Neuromorphic Wireless Split Computing with Multi-Level Spikes

TL;DR

The paper addresses scaling neuromorphic inference by splitting spiking neural networks across two wireless devices and encoding information with multi level spikes. It introduces OFDM based digital and analog modulation schemes to transmit $m$ bit payload spikes between encoder and decoder SNNs, and analyzes the resulting trade offs under pilot based channel estimation and various power constraints. Through simulations and real world SDR experiments, it demonstrates that there is an optimal spike payload $m$ that depends on SNR and resource availability, with digital modulation performing better at high SNR and large $m$ and analog modulation excelling at lower SNR and smaller $m$, while highlighting the need for channel aware end to end tuning for analog links. The work provides a first comprehensive study of neuromorphic wireless split computing with multi level SNNs and offers practical guidance for deploying such systems on OFDM based radios, including insights into energy and latency-accuracy trade offs for edge inference.

Abstract

Inspired by biological processes, neuromorphic computing leverages spiking neural networks (SNNs) to perform inference tasks, offering significant efficiency gains for workloads involving sequential data. Recent advances in hardware and software have shown that embedding a small payload within each spike exchanged between spiking neurons can enhance inference accuracy without increasing energy consumption. To scale neuromorphic computing to larger workloads, split computing - where an SNN is partitioned across two devices - is a promising solution. In such architectures, the device hosting the initial layers must transmit information about the spikes generated by its output neurons to the second device. This establishes a trade-off between the benefits of multi-level spikes, which carry additional payload information, and the communication resources required for transmitting extra bits between devices. This paper presents the first comprehensive study of a neuromorphic wireless split computing architecture that employs multi-level SNNs. We propose digital and analog modulation schemes for an orthogonal frequency division multiplexing (OFDM) radio interface to enable efficient communication. Simulation and experimental results using software-defined radios reveal performance improvements achieved by multi-level SNN models and provide insights into the optimal payload size as a function of the connection quality between the transmitter and receiver.

Figures (11)

• Figure 1: (a) Neuromorphic wireless split computing architecture based on multi-level SNNs: Spikes exchanged between a transmitter and a receiver over a wireless channel include a payload of $m$ bits. (b) While the accuracy of a centralized implementation increases monotonically with the spike payload $m$shrestha2024efficienttheilman2024spiking, in the presence of communication constraints there is generally an optimized value of $m$ that balances the informativeness of each spike with the reduced accuracy of higher-rate transmission.
• Figure 2: Neuromorphic wireless split computing with multi-level spikes: (a) An SNN is split into an encoding SNN and a decoding SNN, which are connected over a wireless channel following a spilt computing architecture. (b) Unlike prior works skatchkovsky2020endchen2023neuromorphicchen2022neuromorphic10682971, the SNNs implement spiking neurons that communicate using multi-level spikes shrestha2024efficienttheilman2024spiking, adopting a multi-level leaky integrate-and-fire (M-LIF) neuron model. (c) The output of the encoding SNN is transmitted using either analog or digital modulation. In the analog implementation, each output neuron of the encoding SNN is assigned separate OFDM subcarriers. In contrast, in the digital implementation, the AER protocol is used to embed information about the neurons' identities. Overflow bits that do not fit the allocated OFDM symbols are discarded.
• Figure 3: Surrogate derivative \ref{['eq:gradient-2']} used for training SNN models with M-LIF neurons ($m=2$ and $\Gamma=1$).
• Figure 4: Timeline of the proposed neuromorphic wireless split computing system. Time is discretized into sensing slots $t=1,2,\ldots,T$, corresponding to the time period over which the neuromorphic sensor accumulates information before reporting the presence or absence of events, along with the corresponding payloads. The spikes produced at time slot $t-1$ are processed by the Tx, and the outputs of the encoding SNNs are transmitted over the air using $N^{\rm OFDM}$ OFDM symbols to the Rx during the following, $t$-th, sensing slot. The decoding SNN at the Rx then processes the received signals to produce an inference decision. Each sensing time step $t$ is typically much longer than the duration of an OFDM symbol.
• Figure 5: The experimental setup includes a DVS sensor, a transmitter and a receiver. The screen visualizes the event-based input of the DVS sensor (left), along with the corresponding received OFDM signal at the receiver and the gesture type detected by the decoding SNN (right).