Technical Summary: General Self-Prediction Enhancement for Spiking Neurons

1. Problem Statement

Spiking Neural Networks (SNNs) offer significant energy efficiency through event-driven, sparse computation, making them ideal for resource-constrained edge hardware. However, training high-performance SNNs remains challenging due to the non-differentiable nature of binary spike events, which complicates standard backpropagation. While existing methods like surrogate gradient training and ANN-to-SNN conversion have achieved progress, they often face trade-offs between convergence speed, accuracy, and biological plausibility.

Crucially, mainstream SNNs largely ignore predictive coding, a core cortical mechanism where the brain predicts sensory inputs and encodes deviations (prediction errors) to optimize perception. Current SNNs typically lack neuron-intrinsic mechanisms to perform such local self-prediction, often relying on costly global recurrent connections or ignoring temporal dynamics entirely.

2. Methodology

The authors propose a Self-Prediction Enhanced Spiking Neuron method. This approach integrates a predictive coding mechanism directly into the dynamics of individual spiking neurons, compatible with various neuron models (e.g., LIF, PLIF, CLIF, IF) and network architectures.

Core Mechanism

The method introduces an auxiliary prediction current ($m^p_l[t]$) that is dynamically generated from a neuron's own history of inputs and outputs. This current modulates the membrane potential as an auxiliary input, effectively acting as a local self-prediction error signal.

The dynamics are defined as follows:

1. Prediction Error Calculation: The system computes a local error term based on the discrepancy between the current input ($x_{l-1}[t]$) and the scaled output spike ($s_l[t]/\tau_l$).
   $$\text{Error} \approx x_{l-1}[t] - \frac{s_l[t]}{\tau_l}$$
2. Prediction Current Update: This error is filtered through a low-pass mechanism (moving average) to generate the prediction current:
   $$m^p_l[t] = (1 - \tau^p_l)m^p_l[t-1] + \tau^p_l \left( x_{l-1}[t] - \frac{s_l[t]}{\tau_l} \right)$$
   Here, $\tau^p_l$ controls the update rate.
3. Membrane Modulation: The prediction current is added to the synaptic input current ($I_l[t]$) before the membrane potential update:
   $$I_l[t] = x_{l-1}[t] + m^p_l[t-1]$$

Operational Logic

The prediction current serves four specific functions based on the state of the neuron:

• Correct Prediction (Spike expected & emitted): The current slightly elevates the membrane potential, pre-activating the neuron for the next expected event.
• Correct Prediction (No spike expected & none emitted): The current remains low or negative, suppressing noise-induced spurious spikes.
• Missed Spike (Input high, no spike): The current increases positively, amplifying the signal to enhance sensitivity to future similar inputs.
• Unexpected Spike (Input low, spike emitted): The current updates negatively, encouraging the network to suppress over-responsive behavior in the future.

Gradient Analysis

The authors demonstrate that this mechanism creates continuous gradient paths that alleviate the vanishing gradient problem common in SNNs. Unlike standard backpropagation through time, which suffers from gradient sparsity when spikes are infrequent, the self-prediction current ($m^p_l[t]$) acts as a smooth, continuous memory unit.

• Direct Pathway: Gradients flow from future states ($m^p_l[t+1]$) directly to the current input ($x_{l-1}[t]$) via the prediction current, bypassing the sparse spike event at the current time step.
• Indirect Pathway: Gradients flow through the prediction current to the spike output, allowing the network to adaptively correct prediction errors using surrogate gradients.

Biological Plausibility

The design is grounded in biological observations, specifically mimicking the role of distal dendrites in pyramidal neurons. In biology, distal dendrites receive top-down inputs that modulate somatic responses without directly triggering spikes. Similarly, the prediction current acts as a subthreshold modulatory signal. The error-driven update rule mirrors synaptic plasticity, where prediction errors (e.g., strong input failing to trigger a spike) drive adjustments in synaptic efficacy.

3. Key Contributions

• Novel Neuron Mechanism: Proposes a prediction current based on self-prediction error that enhances arbitrary spiking neurons without requiring complex architectural changes.
• Theoretical Analysis: Provides a comprehensive analysis of the method's dynamics, gradient propagation properties (showing improved stability and reduced vanishing gradients), and biological interpretability.
• Broad Empirical Validation: Evaluates the method across diverse network architectures (ResNets, Spiking ResNets), neuron models (IF, LIF, PLIF, CLIF), and datasets (CIFAR-10, ImageNet-100, ImageNet-1K).
• Task Versatility: Validates the approach not only on image classification but also on sequential classification (Sequential CIFAR-10) and reinforcement learning tasks (MuJoCo benchmarks), demonstrating robustness across different temporal and decision-making contexts.

4. Experimental Results

The authors conducted extensive experiments to validate the method:

• Image Classification (CIFAR-10): Across five network architectures and four neuron types, the self-prediction enhancement consistently improved accuracy. For example, on SEW ResNet-34 with LIF neurons, accuracy increased from 95.42% to 95.86% (T=4) and 94.54% to 96.16% (T=8). Gains were observed in most configurations, with the method proving particularly effective for LIF, PLIF, and CLIF neurons.
• Large-Scale Datasets (ImageNet-100/1K): On ImageNet-100, S-LIF improved accuracy from 65.80% to 69.06%. On ImageNet-1K, S-PLIF improved accuracy from 64.25% to 64.45%, confirming generalization to complex scenarios.
• Sequential Classification: On the Sequential CIFAR-10 task (long-term temporal dependencies), the method yielded consistent improvements throughout training. For instance, S-LIF achieved 82.38% accuracy compared to 80.11% for the baseline. Ablation studies confirmed that preserving the full computational graph (including gradients through the spike signal) is crucial for performance.
• Reinforcement Learning: In continuous control tasks (Ant, HalfCheetah, Hopper, Walker2d), the method was integrated with TD3 algorithms. The self-prediction enhanced PLIF (S-PLIF) consistently outperformed baseline PLIF and, in several cases (e.g., Ant-v4 with Proxy Target framework), surpassed the performance of standard Artificial Neural Networks (ANNs). The average performance gain over ANN baselines was 3.28% for S-PLIF in the Proxy Target setting.

5. Significance and Claims

The paper claims that the proposed method offers a simple yet effective way to enhance SNNs by bridging the gap between computational efficiency and biological plausibility.

• Training Stability: By creating continuous gradient paths, the method mitigates vanishing gradients and stabilizes training, leading to higher accuracy across diverse architectures.
• Biological Alignment: The mechanism aligns with cortical principles of predictive coding, dendritic modulation, and error-driven plasticity, providing a more biologically grounded model for spiking neurons.
• General Applicability: The method is not restricted to specific network types or tasks; it functions as a general enhancement applicable to various neuron models and domains, including vision, sequence processing, and control.

The authors conclude that this approach establishes a promising direction for enhancing Spiking Neural Networks, enabling them to achieve performance comparable to or exceeding ANNs while maintaining their inherent energy efficiency and biological interpretability.