Spiking Neural Network PHY

Unlike Artificial Neural Networks (ANNs) that perform dense matrix multiplications on every timestep, SNNs operate on sparse binary spike trains. A neuron computes only when it receives an incoming spike, leading to massive computational savings. In a PHY context, this means the receiver's processing elements remain dormant during periods of silence or low signal activity, directly converting spectral sparsity into hardware energy efficiency. This is implemented using Leaky Integrate-and-Fire (LIF) or Izhikevich neuron models, where the membrane potential accumulates weighted input spikes and emits an output spike upon crossing a threshold.

SNNs possess native short-term memory through the decaying membrane potential of each neuron, eliminating the need for external memory modules like LSTMs or attention buffers for many temporal tasks. This makes them intrinsically suited for processing time-series RF data such as I/Q samples. The neuron's time constant can be tuned to match the coherence time of the wireless channel, enabling the network to naturally track phase drift and Doppler shifts without explicit complex-valued Kalman filtering. This temporal processing capability is critical for direct sequence detection in high-mobility scenarios.

The binary spike function is non-differentiable, preventing standard backpropagation. SNNs are trained using Surrogate Gradient (SG) methods, where the Heaviside step function is replaced with a smooth, differentiable proxy (e.g., a fast sigmoid or arctan) during the backward pass. This allows end-to-end training with frameworks like PyTorch and SpikingJelly. For PHY applications, this means an SNN-based receiver can be jointly optimized for channel estimation and symbol detection using stochastic gradient descent, learning to output spike-coded soft decisions that map directly to log-likelihood ratios (LLRs).

Information in SNNs is encoded in spike patterns. The two dominant paradigms are:

• Rate Coding: The firing rate of a neuron over a time window represents a value. This is robust but slow, requiring many timesteps for precision.
• Temporal Coding: The precise timing of individual spikes encodes information. Time-to-First-Spike (TTFS) coding is extremely efficient, as a neuron only needs to fire once. For PHY tasks like Automatic Modulation Classification (AMC), TTFS can encode the phase and amplitude of a received symbol into the precise firing delay of an input neuron, enabling ultra-low-latency classification with minimal spike counts.

The ultimate value proposition of SNNs is realized on dedicated neuromorphic processors like Intel's Loihi 2 or IBM's NorthPole. These chips implement in-memory computing and asynchronous digital logic that natively execute SNN dynamics without the von Neumann bottleneck. For a software-defined radio (SDR) PHY, an SNN compiled to Loihi 2 can perform continuous channel equalization at microwatt power levels, enabling persistent spectrum sensing in battery-constrained edge devices. The Lava software framework provides an open-source platform for mapping trained SNN PHY models to these asynchronous hardware architectures.

The conversion of high-precision floating-point I/Q samples into binary spike trains acts as a natural information bottleneck, providing inherent regularization and robustness to noise. This quantization effect, when combined with the stochastic nature of spike generation, makes SNN receivers remarkably resilient to adversarial perturbations and low-resolution Analog-to-Digital Converters (ADCs). The network learns to extract task-relevant features directly from the spike domain, effectively performing joint source-channel decoding that is optimized for the final decision metric, such as bit error rate.