Explore the cutting-edge research on co-designing machine learning models, computational precision, and circuits in the energy-accuracy trade-off space presented by Prof. Marian Verhelst at the tinyML Summit 2019. Delve into circuit-level choices and implications, architecture-level decisions, and algorithm-level precision considerations. Discover parameterized hardware energy/latency/area models and energy-based cross-layer optimization techniques. Learn about the need for flexible systems with cross-layer frameworks and examine cascaded networks for efficient face recognition, keyword, and speaker recognition. Gain insights into the future of embedded Deep Neural Networks in this informative 23-minute conference talk from the MICAS laboratories at KU Leuven's Electrical Engineering Department.
From Small to Tiny: Co-designing ML Models, Computational Precision, and Circuits - Energy-Accuracy Trade-offs
Overview
Syllabus
Intro
Circuit level choices
Circuit level implications
Architecture level choices (2)
Algorithm level choices: precision
Algorithm level choices: implications
Parametrized HW energy/latency/area model
Energy-based cross-layer optimization
Needs for flexible systems with cross-layer framework
Cascaded networks for efficient face recognition
Cascaded ML models for efficient keyword & speaker recognit
Towards embedded Deep Neural Networks
Taught by
tinyML
