Time-Multiplexed Low-Rank Posteriors for Spiking Networks

Spiking neural networks (SNNs) are networks whose neurons talk in sparse binary pulses rather than dense floating-point numbers. On neuromorphic chips they classify event-camera streams at milliwatt power. Like ordinary networks, they are overconfident on inputs they have never seen. The standard Bayesian fix runs the network many times with different sampled weights and measures how much the answers disagree. On a neuromorphic chip that fix is a poor fit for two reasons: every extra pass multiplies the spike count and therefore the energy, and the chips have no on-chip random-number units for sampling weights. We propose to fit a weight posterior in a very small subspace, at most 16 directions, by evaluating the actual spiking network forward only. We then compress that posterior into a handful of fixed, deterministic weight perturbations ("sigma points"), each realized as a tiny rank-$r$ side branch of the spiking layers. During one forward pass, the network cycles through these branches across its internal time steps. The result is a sampling-free, single-pass, Bayesian-grounded uncertainty estimate whose energy overhead is set by the branch width rather than by a sample count.

The bet is that the low intrinsic dimensionality of posteriors removes both obstacles at once. Low-dimensional posteriors are well established for dense networks: Subspace Inference for Bayesian Deep Learning found that the choice of subspace matters more than the inference algorithm used inside it, and Bayesian Semi-structured Subspace Inference found that fewer than 16 dimensions can match full-space sampling. In so few dimensions the posterior can be fitted from forward-loss evaluations of the true, non-differentiable spiking network. That avoids building curvature from surrogate gradients, which are only an approximation to the true derivative. In so few dimensions the posterior can also be summarized by a short list of deterministic points that fixed hardware can store.

Core Claim

A $d$-dimensional subspace posterior ($d \le 16$) over low-rank adapter directions of a trained SNN can be fitted without surrogate curvature and deployed as $K = 2d$ deterministic rank-$r$ branches cycled across time steps. Within a fixed energy budget, measured in synaptic operations, this deployment should give better calibration and out-of-distribution detection than temporal dropout or post-hoc calibration. It should also approach the quality of $K$ separate forward passes, at roughly $1 + Kr/n$ of the single-pass cost, where $n$ is the layer width. That ratio counts the extra side-branch neurons relative to the base layer.

The closest prior work is Efficient Uncertainty Estimation in Spiking Neural Networks via MC-dropout. That paper (AOT-SNN) also uses the time steps of a single pass as surrogate samples. Its samples are random activation-dropout masks, which require random-number generation on the chip, and it has no posterior over weights. Bayesian continual learning via spiking neural networks places full-space variational posteriors over every synapse. That doubles storage, and the authors themselves note it needs sampling hardware that existing chips lack. The low-rank Bayesian adapters Bayesian Low-rank Adaptation for Large Language Models (Laplace-LoRA) and BLoB supply the adapter parameterization we reuse. They were developed for Transformers, and Laplace-LoRA relies on Kronecker-factored curvature from exact gradients, while BLoB draws $N=10$ random weight samples at inference. We differ from all four on the same point: the deployed object is a fixed, deterministic set of low-rank weight perturbations whose spread reflects a weight posterior.

Research Program

Subspace construction. We start from a deterministic SNN trained with surrogate gradients: leaky integrate-and-fire spiking layers, where each neuron sums its inputs with a decay and fires when a threshold is crossed, followed by a non-spiking membrane-potential readout as in AOT-SNN. We attach rank-$r$ adapters $\Delta W_\ell = B_\ell A_\ell$ to each spiking layer $\ell$ and to the readout, and run a short phase of averaged fine-tuning in which only the adapters move. Principal components of the adapter trajectory, following Minimal Ranks, Maximum Confidence, give $P$, a projection onto $d$ directions. Restricting $P$ to adapter coordinates guarantees that any point $w_0 + P\theta$, with $w_0$ the trained weights and $\theta \in \mathbb{R}^d$ the subspace coordinates, changes each layer by at most rank $r$. The adapter symmetry problem identified in Learning Adapter Rank via Symmetry Breaking does not affect us, because inference happens on $\theta$, not on the individual adapter factors.

Posterior fitting without surrogate curvature. Because $d$ is small, we fit $p(\theta \mid \mathcal{D})$ in two ways. The first is a Laplace approximation, a Gaussian centred at the best-fitting point, whose $d \times d$ Hessian is estimated by finite differences of the true spiking network's minibatch loss. The second is elliptical slice sampling in $\theta$, which also needs only forward evaluations. We then compare both against a Laplace fit whose curvature comes from surrogate gradients. This comparison directly tests the open question of whether surrogate gradients distort curvature. A temperature on the posterior is tuned on validation negative log-likelihood (NLL), which measures how much probability the model assigns to the correct answer.

Deterministic deployment. From the Gaussian $\mathcal{N}(\hat\theta, \Sigma)$ we form $K = 2d$ sigma points, $\theta_k = \hat\theta \pm \sqrt{d}\, \Sigma^{1/2} e_j$, where $e_j$ is the $j$-th unit vector. Each sigma point defines a fixed rank-$r$ branch per layer, wired in parallel with the frozen base weights. In time step $t$, only branch $k(t)$ is enabled. The readout keeps $K$ separate integrators, so each sigma point gets its own logit vector. Predictive probabilities are the average over sigma points, and disagreement among them supplies the epistemic signal.

