Orlando Moreira

GrAI Matter Labs

Training AI for event-driven processors

Abstract

Event-driven processors (EDP) exploit the sparsity in Deep Neural Networks’ activation maps; on an EDP, neurons with zero-valued activations result in no processor activity. To improve the performance of event-driven processors it is key to minimize non-zero activations.
STAR is a new training schedule that achieves this by combining activation regularization and thresholding to improve accuracy and reduce activations.
Performance measurements on an EDP from GrAI Matter Labs show that STAR leads to dramatic improvements in latency, frame rate, and energy efficiency. As such, STAR makes event-driven processors a compelling alternative to GPUs for Edge computing for Edge/Embedded applications.

Biography

Orlando Moreira is Chief Architect and Fellow at GrAI Matter Labs, a neuromorphic chip design startup. He graduated in Electronics and Telecommunications Engineering from the University of Aveiro. He holds a Ph.D. degree in Electrical Engineering from the Eindhoven University of Technology, with a thesis on real-time analysis and scheduling for streaming in embedded multiprocessors.

During his career, he worked at Philips Research (2000-2007), NXP Semiconductors (2007-2008), ST-Ericsson/Ericsson (2008-2015), and Intel (2015-2018). He published research on compilers, computer architecture, optimized training for edge processors, design flow automation, reconfigurable computing, real-time multiprocessor scheduling, resource management, and temporal analysis of data flow.