Timing-Based Backpropagation in Spiking Neural Networks Without Single-Spike Restrictions

Technical Summary: Timing-Based Backpropagation in Spiking Neural Networks Without Single-Spike Restrictions

Problem Statement
Spiking Neural Networks (SNNs) offer a biologically plausible framework for information processing by encoding data in precise spike timings. However, supervised learning in SNNs faces significant challenges, primarily the non-differentiability of the binary spike function. Existing approaches generally fall into two categories:

1. Rate-based methods: These approximate gradients using surrogate gradients (SG) and time discretization. While effective, they often fail to capture exact temporal coding properties, treating spikes more like continuous activation functions.
2. Timing-based methods: These compute exact gradients based on spike timing (e.g., SpikeProp, Mostafa). However, conventional timing-based methods typically impose a single-spike restriction, where each neuron fires at most once per input pattern (often using Time-To-First-Spike or TTFS coding). This restriction limits the computational capacity of SNNs and contradicts biological evidence suggesting that multi-spike firing is essential for efficient neural coding. Previous attempts to extend timing-based learning to multi-spike scenarios have not demonstrated performance superior to single-spike restricted models.

Methodology
The authors propose a novel backpropagation algorithm for feedforward SNNs that allows neurons to fire multiple times while maintaining exact gradient computation based solely on spike timing, without relying on surrogate gradients or heuristic approximations.

• Neuron Model: The study utilizes the Leaky Integrate-and-Fire (LIF) model in continuous time. A critical distinction from some existing multi-spike studies is the reset mechanism: only the membrane potential is reset to zero upon firing, while the postsynaptic current is not reset. This preserves the history of synaptic inputs, allowing past information to influence future spikes within the same neuron.
• Mathematical Formulation:
  • The authors define two spike order concepts: local spike order (the sequence of spikes for a single neuron) and global spike order (the sequence of all spikes across a layer).
  • By transforming time $t$ into a variable $z = \exp(t/\tau_I)$, the differential equations governing membrane potential and current are converted into algebraic forms. This allows for the explicit and analytical calculation of the next spike timing $z^{(n+1)}_j$ based on previous spike timings and synaptic weights.
  • The forward pass computes spike timings sequentially. The backward pass propagates errors through these exact spike timings, including the "reset paths" (gradients flowing through the reset of membrane potential), which are often ignored in surrogate gradient methods.
• Loss Functions:
  • Membrane Potential Loss: For classification, the membrane potential of output neurons is observed at a preset end time $t_{out}$ (without firing), and cross-entropy is calculated.
  • Dead Neuron Penalty Loss: To address the issue of "dead neurons" (neurons that never fire due to poor initialization, rendering them untrainable in timing-based methods), a penalty loss is introduced. This loss penalizes neurons that fail to fire for a sufficient number of input patterns, effectively pulling them back into the learning process.
• Efficiency: A two-stage partitioned algorithm is proposed to handle the computational and memory intensity of continuous-time dynamics. This approach decouples computation from specific time steps, allowing for parallel processing based on spike counts, which is optimized for GPU/TPU execution.

Key Contributions

1. Exact Multi-Spike Backpropagation: The paper presents the first supervised learning algorithm for multi-spike SNNs that computes exact gradients using only spike timing, eliminating the need for surrogate gradients or rate-based approximations.
2. Performance Benchmarking: Using the MNIST dataset, the authors demonstrate that multi-spike SNNs trained with this method outperform existing single-spike restricted timing-based SNNs and achieve accuracy comparable to non-convolutional Artificial Neural Networks (ANNs) and rate-based SNNs.
3. Characterization of Leakage Time Constants: The study reveals a unique relationship between the synaptic leakage time constant ($\tau_I$) and network performance in multi-spike SNNs. Unlike single-spike models, multi-spike networks exhibit an optimal leakage time constant (found to be $\tau_I = 0.8$ in their experiments) that maximizes test accuracy.
4. Spike Count Dynamics: The research shows that the distribution of spike counts per neuron is highly sensitive to leakage constants. Smaller time constants (higher leakage) lead to higher average spike counts and a shift toward multimodal distributions, whereas larger time constants cause the network to converge toward a lower, stable spike count.

Results

• Accuracy: On the MNIST dataset (784-400-10 architecture), the proposed multi-spike timing-based SNN achieved a test accuracy of 98.43 ± 0.05% with an optimal leakage constant. This surpasses the best single-spike timing-based results (97.96%) and is comparable to state-of-the-art rate-based SNNs and ANNs.
• Optimal Parameters: The experiments identified a clear optimum for the leakage time constant in multi-spike networks. In contrast, single-spike networks showed a monotonic convergence without a distinct peak in performance relative to the leakage constant.
• Sparsity: Multi-spike SNNs were found to be sparser (fewer spikes per neuron on average) than single-spike restricted SNNs across a wide range of leakage constants, suggesting more efficient information encoding.

Significance and Claims
The paper claims to provide a solid foundation for gradient-based learning in timing-based SNNs by removing the artificial single-spike constraint. The authors argue that their method:

• Validates Biological Plausibility: By allowing multiple spikes and computing exact gradients, the model better reflects biological neural systems where information is encoded in multi-spike timing.
• Improves Computational Capacity: The extension to multi-spike firing naturally enhances the information processing capacity of SNNs without sacrificing the precision of temporal coding.
• Solves Initialization Sensitivity: The introduction of the "dead neuron penalty" mitigates the dependency on initial weight variables, a known hurdle in timing-based learning where non-firing neurons cannot receive gradients.
• Energy Efficiency: The authors note that their model, which relies on precise spike timing, has the potential to contribute to the development of highly energy-efficient computing systems, as the learning process itself can be energy-efficient.

The study concludes that while the timing-based method faces inherent challenges regarding sudden spike variations due to weight updates, the proposed algorithm successfully demonstrates that multi-spike SNNs can achieve superior performance and exhibit unique learning dynamics distinct from their single-spike counterparts.