Event-Driven Convolution-Based Processing

Event: CVPR 2025 Workshop on Neuromorphic Vision · Duration: 25 min · ▶ Watch on YouTube

Abstract

The presentation delves into event-driven convolution-based processing, drawing inspiration from the brain’s efficient information encoding. It covers the evolution of event-driven sensors like the DVS and sensitive-DVS, and their integration into complex neuromorphic architectures. The speaker demonstrates how event-driven convolutions can be implemented on ASICs, FPGAs, and SpiNNaker, enabling fast, low-latency operations for tasks such as object tracking, winner-take-all (WTA) operations, and stereo vision. Furthermore, the talk introduces advanced learning applications like Spatio-Temporal Back Propagation (STBP) on spiking deep convolutional networks (ConvNets) and stochastic binary STDP, showcasing methods to reduce spike count, hardware resources, and energy consumption while maintaining high accuracy.

Speakers

  • Bernabé Linares-Barranco — Instituto de Microelectrónica de Sevilla (IMSE-CNM), CSIC & Univ.Sevilla

Talks (1)

  • 00:00:00 — Bernabé Linares-Barranco: Event-Driven Convolution-Based Processing
    • This talk explores event-driven convolution-based processing, highlighting its efficiency, low latency, and applications in neuromorphic systems, including custom ASICs, FPGAs, and SpiNNaker platforms.

Key Takeaways

  • Event-driven convolutions offer fast, pseudo-simultaneous processing, filtering temporal and spatial noise effectively.
  • They can implement Winner-Take-All (WTA) and competition kernels, suitable for various neuromorphic applications.
  • Implementations on ASICs, FPGAs, and SpiNNaker demonstrate the versatility and real-time capabilities of these systems.
  • Event-driven convolutions are highly beneficial for stereo event-matching, enabling efficient 3D reconstruction.
  • Efficient low-rate STBP on SpiNNaker with N-MNIST and 1-bit weight stochastic STDP significantly reduce hardware resources and energy consumption while maintaining accuracy.

Methods / Models / Datasets Mentioned

  • DVS (Dynamic Vision Sensor)
  • sensitive-DVS
  • CAVIAR Project
  • Address-Event-Representation (AER)
  • ASICs
  • FPGAs
  • SpiNNaker
  • STBP (Spatio-Temporal Back Propagation)
  • Stochastic Binary STDP
  • Convolutional AER Vision Architecture
  • Multi-kernel Conv.
  • LeNet (ConvNet)
  • Dense400 (Fully Connected)
  • N-MNIST dataset
  • Gabor filters
  • Mexican hat kernel
  • AER-NODE
  • USB-AERmini2
  • PyTorch
  • Deep SNN

Topics

Event-driven processing · Neuromorphic computing · Spiking Neural Networks (SNNs) · Convolutional Neural Networks (CNNs) · Hardware efficiency · Low-latency processing · Bio-inspired vision · Stereo vision · STDP (Spike-Timing-Dependent Plasticity) · STBP (Spatio-Temporal Back Propagation)

Notes

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