Explore the bio-inspired approach to designing power-efficient hardware for AI in this tinyML Talks France presentation. Delve into the challenges of embedded AI, focusing on energy dissipation and data movement. Learn about current hardware accelerator designs, weight quantization techniques, and activation sparsity. Compare state-of-the-art technology with biological efficiency, and discover how brain-inspired circuits and technologies like Non-Volatile Memories can enhance data locality. Examine the potential of spiking neural networks and their implementation using novel technologies. Gain insights into inmemory compute, future applications, and proof-of-concept designs. Discuss classification accuracy, scaling possibilities, and multicore implementations. Conclude with an overview of embedded RAM, tinyML resources, and research directions in compression, image classification, and unsupervised learning for power-frugal AI hardware design.
TinyML Talks France - How to Design a Power Frugal Hardware for AI - The Bio-Inspiration Path
Overview
Syllabus
Introduction
Welcome
Start
AI at the edge
Trends in computing
Inmemory compute
Nonwater memory
Memory requirements
Future applications
Bioinspired
Proof of concept
topology
classical domain
receptive fields
resistive rams
analog nuance
fabrication
classification accuracy
demo
scaling
multiple values
viability
unit topologies
multicore implementation
conclusion
embedded RAM
tinyML resources
Mass testing
Compression
Other labs
Image classification
Bioinspired path
Depth analysis
Unsupervised learning
Rules for unsupervised learning
Flexibility
Research
Sponsors
Taught by
tinyML
