Synchrony-Gated Plasticity with Dopamine Modulation for Spiking Neural Networks

AI summary

Overview

Research area: Spiking neural networks (SNNs) — specifically training methods for spiking transformers that blend supervised surrogate-gradient learning with biologically inspired local plasticity.

Technical level: Advanced. The paper assumes familiarity with surrogate backpropagation, spike-timing-dependent plasticity (STDP), Hebbian learning, spiking transformer architectures, and neuromodulation concepts such as three-factor learning rules.

Scope: The paper introduces DA-SSDP, a batch-level spike-synchrony plasticity rule gated by a dopamine-inspired scaling factor, and evaluates it as a drop-in training-time addition to the SpikingResformer backbone on CIFAR-10, CIFAR-100, CIFAR10-DVS, and ImageNet-1K.

What This Paper Is About

Spiking neural networks are trained almost entirely with surrogate backpropagation, and existing training-time regularizers (weight decay, dropout, label smoothing) operate on weights or real-valued activations while ignoring the spike event structure that actually drives computation. Biologically inspired local plasticity rules capture spike timing well, but they are hard to scale to deep multi-layer networks because they require detailed spike-timing logs, consume large amounts of memory, and their update directions can conflict with the supervised loss. This paper proposes a low-memory, synchrony-based local signal that is calibrated against the task loss during a short warm-up, then frozen and applied as a small post-backpropagation weight adjustment to deeper layers — with the goal of adding useful local information without changing the architecture, the optimizer, or the training schedule.

Key Contributions

1. Bringing a synchrony-based signal into model learning. DA-SSDP injects a batch-level synchrony-based complementary learning signal during training. Unlike STDP, it first screens for co-firing synchrony (a binary gate) before applying any bounded Hebbian-style adjustment, so the rule is independent of pre/post spike order. A brief warm-up phase derives a dopamine-inspired scaling factor from the empirical synchrony–loss relationship; after that the gate is fixed and only rescales the magnitude of the local update. When synchrony is task-irrelevant, the gate collapses toward unity and the rule safely reverts to a two-factor baseline with no online loss signal.

2. A scalable training-time synchrony update for deep SNN-transformers. The rule requires only binary spike indicators and first-spike latencies, adding O(C_out C_in) per batch over the pre-synaptic channels C_in and post-synaptic channels C_out of the hooked module, with element-wise per-batch operations and no inference-time cost. It is described as a drop-in implementation that touches only the final classification part and a simple late-stage feature projector, keeping the architecture and training schedule (epochs, batch size, optimizer) unchanged, and saving substantial memory compared with storing detailed spike-timing records.

3. Quantitative validation with safe degradation. Across CIFAR-10, CIFAR-100, ImageNet-1K, and the event-stream dataset CIFAR10-DVS, DA-SSDP shows consistent accuracy gains under the same training setup. The reported test-set median batch synchrony S_b rises from 3×10⁻⁴ to 1.0×10⁻² (roughly 33×).

4. Placement and hyperparameter robustness analysis. The authors show where the rule can be attached in the network and how it behaves under different potentiation/depression amplitudes, including a case where the two-factor version diverges but DA-SSDP remains stable.

Main Findings

• Accuracy gains over the same baseline: With identical model structures and comparable training configurations, DA-SSDP yields +0.42% on CIFAR-10, +0.99% on CIFAR-100, and +0.73% on ImageNet-1K. The abstract also lists CIFAR10-DVS at +0.1%. Model size and inference cost are unchanged; the rule adds only train-time computation at the hooked layers.

• Reporting convention: Table 1 reports Top-1 accuracy as mean (±) standard deviation over five runs. On CIFAR-10, SpikingResformer-CIFAR (10.83 M parameters, 4 time steps) goes from 95.95% to 96.15 ± 0.09 with SSDP without DA and 96.22 ± 0.1 with DA-SSDP. On CIFAR-100, the same architecture family (10.83 M parameters, 4 time steps) goes from 78.73% to 79.26 ± 0.22 (SSDP, no DA) and 79.48 ± 0.24 (DA-SSDP), with energy estimates of 0.493 mJ, 0.504 mJ, and 0.494 mJ respectively. On ImageNet, SpikingResformer-L (60.38 M parameters, 4 time steps) goes from 78.77% to 79.12 ± 0.23 (SSDP, no DA) and 79.29 ± 0.21 (DA-SSDP), with energy 8.76 mJ, 8.89 mJ, and 8.76 mJ.

• The simplified version does not degrade performance: The authors report that when the gate settles at one, simplifying DA-SSDP to the basic two-factor synchrony mechanism, the small weight updates (driven by concurrent spike firing and a Gaussian latency function) did not degrade performance and sometimes gave a small accuracy boost, acting as a regularizer during training.

