Continual On-Device Learning on Multi-Core RISC-V Microcontrollers

Explore the latest advancements in Continual On-device Learning for Multi-Core RISC-V MicroControllers in this conference talk from tinyML EMEA 2022. Delve into the challenges of adapting Deep Learning models on deployed sensors and discover how Continual Learning methods can be efficiently implemented on low-power platforms. Examine the trade-offs between memory, energy consumption, and accuracy in back-propagation techniques using Latent Replays. Learn about PULP-TrainLib, a high-performance compute library for MCU-based learning, and its impressive performance improvements over existing solutions. Gain insights into quantization strategies, memory optimization, and the practical applications of on-device learning for tasks such as license plate recognition and speech enhancement.

Intro
An (Open) HW perspective
PULP-NN: Accelerating DNN Inference
License Plate Recognition
Speech Enhancement
Data challenge for the real-world
Catastrophic Forgetting
Continual Learning (CL)
CL with Latent Replays
Quantization and Memory Cost
Learning Kernels Latency on PULP
Learning a new object class