Spiking Neural Networks for Event-based Vision

Event: CVPR Event-based Vision Workshop, 17th June 2019, Long Beach, CA · Duration: 23 min · ▶ Watch on YouTube

Abstract

This talk traces the development of spiking neural networks (SNNs) for event-based vision, starting with early work on visual recognition using spatial filters and visual motion estimation with synaptic delays. It highlights the transition from hand-tuned models on small datasets to a more automated deep learning approach for SNNs. The presentation introduces SLAYER, a novel backpropagation algorithm designed for SNNs, addressing challenges like non-differentiability and error assignment in the temporal domain. Finally, it showcases SLAYER’s performance on various datasets, including NMNIST and DVS Gesture, and presents initial results on Intel’s Loihi neuromorphic chip.

Speakers

  • Garrick Orchard — Intel

Talks (1)

  • 00:00:00 — Garrick Orchard: Spiking Neural Networks for Event-based Vision
    • A presentation on the evolution of spiking neural networks for event-based vision, from basic feature detection and motion estimation to deep learning approaches and hardware implementation on Loihi.

Key Takeaways

  • Early SNN models for event-based vision used coincidence detection with spatial filters and synaptic delays for feature extraction and motion estimation.
  • The SLAYER algorithm provides a thoroughly derived and powerful framework for configuring SNNs, enabling deep spiking networks (10+ layers) to learn both weights and delays.
  • SLAYER addresses the non-differentiability of spike generation and the temporal error assignment problem in SNNs, allowing for learning of precise spike timings and/or rates.
  • SLAYER achieves state-of-the-art results on the NMNIST dataset and competitive performance on the DVS Gesture dataset compared to other SNN approaches.
  • Preliminary results demonstrate that SLAYER-trained algorithms translate well to limited precision neuromorphic hardware like Intel’s Loihi chip, maintaining high accuracy.

Methods / Models / Datasets Mentioned

  • HMAX
  • HFirst
  • SLAYER
  • DART
  • TrueNorth
  • MFCC-SOM-SNN
  • Spiking CNN and HMM

Topics

Spiking Neural Networks (SNNs) · Event-based Vision · Visual Recognition · Motion Estimation · Deep Learning for SNNs · Backpropagation in SNNs · SLAYER algorithm · Neuromorphic Hardware · Loihi · NMNIST

Notes

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