• Event-stream data behaves differently: On CIFAR10-DVS (SpikingResformer-CIFAR, 17.31 M parameters, 10 time steps, 2.403 mJ), the baseline is 84.4%, with both SSDP without DA and DA-SSDP reaching 84.5 ± 0.1 (DA-SSDP energy 2.456 mJ, SSDP energy 2.421 mJ). The paper explains that many pre/post channel pairs never co-fire within T steps, so the synchrony gate λ is often zero, |Δt| is usually large, and the Gaussian weight g(Δt) is close to zero. Batch synchrony S_b varies only slightly, the loss correlation is weak, the fitted slope k is close to zero, and the dopamine gate is essentially neutral (G_b ≈ 1). This explains why the improvement on CIFAR10-DVS is modest compared with frame-based datasets, and indicates that the rule is most effective when the data provide strong, task-relevant spike synchrony.

• Placement matters: On CIFAR-100 with a baseline of 78.73% Top-1, attaching the rule to the last DSSA block gives +0.58% (two-factor SSDP) and +0.87% (DA-SSDP); attaching it only to the classifier gives +0.23% and +0.39% respectively; using both hooks gives the best result at 79.72% (+0.99%). Applying the update in the Prologue convolution destabilized training in their setting and diverged under the same schedule for both SSDP and DA-SSDP.

• Better stability under extreme potentiation/depression settings: Under high amplitudes (A+ = A− = 10⁻³), two-factor SSDP fails to converge, but DA-SSDP reaches 79.21%. At moderate amplitudes DA-SSDP matches or slightly exceeds SSDP (79.48% versus 79.40% at 1.5×10⁻⁴ / 5×10⁻⁵). The best accuracy is obtained with A+ = 1.5×10⁻³ and A− = 5×10⁻⁴, where DA-SSDP reaches 79.72% versus 78.86% for SSDP. The paper's Table 3 also lists SSDP at 0.00015/0.00005 reaching 79.48%, SSDP at 0.00010/0.00010 reaching 79.03%, DA-SSDP at 0.00015/0.0001 reaching 79.22%, and DA-SSDP at 0.00015/0.00005 reaching 79.40%.

• Spatial effect: Per-patch feature-gain maps computed as the absolute change in post-projection activation between the DA-SSDP and baseline checkpoints show hotspots that align with high-attention areas from the baseline attention maps, suggesting a re-weighting of already attended tokens rather than a wholesale redirection. Because the update is computed at the channel level after aggregating spikes across spatial locations, the gate and synchrony signal do not encode where in the image a spike occurred.

Methodology in Plain English

For each hooked module, the method looks at a mini-batch of size B with C_in input channels and C_out output channels. It keeps only cheap information from the forward pass: a binary indicator of whether each channel fired at least once in the time window, and the time of the first spike for each channel (channels that never fire are assigned t = T, the window length). Only the first spike is stored, based on prior findings that the first spike carries most of the critical information while later spikes contribute little and may add redundancy.

Two things are then combined per pre/post channel pair. A binary synchrony gate λ equals 1 only when both the pre- and post-synaptic channels fired at least once, which decides the sign of the update: co-firing leads to potentiation, no co-firing leads to depression. A Gaussian kernel of the absolute first-spike lag, g = exp(−Δt²/(2σ²)), scales only the magnitude, so nearby spikes produce stronger updates and the rule does not depend on spike order. Summing these terms across the batch and clipping the result to [−1, 1] gives the weight adjustment. Averaging the binary co-firing mask over all channel pairs gives a single scalar synchrony score S_b per sample, in the range [0, 1], which is comparable across layers with different channel sizes.

The dopamine part works in two stages. During a warm-up period (E_warm = 100 epochs, or 80 for CIFAR10-DVS), the network is not modified; instead the method collects per-batch pairs of synchrony S_b and supervised loss ℓ_b, standardizes both, and fits a single slope k as the negative empirical correlation between standardized synchrony and loss. That slope is then frozen, along with the synchrony mean and standard deviation. From then on, each batch gets a scalar gate G_b = clip(1 + k(S_b − μ_S)/σ_S, 0, 2) that uniformly rescales all the local updates in that batch without changing their direction. If synchrony barely varies or barely correlates with loss, k ≈ 0 and the gate is essentially neutral, reducing the method to a harmless two-factor baseline.

The updates are applied as a post-update hook: each training iteration starts with the usual surrogate-gradient update, then the binary activity markers and first-spike timings are used to compute an additive weight adjustment ΔW that is applied directly to the relevant weights. In the SpikingResformer backbone, two hooks are used — the linear classifier, and the 1×1 projection convolution of the last DSSA block in Stage 3, where the DSSA input channels are treated as pre-synaptic and the projection outputs as post-synaptic. The two DA-SSDP instances keep separate parameters and separate warm-up data. Kernel parameters are A+ = 1.5×10⁻³, A− = 1.0×10⁻⁴, and a learnable σ. Backbones use T = 4 steps on CIFAR-10/100 and ImageNet-1k, and T = 10 on CIFAR10-DVS, with each configuration repeated five times with different seeds. After calibration, (k, μ_S, σ_S) are fixed, while A+, A−, and σ remain trainable.

