Spiking Neural Network (SNN)

Unlike conventional artificial neural networks that perform dense matrix multiplications on every forward pass, SNNs operate on a sparse, event-driven principle. A neuron computes and emits a spike only when its accumulated membrane potential crosses a firing threshold. This means zero computation occurs during silence, leading to massive energy savings on neuromorphic hardware. For RF signal processing, this allows a receiver to remain in a near-dormant state until a valid preamble or signal of interest is detected, enabling always-on spectrum monitoring at microwatt power budgets.

The Leaky Integrate-and-Fire (LIF) model is the workhorse neuron of SNNs. It behaves like an RC circuit:

• Integration: Incoming spike trains charge the neuron's membrane potential over time.
• Leakage: The potential decays exponentially toward a resting state, enforcing a notion of recency.
• Fire & Reset: Upon crossing a threshold, the neuron emits a spike and resets its potential. This intrinsic temporal memory allows a single LIF neuron to process spatiotemporal patterns directly, making SNNs naturally suited for classifying dynamic RF signatures like transient interference or frequency-hopping sequences without requiring explicit recurrent connections.

Spike-Timing-Dependent Plasticity (STDP) is a biologically plausible, unsupervised local learning rule. It adjusts synaptic weights based solely on the precise relative timing of pre- and post-synaptic spikes:

• Long-Term Potentiation (LTP): If a pre-synaptic spike consistently precedes a post-synaptic spike, the connection is strengthened.
• Long-Term Depression (LTD): If the order is reversed, the connection is weakened. This causal learning mechanism allows SNNs to discover temporal correlations in data streams without global backpropagation. For on-device RF fingerprinting, STDP enables continuous, low-power adaptation to subtle hardware drift in transmitters without cloud connectivity.

The discontinuous nature of the spike emission function—a step function—has a gradient of zero almost everywhere, breaking standard backpropagation. Surrogate gradient methods solve this by substituting the non-differentiable spike function with a smooth, differentiable proxy during the backward pass only. Common surrogates include the fast sigmoid or arctangent functions. This technique, known as Backpropagation Through Time (BPTT) with surrogate gradients, enables direct supervised training of deep SNNs using GPU-accelerated frameworks like smTorch, achieving competitive accuracy with conventional RNNs on sequential RF tasks like automatic modulation classification.

SNNs encode information in two fundamentally different ways:

• Rate Coding: Information is represented by the average firing rate of a neuron over a time window. It is robust to noise but slow and energy-inefficient.
• Temporal Coding: Information is encoded in the precise timing of individual spikes, such as time-to-first-spike or phase-of-firing. A single spike can carry significant information, enabling ultra-low-latency decisions. For latency-critical RF applications like radar pulse deinterleaving, temporal coding allows an SNN to classify a waveform after receiving only a handful of spikes, achieving inference latencies measured in microseconds on neuromorphic hardware like Intel's Loihi 2.

SNNs achieve their theoretical energy efficiency only when deployed on neuromorphic processors that natively implement event-driven computation. These chips, such as Intel's Loihi 2 and IBM's NorthPole, replace synchronous clock-driven logic with asynchronous digital circuits:

• Address Event Representation (AER): Spikes are communicated as lightweight packets containing a neuron address and timestamp.
• In-Memory Computing: Synaptic weights are stored locally at the neuron, eliminating the von Neumann bottleneck. This co-design enables RF edge devices to perform complex signal classification directly on the sensor, operating within a strict milliwatt power envelope unsuitable for conventional deep learning accelerators.

A biologically plausible unsupervised learning rule where synaptic weights are adjusted based on the precise relative timing of pre- and post-synaptic spikes. If a pre-synaptic neuron fires just before a post-synaptic neuron, the connection is strengthened (long-term potentiation). If the order is reversed, it is weakened (long-term depression). This temporal causality replaces traditional backpropagation in many local learning implementations.

• Mechanism: Weight change is a function of the time difference between spikes.
• Advantage: Enables continuous, online learning without global error signals.
• Use Case: Pattern recognition in event-based sensory streams.

The most common neuron model used in computational SNNs due to its balance of biological fidelity and computational efficiency. The neuron's membrane potential integrates incoming spike currents over time and leaks toward a resting state. When the potential crosses a firing threshold, the neuron emits a discrete spike and resets.

• Key Dynamics: Governed by a membrane time constant controlling the leak rate.
• Computation: Solves a simple differential equation, making it hardware-friendly.
• Contrast: Simpler than the Hodgkin-Huxley model but captures essential temporal dynamics.

Bio-inspired cameras, such as Dynamic Vision Sensors (DVS) , that abandon fixed frame rates. Each pixel independently and asynchronously logs a spike event only when it detects a logarithmic change in brightness. This creates a native sparse spatiotemporal spike stream that feeds directly into SNNs without conversion.

• Latency: Microsecond-level reaction time vs. millisecond-level for traditional cameras.
• Dynamic Range: Exceeds 120 dB, handling extreme lighting conditions.
• Data Efficiency: Eliminates redundant static background data, reducing bandwidth.

A method for training SNNs directly using backpropagation by overcoming the non-differentiability of the spike function. During the forward pass, the neuron uses a hard threshold. During the backward pass, the derivative of the threshold is replaced with a smooth surrogate function, such as a fast sigmoid or arctan, allowing gradients to flow.

• Purpose: Enables gradient descent on discrete spike events.
• Frameworks: Implemented in libraries like snnTorch and Norse.
• Benefit: Achieves competitive accuracy with traditional ANNs on complex tasks.

Two fundamental encoding schemes for translating real-valued input into spike trains. Rate coding represents information by the firing frequency over a time window, averaging spikes. Temporal coding represents information by the precise timing of individual spikes, often using latency to first spike or phase shifts.

• Rate Coding: Robust to noise but slow and energy-inefficient.
• Temporal Coding: Ultra-fast and sparse but sensitive to jitter.
• Trade-off: Temporal codes unlock the full low-latency potential of SNNs but require precise hardware.