Event-Driven Backpropagation

• Event-driven backpropagation is a framework that computes exact gradients for spiking neural networks by triggering updates solely at discrete spike events, preserving temporal and spatial sparsity.
• It leverages adjoint methods and event-triggered jumps to handle discontinuities in neural dynamics, optimizing both computational load and memory usage.
• Empirical results on platforms like BrainScaleS-2 and SpiNNaker2 show significant energy savings and improved accuracy by processing only essential spike-event data.

Event-driven backpropagation is a class of algorithms that computes gradients for training neural networks—especially spiking neural networks (SNNs)—using only discrete spike events as triggers for calculation and communication, instead of relying on dense, time-stepped activity or continuous neurodynamical traces. Its central objective is to reconcile the non-differentiable, asynchronous nature of spiking systems with the requirements of gradient-based optimization, achieving both computational and memory efficiency while preserving the temporal and spatial sparsity inherent to neural event streams. Recent theoretical and practical advances have resulted in exact event-driven gradient rules for SNNs, rigorous hardware-compatible formulations, and several scalable implementations across diverse neuromorphic architectures.

1. Mathematical Foundations of Event-Driven Backpropagation

The event-driven backpropagation paradigm is fundamentally rooted in treating spikes as the atomic events for both forward state evolution and backward credit assignment. In the canonical continuous-time SNN setting, neural states are governed by ODEs with spike-triggered resets, typically:

$$\mathrm{(free\ dynamics):} \quad \begin{cases} \tau_{\rm m}\,\dot V_n(t) = -\,V_n(t) + I_n(t),\ \tau_{\rm s}\,\dot I_n(t) = -\,I_n(t), \end{cases}$$

where $V_n$ denotes membrane potential and $I_n$ synaptic current. Spikes are generated when $V_n$ reaches threshold, at which $V_n$ is reset and $I$ receives postsynaptic updates through the weight matrix $W$ (Wunderlich et al., 2020).

The supervised loss $\mathcal L$ is generalized to combine both event-time and continuously-accumulated components:

$$\mathcal L = l_{\rm p}(\{t_k\}) + \int_0^T l_V(V(t), t)\,dt,$$

allowing for loss functions on sets of spike times and smoothly varying voltage-based objectives.

To backpropagate errors through this hybrid system, event-driven rules are derived via the adjoint method with explicit handling of discontinuities at spike events. Two time-dependent adjoint variables $\lambda_V(t)$ and $V_n$0 integrate backward in time:

$V_n$1

with prescribed terminal and jump conditions at each event (Wunderlich et al., 2020).

Through these ODEs and event-time jumps, the gradient with respect to each synaptic parameter $V_n$2 is sparsified to spike events:

$V_n$3

yielding a gradient accumulation only at times when the pre-synaptic neuron $V_n$4 emits a spike—an explicit temporal and spatial sparsity (Wunderlich et al., 2020, Pehle et al., 2023, Béna et al., 2024).

2. Algorithmic Realizations and Variants

Event-driven backpropagation principles have been implemented in several algorithmic forms:

• EventProp: The prototypical event-driven algorithm, computes exact gradients using continuous-time adjoints and maintains sparsity by updating gradients only at spike events. Forward and backward passes are performed as event-driven simulations, accumulating gradients whenever a spike triggers (Wunderlich et al., 2020, Pehle et al., 2023, Béna et al., 2024).
• Spike-timing-driven (STD-ED) and Membrane-potential-driven (MPD-ED) rules: These replace time-stepped surrogate derivatives with updates triggered strictly by spike occurrence or membrane crossing, further optimizing for hardware execution by masking out non-event timepoints (Wei et al., 2024).
• Random feedback alignment (eRBP): An event-driven rule for spiking networks that propagates error signals through fixed, random feedback weights rather than exact inverse pathways, with updates triggered by presynaptic events and local membrane state (Kaiser et al., 2019).
• Spatio-temporal plausibility extensions: Methods introduce spatial and temporal adjustments to gradients, e.g., scaling updates according to the instantaneous membrane voltage at spike and allowing errors to propagate across spikes via residual terms, thereby enhancing biological plausibility and temporal credit assignment (Shen et al., 2021).
• SpikeTrain-level backpropagation for RSNNs: Gradients are computed at the level of spike-train accumulations (e.g., spike-train post-synaptic potentials) rather than dense time-series, enabling efficient exact gradients in deeply recurrent SNNs (Zhang et al., 2019).
• Discretized error propagation (SpikeGrad): Both activations and backpropagated errors are discretized into signed spike events, with the entire training loop constructed from event-triggered comparator and accumulator operations (Thiele et al., 2019).

3. Hardware Implementations and Scalability

Event-driven backpropagation is specifically engineered for neuromorphic hardware, leveraging sparse, asynchronous event communication for both forward and backward propagation:

• BrainScaleS-2: EventProp has been implemented on this analog-digital mixed-signal platform, requiring only spike times and optional voltage measurements to compute exact gradients. Information efficiency is improved by an order of magnitude compared to surrogate-gradient approaches, directly translating to energy savings and enabling larger-scale learning experiments (Pehle et al., 2023).
• SpiNNaker2: Batch-parallelized, on-chip training using discretized EventProp equations exploits the manycore event-driven hardware for real-time operation while preserving spike and error sparsity. Event packets carry error signals alongside spikes, and each event triggers localized updates without dense synchronization (Béna et al., 2024).
• On-chip realizations: Event-driven rules support fully local learning, asynchronous event-triggered updates, and robust operation with device mismatch, as confirmed in mixed-signal circuit prototypes (Cartiglia et al., 2021). Simulations indicate substantial energy and latency reduction over traditional approaches (Shen et al., 2021, Wei et al., 2024).

