Neuromorphic Integrated Sensing and Communications

Jiechen Chen, Nicolas Skatchkovsky, Osvaldo Simeone

TL;DR

This paper introduces Neuromorphic Integrated Sensing and Communications (N-ISAC), a novel system that uses a single impulse radio (IR) signal for both transmitting digital data and detecting radar targets. By employing a Spiking Neural Network (SNN) at the receiver, N-ISAC efficiently decodes data and senses the environment simultaneously, offering significant advantages over traditional separate sensing and communication systems.

What Problem Does This Paper Address?

Modern wireless systems are increasingly tasked with not only communicating information but also sensing their surroundings. This is particularly relevant for applications like autonomous vehicles or the Internet of Things (IoT), where understanding the environment is crucial for safe and efficient operation. Traditionally, these two functions – sensing and communication – have been handled by separate systems, requiring dedicated hardware and radio resources, leading to inefficiencies in terms of energy consumption and spectral usage. The paper addresses this by proposing an Integrated Sensing and Communications (ISAC) approach, specifically leveraging the unique properties of impulse radio (IR) and the energy-efficient processing capabilities of neuromorphic computing. The challenge lies in designing a receiver that can perform both tasks concurrently using the same transmitted signal, optimizing for both communication accuracy and sensing reliability.

Methods

The core of the proposed N-ISAC system is a Spiking Neural Network (SNN) acting as the receiver. This SNN is designed to process signals transmitted using Impulse Radio (IR), specifically Pulse-Position Modulation (PPM).

Here's a breakdown of the key methods:

  • IR Transmission with PPM: Digital data bits are encoded by the position of a pulse within a time slot. This is a low-power and potentially low-latency modulation scheme.
  • Neuromorphic Receiver (SNN): Instead of traditional digital signal processing, an SNN is used. SNNs are inspired by the brain's structure and function, processing information through "spikes" (binary events) and maintaining an internal "membrane potential." This allows for highly energy-efficient processing, especially for event-driven data.
  • Dual-Purpose SNN: The SNN at the receiver has two output neurons: one dedicated to decoding the digital data (estimating the transmitted bit) and another for radar sensing (estimating the presence or absence of a target).
  • Supervised Learning for SNN Optimization: The SNN's parameters (synaptic weights, etc.) are trained using a supervised learning approach. A dataset of transmitted signals, corresponding received signals, and ground truth labels for both data and target presence is used.
  • Joint Loss Function: A novel training criterion is proposed that combines two cross-entropy losses: one for communication (data decoding) and one for sensing (radar detection). A weighting factor ($\beta$) allows for balancing the priority between these two tasks.
  • Surrogate Gradients: To handle the non-differentiable nature of spike generation in SNNs during training, surrogate gradient methods are employed.

What's new here is the integration of neuromorphic computing (SNNs) with IR-based ISAC. While IR and neuromorphic computing have been explored separately for communication or inference, this paper pioneers their combination for simultaneous sensing and communication.

Datasets

The paper describes the generation of synthetic datasets for training and evaluation.

  • Training Data: 60,000 examples were generated. Each example consists of:
  • A sequence of transmitted bits ($x$).
  • The corresponding received signal samples ($y$) after passing through a simulated multipath channel with potential clutter and a target.
  • A binary label ($v$) indicating the presence (1) or absence (0) of a radar target.
  • Test Data: 10,000 examples were generated using the same methodology.

The authors do not introduce a new public dataset but rather generate synthetic data based on established models for IR channels, multipath fading, and radar target detection.

