tinyML Talks: Multi-armed Bandit on System-on-Chip: Go Frequentist or Bayesian?

“Multi-armed Bandit on System-on-Chip: Go Frequentist or Bayesian?”

Sumit J Darak
Associate Professor
IIIT-Delhi

This talk will discuss the different multi-armed bandit-based online learning algorithms and their applications. We will highlight the need to realize these algorithms on edge platforms for wireless radio, Internet of Things (IoT), and robotics applications and the various challenges of mapping these algorithms on SoC. Some of our proposed architectures for three well-known algorithms: Upper Confidence Bound (UCB), Kullback Leibler UCB (KLUCB), and Thompson Sampling will be presented. We will then explore hardware-software co-design and fixed-point analysis for the efficient realization on heterogeneous SoC. To support future richness and flexibility in dynamic environments, intelligent reconfigurable architecture will be discussed.

source