The table below summarizes key reported energy or information-efficiency metrics:

Hardware/System	Gradient Method	Info/Energy Gain	Reference
BrainScaleS-2	EventProp vs. Surrogate	6–10$V_n$5 fewer bits	(Pehle et al., 2023)
ANP-I CMOS chip	MPD-ED vs. STBP	30$V_n$6 less energy	(Wei et al., 2024)
SpiNNaker2	EventProp vs. GPU	$V_n$7100$V_n$8 less power	(Béna et al., 2024)

4. Comparative Analysis: Exactness, Approximation, and Surrogate Gradients

Unlike backpropagation-through-time (BPTT) with surrogate gradients—which require dense memory and approximate the spiking Jacobian via smoothed transfer functions—event-driven backpropagation provides either exact gradients (as in EventProp and ST-RSBP) or provable unbiased weak derivatives (as in functional backpropagation through binning (Chen et al., 13 Feb 2026)):

• Exactness: Event-driven adjoint equations, with proper jump handling at spike events, enable calculation of the exact derivative of the network loss with respect to all parameters, without heuristic approximations or temporal unrolling.
• Surrogates and approximations: Earlier methods, and some scalable variants, substitute the spike nonlinearity with a smooth surrogate, trading off strict biological detail for computational tractability. Recent advances minimize these approximations by localizing them only to spike events or by masking derivatives, thus restoring substantial event-parsimony (Lee et al., 2016, Wei et al., 2024).
• Functional backpropagation: For event binning or similar nonsmooth preprocessing, weak derivatives are synthesized by lifting to continuous functionals and reconstructing cotangent functions, yielding unbiased gradients even through discontinuous operations (Chen et al., 13 Feb 2026).

5. Empirical Results and Performance Benchmarks

Multiple works have reported competitive or state-of-the-art performance for event-driven backpropagation in both supervised classification and neuromorphic hardware benchmarks:

• Yin-Yang (toy classification): EventProp trains a two-layer LIF SNN to $V_n$9 accuracy (Wunderlich et al., 2020).
• MNIST (digit classification): EventProp achieves $I_n$0, matching or surpassing single-layer surrogate-based approaches (Wunderlich et al., 2020).
• CIFAR-10 and DVS-Gesture: STD-ED and MPD-ED exhibit improvements of up to $I_n$1 accuracy and $I_n$2 energy reduction over time-stepped surrogates (Wei et al., 2024).
• Neuromorphic system comparison: EventProp consistently achieves higher accuracy and longer feasible experiment durations by transmitting only spike-event data (Pehle et al., 2023, Béna et al., 2024).
• Sparsity-controlled training: Explicit inclusion of spike count in the loss, controlled by annealing schedules, yields up to $I_n$3 reduction in spikes with negligible accuracy loss (Allred et al., 2020).

6. Biological Plausibility, Local Learning, and Extensions

There is convergent evidence from algorithmic, hardware, and theoretical studies that event-driven backpropagation possesses features increasingly aligned with biological learning:

• Local rules: Weight updates depend exclusively on locally available variables and spike events—either postsynaptic events, local error currents, or compartmental current differences—without the need for global gradients or clocks (Cartiglia et al., 2021, Shen et al., 2021, Greedy et al., 2022).
• Temporal causality: Learning signals propagate backward only as actual spikes or bursts, preserving the temporal and spatial sparsity of biological networks (Wunderlich et al., 2020, Greedy et al., 2022).
• Error multiplexing: Features such as burst multiplexing, short-term plasticity, and dendritic error encoding enable backpropagation-like error transport in multilayer architectures subject to biological constraints (Greedy et al., 2022).
• Hardware robustness: Event-driven, local-learning methods tolerate device mismatch and analog noise, providing a clear path to robust on-chip lifelong learning (Cartiglia et al., 2021, Béna et al., 2024).
• Extension to non-spiking and hybrid operations: Functional backpropagation methods generalize event-driven derivatives to broader classes of discontinuous operations and nonuniform sampling schemes (Chen et al., 13 Feb 2026).

7. Limitations, Open Challenges, and Prospects

While event-driven backpropagation addresses many limitations of dense or surrogate-based learning, challenges persist:

• Scalability to very deep or temporally extended networks: Storage of event data and discretization errors may limit scalability; higher-order solvers and hybrid event-stepped approaches are areas of ongoing development (Béna et al., 2024).
• Application beyond SNNs: Extending exact event-driven gradient rules to network types with more complex dynamics, gating, or plasticity architectures—such as eGRU, meta-learning, and continuous adaptation—remains an open research area (Béna et al., 2024, Greedy et al., 2022).
• Feedback routing and communication bottlenecks: Implementing scalable event-driven backward paths in very large or deeply hierarchical spiking arrays requires further innovation in error signal routing and compression (Cartiglia et al., 2021, Pehle et al., 2023).
• Alignment with strict biological detail: While plausibility has improved, some mechanisms (e.g., symmetry in feedback weights, accumulation of real-valued adjoints) remain neurobiologically idealized; yet recent models, such as BurstCCN, demonstrate approaches that approximate backprop-derived gradients within the constraints of cortical circuitry (Greedy et al., 2022).

Event-driven backpropagation thus constitutes a foundational approach for enabling rigorously trainable, scalable, and energy-efficient spiking and neuromorphic neural systems, providing theoretical and practical advances toward both large-scale artificial intelligence and hardware-embedded adaptive computation.