The known risk is that hidden membrane potentials leak across time steps. Consecutive "samples" therefore blur together, which would shrink the variance. We will measure this blurring directly by comparing cycled variance with true $K$-pass variance. We will also compare per-step cycling with block cycling, in which each sigma point holds for a contiguous block of steps longer than the leak time constant $1/(1-\beta)$, where $\beta$ is the membrane decay factor.

Baselines

The naive reference point is the deterministic SNN with its raw readout. It has no added cost, and it fixes the energy floor and the overconfidence we are trying to remove. On DvsGesture, Bayesian continual learning via spiking neural networks reports an expected calibration error (ECE, the average gap between stated confidence and actual accuracy) of 0.088 for a frequentist SNN. We will reproduce our own figure on the same architecture rather than reuse theirs.

Temperature scaling on the readout logits, using one scalar fitted on validation data, is the cheapest credible competitor. It is run on the identical network with the identical validation split.

AOT-SNN is the strongest existing single-pass method and the most direct competitor. It reports ECE of about 0.04 on CIFAR-10 and 65.15% accuracy on CIFAR-100 (Sun et al.). We reimplement it on our backbone with the same time-step budget $T$ and the same hyperparameter search budget, and we tune its dropout rate on validation NLL. Beating it means matching its accuracy and in-distribution ECE while improving out-of-distribution AUROC (the probability that an unseen input is scored more uncertain than a familiar one) and ECE under corruption, at equal or lower synaptic operation count. The comparison is also made with no on-chip randomness, which AOT-SNN requires.

Last-layer Laplace on the readout integrator applies the analysis of Is the Last Layer Sufficient for Uncertainty Quantification? to SNNs. It costs almost nothing at inference, because the Jacobian is just the time-averaged feature vector. It is fitted with the same data and the same prior-tuning procedure as our method.

The full-space Bayesian SNN of Bayesian continual learning via spiking neural networks reports ECE 0.064 on DvsGesture with real-valued synapses. It serves as the reference for weight-space Bayesian quality, with roughly double the storage.

A deep ensemble of five SNNs, together with the same subspace posterior evaluated with $K$ full separate passes, forms the quality ceiling. I found no reported result on Lacuna for SNN deep ensembles on these benchmarks, so we run the ensemble ourselves.

Three ablations isolate the key ingredients. The first restricts the subspace to readout directions only, which tests whether spiking-layer directions matter. The second replaces sigma points with random Gaussian samples cycled over time, which tests the value of deterministic placement. The third replaces forward-only curvature with surrogate-gradient curvature.

Why not just…?

Why not just do Laplace on the readout layer, which is analytic and essentially free? This is the most serious threat. Is the Last Layer Sufficient for Uncertainty Quantification? argues that last-layer posteriors match full linearization in dense networks. Laplace-LoRA reports the opposite for adapters, with all-layer Laplace clearly beating last-layer Laplace. The readout-only ablation is the experiment that settles which pattern SNNs follow. Our method includes the readout case as a special case, so the comparison is clean. If readout-only matches the full method under corruption and on held-out classes, we will report that SNN uncertainty lives in the readout and recommend the free method. In that case the remaining contribution is the forward-only curvature analysis.

Why not just use AOT-SNN's temporal dropout? It already achieves single-pass uncertainty. However, its spread comes from random activation masks, not from parameters that fit the data, and it needs a random-number generator at every time step. The head-to-head comparisons on CIFAR-10-C and DvsGesture lighting shift will show whether deterministic weight-space points give stronger out-of-distribution signals. The no-RNG constraint stands as a hardware argument regardless of the outcome.

Why not just ensemble or run $K$ passes? Both multiply spikes and therefore energy by $K$, and ensembles also multiply memory. We include them as quality ceilings. The question is whether time-multiplexing recovers most of their benefit at a fraction of the energy, and the leakage measurement tells us exactly how much is lost.

Why not just temperature scaling? A single scalar cannot make the network more uncertain on inputs that are far from the training data than on familiar ones. We expect it to win on in-distribution ECE and lose on out-of-distribution AUROC. Reporting the two metrics separately makes this visible.

Proposed Experiments

The primary benchmark is DvsGesture, the 11-class event-camera gesture set. A Low Power, Fully Event-Based Gesture Recognition System reports 94.59% accuracy at 178.8 mW on TrueNorth. We use its lighting conditions as a natural distribution shift and hold out gesture classes as out-of-distribution inputs. CIFAR10-DVS adds a harder event-stream task; the only reported numbers on Lacuna are early baselines, the best being 29.67%. Static-image CIFAR-10, CIFAR-100 and CIFAR-10-C allow direct comparison with AOT-SNN. Metrics are accuracy, ECE, NLL and out-of-distribution AUROC, together with spike count and synaptic operations per inference. Energy is converted using accumulate-operation counts, and measured on Loihi if access permits; otherwise we label it simulated.

Three checks decide the proposal. First, the cycled sigma-point model must beat AOT-SNN on out-of-distribution AUROC and shifted-condition ECE within 1.2× its synaptic operations. Second, it must retain most of the $K$-pass variance; we will report the ratio of cycled to $K$-pass predictive variance and choose block cycling if per-step leakage collapses it. Third, multi-layer subspace directions must beat readout-only Laplace under shift. If this third check fails, the conclusion is that readout-only is enough, and we say so.