Event-Based Backpropagation for SNNs

• Event-based backpropagation algorithms are defined as frameworks that compute gradients only during discrete spike events using adjoint methods or surrogate gradients.
• They leverage the intrinsic sparsity of spiking neural networks to reduce computational and memory costs, significantly improving energy efficiency.
• These methods naturally scale to neuromorphic hardware platforms, achieving competitive accuracies on benchmarks like MNIST and CIFAR-10 while minimizing communication overhead.

Event-based backpropagation algorithms provide a framework for training spiking neural networks (SNNs) and neuromorphic systems by propagating errors through discrete events (spikes) rather than through dense, synchronous state trajectories. This paradigm leverages the sparse, event-driven nature of SNNs to achieve efficient, scalable, and biologically inspired learning, especially tailored to both digital and analog neuromorphic hardware. Central to this approach is the derivation of exact or approximate gradients with respect to parameters, using temporally and spatially sparse computations, adjoint sensitivity methods, surrogate gradients, and event-driven communication protocols.

1. Mathematical Formulation of Event-Based Backpropagation

Event-based backpropagation formalizes gradient computation for SNNs by explicitly modeling the discontinuities and sparse updates that arise from the discrete nature of spike events. For a spiking network of leaky integrate-and-fire (LIF) neurons, the membrane dynamics are given by

$$\frac{dV}{dt} = \frac{-V + I}{\tau_{\mathrm{mem}}},\quad \frac{dI}{dt} = \frac{-I}{\tau_{\mathrm{syn}}}$$

where $V$ is the membrane potential, $I$ is the synaptic current, and $\tau_{\mathrm{mem}}$, $\tau_{\mathrm{syn}}$ are the respective time constants. Spikes are emitted when $V$ reaches a threshold $\theta$, triggering an instantaneous reset and synaptic current jump across the network.

EventProp, a canonical event-based backpropagation algorithm, applies the adjoint method with explicit jump conditions to properly handle the threshold-and-reset discontinuities at spike times. The adjoint variables $(\lambda_V, \lambda_I)$ evolve backward in time according to: $$\tau_{\mathrm{mem}}\,\lambda_V'(t) = -\lambda_V(t) - \frac{\partial l_V}{\partial V}, \quad \tau_{\mathrm{syn}}\,\lambda_I'(t) = -\lambda_I(t) + \lambda_V(t)$$ where $l_V$ is any membrane-voltage-based loss component. At each spike event $V$0 of neuron $V$1, the adjoint $V$2 undergoes a discontinuous jump (precise formula given in (Wunderlich et al., 2020)). The synaptic weight gradients are then computed as

$V$3

summed over all presynaptic spikes for neuron $V$4 and corresponding postsynaptic indices $V$5.

This formulation produces exact gradients—up to discretization error—without requiring approximation of the spike nonlinearity, and restricts the backward computation to the sparse set of recorded spike events (Wunderlich et al., 2020).

2. Algorithm Design and Technical Variants

Event-based backpropagation algorithms can be instantiated in several forms, varying by the treatment of neural dynamics, the handling of non-differentiability at spike events, and the computational model (continuous time vs. discretized time):

• Adjoint-based event propagation (EventProp): Computes gradients via exact adjoint ODEs with event-driven jump conditions, suitable for continuous-time SNN dynamics (Wunderlich et al., 2020, Pehle et al., 2023, Béna et al., 2024).
• Surrogate gradient BPTT: Replaces the non-differentiable spike function with a smooth surrogate (e.g., piecewise linear or sigmoid) in the backward pass, enabling BPTT across time-unfolded spiking networks (Allred et al., 2020, Lee et al., 2016, Kugele et al., 2021).
• Biologically plausible event adjustments: Introduce spatial and temporal corrections to the surrogate gradient, such as normalizing spatial gradients by the membrane potential at spike time and adding a temporal residual pathway, improving energy efficiency and stability (Shen et al., 2021).
• Random feedback and local learning rules: eRBP and other local rules use random projections and event-driven local computations for error propagation, suitable for ultra-efficient neuromorphic implementations (Kaiser et al., 2019).

Algorithmic steps commonly include an event-driven forward pass that records spike events, followed by a backward sweep, in which error signals and their adjoints are propagated (often only at spike times for exact algorithms), culminating in sparse, local weight updates.

3. Computational, Memory, and Energy Efficiency

Event-based backpropagation algorithms exploit temporal and spatial sparsity inherent in SNNs. The total computational and memory cost scales with the number of spike events $V$6 rather than the dense product of network size and simulation steps ($V$7) as in conventional BPTT. Only the events—spike times, neuron indices, and potentially voltage slopes—need to be buffered for the backward pass (Wunderlich et al., 2020, Pehle et al., 2023, Béna et al., 2024).