Why This Matters

The paper targets a specific gap: spiking transformers have advanced rapidly on architecture, but training still relies almost entirely on surrogate backpropagation, and the field lacks a scalable, structure-aware local signal that interacts with spike statistics. Conventional regularizers ignore event structure, while timing-based plasticity rules need dense temporal tracking and can conflict with gradient signals, confining most demonstrations to small-scale or online/RL settings. DA-SSDP reframes population-level spike synchrony — a quantity neuroscience highlights as a description of coordinated activity but which has rarely been used as a scalable supervised training-time signal in deep SNNs — as a low-memory substitute. Its practical selling point is that it improves accuracy without altering the forward pass, the model structure, or the training schedule, and with no inference-time cost.

Potential real-world applications:

• Event-driven vision systems using dynamic vision sensors, where the CIFAR10-DVS experiments give an initial (and, per the authors, limited) picture of how synchrony-based signals behave on sparse event streams.
• Edge and embedded inference on neuromorphic hardware, where the paper's emphasis on unchanged inference cost plus train-time-only overhead matters for deployment budgets.
• Energy-constrained always-on sensing pipelines, given the paper's reporting of energy estimates in millijoules alongside accuracy for SpikingResformer variants.
• General training pipelines for spiking transformer backbones where practitioners want a drop-in regularizer that does not require retuning epochs, batch size, or optimizer.

Industry relevance: Neuromorphic hardware and brain-inspired computing are cited as a growing area driven by demand for energy-efficient computing, with named platforms referenced in the introduction (Loihi, TrueNorth, SpiNNaker, and related work). Because the method leaves inference untouched and requires only binary indicators and first-spike latencies, it is positioned as a low-risk addition to existing spiking-model training stacks rather than a hardware redesign. The paper notes the code is available at https://github.com/NeuroSyd/DA-SSDP and that it was reviewed on OpenReview.

Future Directions

• Handling temporally sparse event data. The CIFAR10-DVS results show the rule falls back to a near-neutral gate when co-firing is rare and first spikes are far apart. A natural next step is a variant that remains informative under sparse, event-driven inputs rather than defaulting to the two-factor baseline.
• Extending placement beyond mid-to-late layers. Because the Prologue placement diverged under the same hyperparameters, and the largest gains came from the last DSSA block, the conditions under which early-layer synchrony becomes usable are left open.
• Making the dopamine slope adaptable after warm-up. The current design freezes (k, μ_S, σ_S) after calibration to avoid fluctuations late in training; whether a controlled update of the gate, or a different warm-up length, would help is not reported.
• Broadening the evaluation scope. Results are reported for one family of benchmarks and one backbone family (SpikingResformer) across five seeds per configuration; behavior on other spiking architectures, other modalities, and additional tasks is not reported.

Target Audience

This paper is most useful to researchers and engineers working on spiking neural networks and neuromorphic computing, particularly those training spiking transformers at scale who are interested in biologically inspired learning rules that coexist with surrogate backpropagation. It will also interest computational neuroscientists studying three-factor plasticity and dopamine modulation, and practitioners looking for a low-overhead, drop-in training-time regularizer for spiking models. Readers without background in surrogate gradients, STDP, or spiking transformer architectures will find the method section demanding.

Authors’ abstract

While surrogate backpropagation proves useful for training deep spiking neural networks (SNNs), incorporating biologically inspired local signals on a large scale remains challenging. This difficulty stems primarily from the high memory demands of maintaining accurate spike-timing logs and the potential for purely local plasticity adjustments to clash with the supervised learning goal. To effectively leverage local signals derived from spiking neuron dynamics, we introduce Dopamine-Modulated Spike-Synchrony-Dependent Plasticity (DA-SSDP), a synchrony-based rule that is sensitive to loss and brings a synchrony-based local learning signal to the model. DA-SSDP condenses spike patterns into a synchrony metric at the batch level. An initial brief warm-up phase assesses its relationship to the task loss and sets a fixed gate that subsequently adjusts the local update's magnitude. In cases where synchrony proves unrelated to the task, the gate settles at one, simplifying DA-SSDP to a basic two-factor synchrony mechanism that delivers minor weight adjustments driven by concurrent spike firing and a Gaussian latency function. These small weight updates are only added to the network`s deeper layers following the backpropagation phase, and our tests showed this simplified version did not degrade performance and sometimes gave a small accuracy boost, serving as a regularizer during training. The rule stores only binary spike indicators and first-spike latencies with a Gaussian kernel. Without altering the model structure or optimization routine, evaluations on benchmarks like CIFAR-10 (+0.42\%), CIFAR-100 (+0.99\%), CIFAR10-DVS (+0.1\%), and ImageNet-1K (+0.73\%) demonstrated consistent accuracy gains, accompanied by a minor increase in computational overhead. Our code is available at https://github.com/NeuroSyd/DA-SSDP.