A Pseudo-Spike Bridge for Federated Learning Across Array Fleets

What the work claims

This is a methods and benchmarking paper. The authors identify a gap that is easy to miss: federated learning assumes every participating device trains the same kind of network, but real fleets mix clients whose compute substrates force different representations. A conventional artificial neural network (ANN) passes continuous real-valued activations; a spiking neural network (SNN) passes binary spikes over time. Averaging model parameters across those two worlds, as standard federated aggregation does, does not just perform poorly, it can collapse training, because there is no shared parameter space to average in.¹

AS-FedBridge's claim is that a lightweight intermediate module, the Bridge, equipped with a Pseudo-Spike Interface, can serve as the common ground. The server aggregates only the bridge parameters; each client keeps its own backbone and a personalized head locally. Across four benchmarks under severe data heterogeneity (a ten-client federation with Dirichlet partitioning at alpha 0.1, five ANN clients and five spiking clients), the framework reports average accuracies of 91.12 percent on CIFAR-10, 71.64 percent on CIFAR-100, 56.12 percent on Tiny-ImageNet, and 82.06 percent on the neuromorphic event dataset CIFAR10-DVS, beating the strongest heterogeneous-federated baselines by up to 2.89 percentage points. Direct FedAvg-style aggregation, by contrast, fails outright in the same setting.

How it works

The mechanism is bidirectional distillation through a small shared network. Each round, a client runs the frozen global bridge on its data. Continuous-valued ANN clients see the bridge as an ordinary module. Spiking clients see a module whose interface projects features into a spike-compatible space: internal states are shaped so they can be interpreted as firing rates, with a polarization regularizer pushing them toward binary, spiking-like behaviour. Knowledge flows in both directions: the bridge distils what the client's backbone has learned, and the client distils the aggregated bridge back into its own network, with matching enforced at the bottleneck so firing-rate semantics survive the round trip.¹

The authors validate the alignment story with three diagnostics rather than accuracy alone. Centered Kernel Alignment (CKA) shows feature geometry stays similar across client types; maximum mean discrepancy shows the distribution gap between continuous and spike representations shrinks; gradient cosine similarity stays positive across all clients, meaning their local updates do not fight each other during aggregation. That last point is the operational one: federated averaging fails when client updates point in conflicting directions, and the bridge is what keeps them aligned.

The ablations quantify each component. Restricting transfer to one direction only costs at least 3.34 accuracy points. Removing the personalized head costs 3.49 points. Removing the pseudo-spike interface itself costs 1.81 points, so the alignment machinery is doing real work rather than acting as a passive relay. The communication accounting is the practical headline: the shared bridge weighs 3.93 MB per client, against 43.20 MB for full-model federated averaging of the same ResNet-18-scale networks, because no private backbone ever leaves its device.

Where a skeptic should push

The most load-bearing assumption is that closing the representational gap is the binding constraint, and the evidence for it is indirect. CKA, MMD, and gradient cosine are similarity diagnostics, not proofs; they show the bridge correlates with good federation, and the ablations strengthen the causal reading, but the paper does not demonstrate why the pseudo-spike projection is the right geometry rather than one of several workable ones. Bridge width sensitivity hints the design is tuned, not free: widening the bridge from the default changes CIFAR-100 accuracy from 71.64 down to 70.17 and 69.40 percent, so capacity and regularization interact in ways the paper maps empirically rather than explains.

Second, everything is an image classification benchmark. CIFAR10-DVS is neuromorphic in format, but it is still a converted vision dataset, not temporal control, not closed-loop stimulation, not neural recording. The accuracy margins over baselines are real but modest (up to 2.89 points), and the strong result against homogeneous alternatives (beating SNN-only federation by 10.58 points) partly reflects how weak single-paradigm federations are under alpha 0.1 partitioning, which is among the harshest non-independent-and-identically-distributed settings in common use. Third, the privacy claim is the standard federated-learning caveat: raw data stays local, but the paper does not evaluate inference or reconstruction attacks against the shared bridge, so the module should be treated as private-data-adjacent, not as a provably safe summary.

What a spike bridge means for the array fleet

The non-obvious implication is that the MEA field already has the heterogeneity this paper models, and it is worse than the paper's. Two laboratories both "doing MEA electrophysiology" may produce data that share no representation at all: one streams continuous 20 kHz waveforms from a CMOS array, another keeps only threshold-crossing timestamps and spike features, a third has an event-based address-event output straight from the headstage. Pooled models across such sites fail today for exactly the reason AS-FedBridge diagnoses: there is no common parameter space to average in, and the failure is silent, showing up as degraded accuracy rather than an error message.¹

The opportunity is governance as much as accuracy. Human-derived organoid and tissue recordings carry consent and provenance constraints that make shipping raw data across institutional borders slow or impossible. A bridge-style protocol offers a template: each site trains locally on its native representation, and only a small aligned module circulates. For rare phenotypes, where no single lab has enough data to train a reliable seizure, disease-state, or pharmacology classifier, that is the difference between a pooled model existing or not. The 3.93 MB versus 43.20 MB communication figure also matters operationally: multi-site collaborations on neural data routinely struggle with data-transfer agreements for terabytes; a federated loop that moves megabytes sidesteps that bottleneck entirely.

The threat side deserves equal weight. A shared bridge that misaligns does not fail loudly; it produces negative transfer, where pooled training is worse than training alone, and the paper's own baselines show the collapse mode is real. There is also a one-way-door risk in the spike direction: once a front end commits to thresholding or event encoding, the continuous signal is gone, and a bridge can only approximate what continuous clients know. A fleet that standardizes prematurely on spike-domain exchange may lock itself out of analyses that need the raw waveform. Finally, the bridge itself becomes an attack and inversion surface, so the governance story needs the same scrutiny as any shared representation, not the benefit of the doubt.

The bottom line

Established in simulation: mixed continuous-and-spiking federations can train a shared model through a small aligned bridge, with verified accuracy gains over heterogeneous-federated baselines and a large reduction in communicated parameters; naive parameter averaging demonstrably fails in the same setting. Hypothesis: the same architecture class can pool analytics across heterogeneous MEA installations without moving raw neural recordings. What would confirm it: a federated study on actual multi-site array data with representation mismatches (continuous versus spike versus event), measuring both accuracy and inversion-attack leakage against the shared module. What would break it: evidence that neural-signal representations diverge more sharply than vision features do, so the bridge needs per-site retraining so often that the federation saves nothing.