platforms.

This line of research tackles the common train-once-deploy-everywhere scheme by investigating if deep learning models can be updated directly on-device after deployment, in short On-Device Leaning (ODL).

PULPTrain-lib: Bringing ODL to MicroControllers

The library collects deep learning primitives for layer-wise forward and backward passes. the target platform is the PULP platform: a MCU-class architecture featuring multiple RISC-V cores with a customized ISA. Our efficient software solution exploits parallelization, loop unrolling and vectorized MAC instrunctions, with full- and half-precision floating point formats. The library is open-source on GitHub.

Continual Learning on MCUs

We formulated a cost model to bring continual learning with latent replays on multi-core MCUs. To reduce the memory requirement, we proposed to a low-bitwidth quantization for the replay tensors up to 7 bits for a lossless solution (missing reference). Thanks to our system architecture and software stack, we could scale the final accuracy after ODL without forgetting depending on the memory size to store the replay buffer.