(Processing Systems Lab)

Machine Learning

Our current work in the area of Machine Learning has focussed on co-design of the back-propagation algorithm with circuit-level observations for low-power accelerators.

MATIC exploits observations in the back-propagation algorithm, together with the nature of memory read failures in SRAM designs. Such a technique allows voltage-overscaling for dramatic reductions in memory power while resulting in minor degradation in classification and regression tasks.

More details on this subject will be provided here:

  1. MATIC: A Low-Power Machine Learning Accelerator using Memory Adaptive Training and In-Situ Canaries