Sensors age. Over months, a chemical sensor's baseline creeps, its gain sags, and a classifier trained at installation quietly gets worse. Over milliseconds, the same sensor also produces glitches, spikes and dropouts that a good model should simply ignore. Work on spiking neural networks, which are networks whose neurons communicate by discrete pulses emitted only when an internal voltage crosses a threshold, has shown that thresholding rejects fast noise well. Work on temporal distribution shift has shown something that pulls the other way: forcing a model to be invariant to everything that changes over time throws away useful signal and loses to plain retraining on pooled data. This proposal starts from the observation that the two literatures disagree mainly about timescale. Noise is fast and should be suppressed. Drift is slow and should be followed.
We propose a spiking classifier whose firing thresholds adapt on two separate clocks. A fast term suppresses within-sample perturbations relative to a reference firing pattern, as existing homeostatic spiking networks already do. A slow term moves that reference itself, without labels, using only confident predictions on recent deployment data. The claim we test is narrow. On a 36-month sensor drift stream, evaluated causally so that only past unlabeled data are used, two-timescale homeostasis should match or beat label-free re-normalization baselines on clean drift. It should clearly beat them once realistic transient glitches are mixed into the stream, because any single normalization window either chases the glitches or lags the drift. The experiments also include a matched non-spiking version, so we can say honestly whether the gain comes from the timescale principle or from spiking itself.
Core Claim
Threshold-adapting spiking neurons already carry a reference pattern of "normal" activity. In HoSNNs, that reference (the Neural Dynamic Signature) is recorded once on clean training data, and each neuron's threshold is pushed toward it. This gives bounded responses under adversarial noise, with CIFAR-10 accuracy under PGD attack rising from 0.54% to 43.33%. Under months-long drift, however, a frozen reference becomes stale. The network then treats the new normal as a perturbation and suppresses it, which is the spiking version of the "robustness paradox" documented for invariant learning on clinical data (Guo et al.). The opposite remedy is to re-normalize inputs to recent statistics, either by per-batch min-max scaling as in the bio-inspired gas SNN or by RevIN. That remedy follows drift, but it matches whatever the recent window contains, including glitches and changes in which classes happen to be present.
Our mechanism separates the two jobs. For hidden neuron $i$, the threshold is
$$\theta_i(t) = \theta_0 + \kappa\,\big(\bar a_i^{\text{fast}}(t) - r_i(s)\big),$$
where $\theta_0$ is the base threshold, $\kappa$ is a gain, $\bar a_i^{\text{fast}}(t)$ is the neuron's firing rate averaged over a few milliseconds within the current sample, and $r_i(s)$ is its reference rate at deployment sample index $s$. After each sample, the reference moves slowly:
$$r_i^{c}(s+1) = r_i^{c}(s) + \frac{g(u_s)}{\tau_{\text{slow}}}\big(\bar a_i(s) - r_i^{c}(s)\big).$$
Here $c$ is the class the network predicted for sample $s$, $\bar a_i(s)$ is the neuron's average rate on that sample, $\tau_{\text{slow}}$ is a horizon of hundreds to thousands of samples, and $g(u_s)\in[0,1]$ is a gate. The gate is computed from single-pass uncertainty $u_s$, estimated by applying dropout at each internal timestep as in AOT-SNN. During fast suppression, the network uses $r_i^{c}$ mixed by its current class probabilities.
Keeping one reference per predicted class is what separates this from re-normalizing marginal statistics. Following drift class by class should keep the gaps between classes intact, which protects discrimination. Matching pooled statistics can collapse those gaps when the mix of classes changes between time windows.
How This Differs From Closest Work
Four papers motivate the proposal by exposing the gap. Using context to make gas classifiers robust to sensor drift tracks drift with a recurrent context pathway, reaching 88.2% versus 86.9% for a plain feedforward network, but it needs labeled samples from past batches. Our update is label-free. The bio-inspired gas SNN reports 99.26% average accuracy over the 10 drift batches, but it relies on batch-level min-max normalization before spike encoding. It never tests whether the spiking network handles drift on its own. We treat that normalization as a baseline to be isolated, not as part of our method. Zero-shot temporal resolution adaptation recalibrates spiking decay factors in closed form when the sampling step changes, restoring MSWC accuracy from 38.8% to 93.6%. It addresses only changes in sampling rate, not drift in the input distribution, and its linear assumptions break at high firing rates. The RAT SNN neural decoder handles cross-session neural drift through supervised staged fine-tuning, not online tracking.
Two lines of work constrain the design. Wild-Time and the MIMIC-IV study by Guo et al. show that invariance penalties lose to pooled ERM, the standard practice of training on all available labeled data. This rules out any design that suppresses slowly changing features. The systematic review finds that recalibration repairs probability calibration but not ranking ability, so discrimination is the metric that matters here.
