Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EON-1: A Brain-Inspired Processor for Near-Sensor Extreme Edge Online Feature Extraction

Alexandra Dobrita, Amirreza Yousefzadeh, Simon Thorpe, Kanishkan Vadivel, Paul Detterer, Guangzhi Tang, Gert-Jan van Schaik, Mario Konijnenburg, Anteneh Gebregiorgis, Said Hamdioui, Manolis Sifalakis

TL;DR

EON-1 is proposed, a brain-inspired processor for near-sensor extreme-edge online feature extraction that integrates a fast online learning and adaptation algorithm that has results of only 1% energy overhead for learning, by far the lowest overhead when compared to other SoTA solutions, while attaining comparable inference accuracy.

Abstract

For Edge AI applications, deploying online learning and adaptation on resource-constrained embedded devices can deal with fast sensor-generated streams of data in changing environments. However, since maintaining low-latency and power-efficient inference is paramount at the Edge, online learning and adaptation on the device should impose minimal additional overhead for inference. With this goal in mind, we explore energy-efficient learning and adaptation on-device for streaming-data Edge AI applications using Spiking Neural Networks (SNNs), which follow the principles of brain-inspired computing, such as high-parallelism, neuron co-located memory and compute, and event-driven processing. We propose EON-1, a brain-inspired processor for near-sensor extreme edge online feature extraction, that integrates a fast online learning and adaptation algorithm. We report results of only 1% energy overhead for learning, by far the lowest overhead when compared to other SoTA solutions, while attaining comparable inference accuracy. Furthermore, we demonstrate that EON-1 is up for the challenge of low-latency processing of HD and UHD streaming video in real-time, with learning enabled.

EON-1: A Brain-Inspired Processor for Near-Sensor Extreme Edge Online Feature Extraction

TL;DR

EON-1 is proposed, a brain-inspired processor for near-sensor extreme-edge online feature extraction that integrates a fast online learning and adaptation algorithm that has results of only 1% energy overhead for learning, by far the lowest overhead when compared to other SoTA solutions, while attaining comparable inference accuracy.

Abstract

For Edge AI applications, deploying online learning and adaptation on resource-constrained embedded devices can deal with fast sensor-generated streams of data in changing environments. However, since maintaining low-latency and power-efficient inference is paramount at the Edge, online learning and adaptation on the device should impose minimal additional overhead for inference. With this goal in mind, we explore energy-efficient learning and adaptation on-device for streaming-data Edge AI applications using Spiking Neural Networks (SNNs), which follow the principles of brain-inspired computing, such as high-parallelism, neuron co-located memory and compute, and event-driven processing. We propose EON-1, a brain-inspired processor for near-sensor extreme edge online feature extraction, that integrates a fast online learning and adaptation algorithm. We report results of only 1% energy overhead for learning, by far the lowest overhead when compared to other SoTA solutions, while attaining comparable inference accuracy. Furthermore, we demonstrate that EON-1 is up for the challenge of low-latency processing of HD and UHD streaming video in real-time, with learning enabled.
Paper Structure (18 sections, 11 figures, 3 tables, 1 algorithm)

This paper contains 18 sections, 11 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Algorithm and refinements
  3. Overview of ultra-rapid feature extraction in the visual cortex
  4. Methods
  5. Low-latency, 1-spike--based inference
  6. Fast online learning (few-shots)
  7. Classifier
  8. Hardware architecture
  9. Edge-Filters and Lateral-Inhibition
  10. Synaptic Weights Memory and Neuron Parameters Memory
  11. Inference engine
  12. Learning engine
  13. Classifier circuit
  14. Results and discussion
  15. Experimental setup
  16. ...and 3 more sections

Figures (11)

  • Figure 1: The neural network structure for this work. It includes a layer of edge-filtering convolution, a lateral inhibition layer, and a layer of fully connected neurons equipped with binary STDP training.
  • Figure 2: The lateral inhibition layer produces binary spikes that can be compressed into a spike vector. Each element of the resulting vector designates the index of the source edge filter that generated the first spike for the respective pixel position.
  • Figure 3: Example of a spike vector and four weight vectors. Bold elements in the weight vectors are the one that match the spike vector. Assuming $T_{Fire} = 2$, only neuron 3 fires.
  • Figure 4: An example of our stochastic binary STDP algorithm. Top: neuron $3$ is eligible for learning and it has two ineffective synapses (blue). Ineffective spikes positions are highlighted in orange. Bottom: one ineffective weight is randomly selected and swapped with the ineffective spike.
  • Figure 5: Base hardware architecture of the EON-1 processor, equipped with an inference engine and a learning engine which embeds the proposed STDP-based learning rule. This architecture can be flexibly and trivially scaled-up by vectorizing either of: the IF units, the edge-filter units, and/or the sequential-learning processes (depending on design-space requirements)
  • ...and 6 more figures