Key Insights and Ideas

  • Synergy between IR and Neuromorphic Computing: The paper demonstrates that the event-driven nature of IR signals (pulses occurring at specific times) aligns well with the spike-based processing of SNNs. This natural fit allows for efficient extraction of information for both communication and sensing.
  • Single Receiver for Dual Tasks: The most significant insight is that a single SNN architecture can be trained to perform both data decoding and radar detection simultaneously, eliminating the need for separate hardware and processing chains.
  • Adaptive Energy Consumption: The SNN receiver can adapt its computational activity based on the input signal's sparsity. When the IR signal is less active (e.g., during idle frames), the SNN consumes less energy, highlighting a key advantage of neuromorphic systems for energy efficiency.
  • Trade-offs are Manageable: The paper shows that by adjusting the weighting factor ($\beta$) in the joint loss function, the system can be tuned to prioritize either communication or sensing, demonstrating a flexible trade-off management.
  • SNN Capacity Matters: The performance of the N-ISAC system, particularly its ability to optimize both tasks, is shown to be dependent on the SNN's capacity (number of hidden neurons). A larger SNN (10 hidden neurons) significantly outperforms a smaller one (6 hidden neurons) and conventional SSAC systems.

Key Results

  • Performance Gains over SSAC: The proposed N-ISAC system consistently outperforms conventional Separate Sensing and Communication (SSAC) systems in terms of both normalized test throughput (communication accuracy) and radar test detection error.
  • Trade-off Curve: Figure 4 shows that N-ISAC can achieve a better trade-off between throughput and detection error compared to SSAC. For instance, with 10 hidden neurons, N-ISAC can achieve higher throughput at a given detection error rate, or lower detection error at a given throughput.
  • Bandwidth Expansion Benefits: Figure 5 illustrates that increasing the bandwidth expansion factor ($L_b$) benefits both ISAC and SSAC by reducing inter-slot interference. However, N-ISAC shows a greater capacity to leverage this expansion for improved performance, especially with a larger SNN.
  • SNN Size Impact: With 10 hidden neurons, the N-ISAC system's performance becomes less sensitive to the weighting factor $\beta$, indicating a more robust optimization. This suggests that a sufficiently capable SNN can inherently balance the two tasks.
  • Energy Efficiency: Figure 6 visually demonstrates the energy efficiency potential, showing a significantly reduced spike count in the SNN during idle transmission periods.

Limitations and Future Work

  • Synthetic Data: The experiments are conducted using synthetic data. While based on realistic models, real-world channel conditions and noise characteristics might introduce further challenges.
  • Fixed Target Delay: The current model assumes a known and fixed target delay ($T_0$). The paper mentions that generalizing to a time-varying target presence/absence pattern is straightforward, but extending to unknown or variable target delays would be a valuable future direction.
  • Limited SNN Architectures: The paper focuses on a specific SNN architecture with a single hidden layer. Exploring more complex SNN architectures or different neuron models could potentially yield further improvements.
  • Hardware Implementation: While the paper discusses the potential for efficient hardware implementation on neuromorphic chips, the experimental results are simulation-based. Real-world hardware implementation and testing would be a crucial next step to validate the energy efficiency claims.
  • Future Work: The authors suggest exploring more complex scenarios, such as dynamic target presence, and investigating different SNN architectures and training methodologies. The potential for N-ISAC in various IoT applications is also highlighted.

Summary updated

Abstract

Neuromorphic computing is an emerging technology that support event-driven data processing for applications requiring efficient online inference and/or control. Recent work has introduced the concept of neuromorphic communications, whereby neuromorphic computing is integrated with impulse radio (IR) transmission to implement low-energy and low-latency remote inference in wireless Internet-of-Things (IoT) networks. In this paper, we introduce neuromorphic integrated sensing and communications (N-ISAC), a novel solution that enables efficient online data decoding and radar sensing. N-ISAC leverages a common IR waveform for the dual purpose of conveying digital information and of detecting the presence or absence of a radar target. A spiking neural network (SNN) is deployed at the receiver to decode digital data and to detect the radar target using directly the received signal. The SNN operation is optimized by balancing performance metrics for data communications and radar sensing, highlighting synergies and trade-offs between the two applications.
BibTeX
@article{chen2023neuromorphic,
  author = {Jiechen Chen and Nicolas Skatchkovsky and Osvaldo Simeone},
  title = {{Neuromorphic Integrated Sensing and Communications}},
  journal = {IEEE Wireless Communications Letters},
  year = {2023},
  doi = {10.1109/lwc.2022.3231388},
}
Figure