This results in memory and bandwidth requirements of $V$8, representing orders-of-magnitude reduction versus dense-sampling approaches, especially favorable for analog or mixed-signal neuromorphic substrates where continuous monitoring of membrane voltage is prohibitive (Pehle et al., 2023). Empirical measurements on platforms such as BrainScaleS-2 and SpiNNaker2 demonstrate 6–16× reductions in communication and energy per sample, with equivalent or superior learning accuracy compared to surrogate-gradient approaches (Pehle et al., 2023, Béna et al., 2024).

4. Extensions to Hardware and Scalability

Event-based backpropagation maps naturally onto digital and analog neuromorphic platforms. Implementations on BrainScaleS-2 utilize only spike event observations (optionally, voltage at event times), drastically reducing I/O and on-chip logging requirements (Pehle et al., 2023). On SpiNNaker2, forward and backward computations, as well as parameter updates, are parallelized across many-core arrays, with event packets handling spikes and error signals (Béna et al., 2024). On Intel Loihi, a synfire-gated event-based backpropagation protocol allows all learning phases (forward, error, update) to be expressed as spike events and coordinated by dynamic gating chains, supporting fully on-chip, low-power learning (Renner et al., 2021).

These architectures enable mini-batch or even batch-parallel training due to localized memory access patterns and distributed event-driven communication schemes. Bandwidth and DRAM scaling remain limiting factors for extreme network sizes, but off-chip gradient accumulation and batch-splitting can mitigate these constraints (Béna et al., 2024).

5. Comparisons to Surrogate Approaches and Training Results

Event-based backpropagation (EventProp) delivers competitive or superior performance to state-of-the-art surrogate-gradient SNNs, achieving MNIST test accuracy of 97.6% (single-layer SNNs, EventProp) and 96.1% for the Yin-Yang benchmark with strictly event-driven hardware-in-the-loop training (Wunderlich et al., 2020, Pehle et al., 2023, Béna et al., 2024). Surrogate-gradient methods with explicit spike sparsity regularization further improve inference energy/use and maintain iso-accuracy on tasks such as CIFAR-10, with up to 70% spike-count reductions (Allred et al., 2020, Kugele et al., 2021). Spatio-temporal corrections (e.g., temporal residuals (Shen et al., 2021)) yield state-of-the-art accuracy and lower latency compared to classical conversion-based SNN training on neuromorphic benchmarks.

Event-based random backpropagation and local plasticity rules (e.g., eRBP, eligibility-trace methods) yield lower hardware complexity but can lag in performance on temporally sensitive tasks due to the absence of precise credit assignment (Kaiser et al., 2019). Nevertheless, the event-driven structure supports highly efficient neuromorphic deployment for robotics and sensory tasks.

6. Open Challenges and Biological Considerations

While event-based backpropagation provides exact gradients for non-convex, discontinuous, and temporally sparse SNNs, it is not without limitations. The method assumes the recording of all spike times and certain derivatives, which may be infeasible at extremely high event rates. The gradient can diverge at points where the membrane potential velocity approaches zero at threshold (“tangent-to-threshold” events), which may require explicit gradient clipping or regularization (Wunderlich et al., 2020).

Extensions to adaptive thresholds, refractory periods, or multi-compartmental models are possible by generalizing the adjoint, event-jump formalism, but at the cost of recording additional event traces (Wunderlich et al., 2020, Pehle et al., 2023). The biophysical plausibility of exact gradients remains an open question; nevertheless, the event-driven nature closely parallels biological mechanisms, masking (by construction) the need for dense, global synchrony.

7. Outlook and Future Directions

Event-based backpropagation algorithms are positioned to enable fully event-driven, scalable learning in neuromorphic hardware, supporting both supervised and potentially unsupervised and reinforcement paradigms at network scales relevant to cortical computation. Anticipated extensions include multi-chip convolutional and recurrent SNNs, real-time in-hardware adjoint solving, and integration with more advanced biophysical neuron models and temporal codes (Pehle et al., 2023, Béna et al., 2024). Unified frameworks bridging functional, weak-gradient computation for event-binned representations further broaden the applicability of event-based backpropagation to event-driven vision and control pipelines (Chen et al., 13 Feb 2026). Such developments promise to further close the gap between the theoretical efficiency of event-driven information processing and the empirical demands of contemporary machine learning.