Spiking Neural Networks - Part III: Neuromorphic Communications

Nicolas Skatchkovsky, Hyeryung Jang, Osvaldo SimeoneIEEE Communications Letters 2021 · DOI

TL;DR

This paper explores how Spiking Neural Networks (SNNs) can bridge the gap between efficient on-device AI and low-power wireless communication. It introduces two key applications: federated learning for distributed SNN training and an end-to-end neuromorphic system that integrates sensing, processing, and impulse radio transmission for ultra-low-power remote inference.

What Problem Does This Paper Address?

The paper tackles the growing challenge of deploying AI on resource-constrained edge devices, particularly in wireless communication scenarios. Traditional machine learning approaches, often reliant on powerful centralized computing, face limitations in terms of data privacy, energy consumption, and bandwidth. Simultaneously, wireless communication systems are becoming increasingly complex, demanding more efficient ways to transmit and process data. This work investigates how SNNs, with their inherent low-power, event-driven nature, can offer a compelling solution by enabling distributed learning and efficient end-to-end neuromorphic communication systems.

Methods

The paper presents two main approaches:

  • Federated Learning for SNNs (FL-SNN): This method adapts federated learning (FL) principles for training SNNs across multiple devices without sharing raw data. The key innovation lies in adapting FL's communication rounds to the fine-grained, time-based nature of SNNs, ensuring that local updates are synchronized with the SNN's internal time steps for effective collaborative training. The paper details the mathematical formulation for this process, including the use of an upper bound on the SNN's log-loss function for gradient descent.
  • NeuroJSCC (Neuromorphic Joint Source-Channel Coding): This approach proposes an end-to-end system that integrates neuromorphic sensing, SNNs for processing, and impulse radio (IR) for communication. The entire pipeline, from data acquisition to inference, operates using sparse, spiking signals. This allows for ultra-low-power remote inference by wirelessly transmitting only essential spiking information. The system is trained by maximizing the log-likelihood of the decoding SNN outputting desired spiking signals in response to input signals, effectively jointly optimizing compression and error correction.

Datasets

  • MNIST-DVS dataset: This dataset, consisting of handwritten digits captured by a neuromorphic camera, is used for evaluating both the FL-SNN and NeuroJSCC systems. The paper notes a specific preprocessing procedure that yields spiking signals over a defined time window.

Key Insights and Ideas

  • SNNs as a natural fit for edge AI and wireless: The event-driven and sparse nature of SNNs makes them inherently more energy-efficient than traditional ANNs, aligning well with the constraints of battery-powered edge devices and bandwidth-limited wireless channels.
  • Federated learning can be adapted for SNNs: By carefully scheduling communication rounds within the SNN's temporal dynamics, federated learning can be effectively applied to train SNNs in a distributed and privacy-preserving manner.
  • End-to-end neuromorphic systems offer significant advantages: Integrating neuromorphic sensors, SNNs, and impulse radio creates a holistic system that can achieve remarkable gains in both speed (time to accuracy) and energy efficiency for remote inference tasks.
  • Impulse radio is a promising low-power communication technology: The paper highlights IR's suitability for low-power communication, especially when combined with SNNs for encoding and decoding, as it leverages the timing of sparse spikes.

Key Results

  • FL-SNN: Experiments demonstrate that federated learning can successfully train SNNs for classification tasks. The paper shows that a balance between the frequency of local updates (∆t) and communication rounds (AJ) is crucial for achieving good performance, with lower-variance updates (larger ∆t) and frequent communication (smaller AJ) often yielding better results.
  • NeuroJSCC: The end-to-end neuromorphic system shows significant advantages over traditional approaches. In digit classification tasks, NeuroJSCC achieves higher accuracy with lower latency compared to uncoded transmission and frame-based Separate Source-Channel Coding (SSCC) schemes, especially under low Signal-to-Noise Ratio (SNR) conditions. The paper illustrates this with plots showing improved test accuracy as a function of observation time-steps and SNR. For instance, at an SNR of -8 dB, NeuroJSCC demonstrates a graceful increase in accuracy with more observation time-steps, outperforming other methods.

Limitations and Future Work

The authors acknowledge several open challenges and suggest future research directions:

  • Quantization and communication constraints: Further study is needed on the impact of quantization on SNN performance and how to optimize for communication constraints in neuromorphic FL.
  • Non-stationarity: The effect of non-stationary data and channel statistics on neuromorphic IR-based communication and learning needs to be investigated.
  • Complex environments: Evaluating the performance of these systems in more complex multi-path channels is an important next step.
  • Multi-access scenarios: Investigating the application of these techniques in multi-access communication scenarios is also suggested.
  • Privacy aspects: While FL inherently offers privacy benefits, further analysis of privacy aspects in neuromorphic FL is warranted.

Summary updated

BibTeX
@article{skatchkovsky2021spiking,
  author = {Nicolas Skatchkovsky and Hyeryung Jang and Osvaldo Simeone},
  title = {{Spiking Neural Networks - Part III: Neuromorphic Communications}},
  journal = {IEEE Communications Letters},
  year = {2021},
  doi = {10.1109/lcomm.2021.3050212},
}
Figure