Marian Verhelst (KU Leuven)
Real-time machine learning on constrained devices
The development of smart wearable and IoT devices encompasses unprecedented challenges for hardware designers. Traditionally, hardware design targets to find, at design time, the best trade-off between four conflicting design requirements: energy consumption, throughput, latency and reliability. The recent trend of integrating more and more smartness and machine learning in embedded devices, however, brings stringent bottlenecks on all four parameters: needing low-latency processing of humongous amounts of data in energy-scarce nodes under high-reliability needs. Traditional hardware design approaches, where a static system is designed through gradual refinements, will not work anymore. We have to find ways to efficiently develop cross-domain optimized systems with online agility trough self-learning and self-adaptivity capabilities. This will be illustrated with several recent realizations.