REALIZING THE PROMISE OF SPIKING NEUROMORPHIC HARDWARE

Event: CVPR 2019 Second International Workshop on Event-based Vision and Smart Cameras · Duration: 26 min · ▶ Watch on YouTube

Abstract

Mike Davies from Intel Labs presents the Loihi neuromorphic research chip, designed for efficient, brain-inspired computation. He highlights the limitations of traditional Von Neumann architectures for intelligent workloads and introduces spiking neural networks (SNNs) as a promising alternative. The talk details Loihi’s architecture, its fully digital asynchronous design, and its software development kit. Benchmarking results are presented for diverse problems including keyword spotting, sparse coding (LASSO), spike-based LSTMs, graph search, and constraint satisfaction, showcasing Loihi’s superior energy efficiency and performance, especially for real-time, low-latency, and scalable applications.

Speakers

  • Mike Davies — Director, Neuromorphic Computing Lab Intel Labs

Talks (1)

  • 00:00:00 — Mike Davies: REALIZING THE PROMISE OF SPIKING NEUROMORPHIC HARDWARE
    • An overview of Intel’s Loihi neuromorphic research chip, its architecture, software ecosystem, and benchmarking results across various AI and optimization problems, demonstrating significant energy and performance advantages for real-time, low-latency applications.

Key Takeaways

  • Intel’s Loihi chip demonstrates significant energy efficiency (5-10x lower) and performance advantages for real-time, low-latency AI tasks compared to conventional architectures, particularly for batch size 1 inference.
  • Neuromorphic architectures excel in scalability, maintaining performance per inference even as network sizes grow, unlike traditional systems bottlenecked by memory.
  • Spiking Neural Networks (SNNs) offer a broader algorithmic space beyond traditional deep learning, enabling new approaches for problems like sparse coding, graph search, and constraint satisfaction.
  • The research frontier involves developing new algorithms guided by neuroscience principles and mathematically formalizing them to unlock the full potential of neuromorphic hardware.

Methods / Models / Datasets Mentioned

  • Deep Learning / Artificial Neural Networks (ANNs)
  • DNN to SNN conversion
  • SNN backpropagation
  • Online SNN pseudo-backprop
  • Olfaction-inspired rapid learning
  • Dynamic Neural Fields
  • SLAM
  • Evolutionary search
  • Cortical models
  • Neural Engineering Framework (NEF)
  • Locally Competitive Algorithms (LCA)
  • Stochastic SNNs for solving CSPs
  • Parallel graph search
  • Phasor associative memories
  • Vector symbolic architecture (VSA)
  • Semantic Pointer Architecture (SPA)
  • LASSO
  • FISTA
  • LSTMs
  • LSNNs
  • Dijkstra's Algorithm
  • Watts-Strogatz network model
  • Sudoku
  • 4-coloring of world map
  • TensorFlow
  • Nengo
  • EONS
  • NRP
  • PyNN
  • ROS

Topics

Neuromorphic Computing · Spiking Neural Networks (SNNs) · Loihi Chip · Energy Efficiency · Low Latency · Brain-Inspired AI · Real-time Inference · On-chip Learning · Heterogeneous Architectures · Algorithm Discovery

Notes

Open for commentary — connections to other work, critiques, follow-up reading.