DRAIN forecasts future network weights from a trajectory of labeled domains. It is heavier than our method and it needs labels, so it serves as a contrast rather than a component. Two papers supply method components: HoSNN provides the fast threshold mechanism and AOT-SNN provides the uncertainty gate.
Research Program
We will build a compact spiking classifier with leaky integrate-and-fire neurons and the two-timescale thresholds above. Inputs are the 16-sensor feature vectors from the UCSD gas sensor drift benchmark (13,910 samples, 10 batches over 36 months), converted into spike trains by latency coding. Training uses only batch 1, or batches 1–2 in a secondary setting, with surrogate gradients. During training, the slow reference is held fixed at the class-conditional training rates.
At deployment, batches 2 through 10 are streamed in chronological order. No labels are available, and every normalization statistic, for every method, may use only past samples. This strictly causal protocol is itself a contribution. The published 99.26% figure uses batch-level normalization whose causal status the file does not specify, and the 88.2% figure uses past labels.
The analysis has three parts. The first sweeps $\tau_{\text{slow}}$ from "never update", which recovers HoSNN-style static anchoring, down to "update every sample". The single-timescale extremes should fail in opposite ways. The second asks Gap 1's open question directly: does plain thresholding without slow homeostasis discard drift-correlated but class-informative features? We test this by probing hidden spike counts with a linear readout refit on each batch's labels, a diagnostic only, not available to the deployed method. The third evaluates a forecasting variant on the Electricity and ETTh datasets with a TS-LIF backbone. There, slow homeostasis is compared with wrapping the same backbone in RevIN.
Baselines
The naive reference point is the frozen model trained on batch 1 with no adaptation, evaluated on each later batch. It measures how much drift actually hurts under our protocol.
The second baseline is the same spiking network with causal min-max normalization. We use the gas SNN's normalization recomputed over a sliding window of past samples, with the window length tuned on batch 2 only. This is the strongest label-free competitor, because it is exactly the ingredient the published 99.26% result depends on.
The third is a non-spiking multilayer perceptron with test-time adaptation of its normalization statistics, using the SAR codebase and the ttab harness. Both receive the same causal stream and the same hyperparameter budget, selected on batch 2.
The fourth is the labeled-context model of Warner et al., which reports 88.2% mean accuracy on its 5-gas protocol. We reimplement it on our split and treat it as a privileged-information reference, since it sees labels we do not. The fifth is a labeled oracle: pooled ERM retrained after each batch. It is the upper bound that Wild-Time and Guo et al. say is hard to beat.
Four ablations isolate the key ingredient: a single timescale with the best tuned $\tau$, marginal rather than class-conditional references, no uncertainty gate, and a non-spiking network that runs the identical two-timescale rule on batch-normalization statistics. The method beats the strongest label-free baseline if its mean accuracy and balanced accuracy over batches 3–10 exceed causal min-max normalization and SAR by a margin larger than the spread across five seeds, with the largest gap on the late batches. On the glitch-injected stream it should also lose markedly less accuracy than they do.
For forecasting, the reference number is RevIN on N-BEATS on ETTh2 at horizon 960, which reports MSE 0.471 versus 6.408 without RevIN (RevIN). We will compare like-for-like by putting RevIN on the TS-LIF backbone. No reported result was found on Lacuna for a spiking forecaster combined with RevIN.
Why not just…?
Why not just normalize each recent window, as the gas SNN and RevIN do? This is the real threat, and on clean drift it may tie us. The experiment that separates the two is glitch injection. We add short bursts of gain spikes and dropouts to random sensors, mixed into the chronological stream. A window short enough to follow 36-month drift is also short enough to be dragged around by bursts, while a long window lags the drift. If windowed normalization matches us even with glitches, the fast/slow separation adds nothing, and we will report that. A second check concerns class composition. If the gases measured differ across batches, marginal normalization should lower balanced accuracy compared with class-conditional references. We will measure this rather than assume it.
Why not just use test-time adaptation of batch-norm statistics on an ordinary network? Rewritten for a standard network, our rule is close to test-time adaptation with a slow, gated, class-conditional update. That is why we run that exact non-spiking ablation. If it matches the spiking version, the contribution is the timescale principle, not spiking, and we will say so. The question left for the spiking version is whether thresholding supplies the fast-noise rejection at lower energy, measured by spike-operation counts. We will not claim an energy advantage without measuring it.
Why not just retrain periodically on pooled labeled data? When labels arrive, this wins, as Guo et al. and Wild-Time show, and it is our upper bound. The problem is that labels in drift settings arrive late or never, as Feng et al. stress. Our question is how much of the oracle's gain a label-free method recovers between retraining cycles. We report that recovered fraction directly.
Decisive Checks
The first check is the $\tau_{\text{slow}}$ sweep on the glitch-injected stream. The idea holds if accuracy is maximized at an intermediate horizon and that optimum beats the best single-timescale normalization baseline. If performance is monotone in $\tau$, the separation adds nothing.
The second check is the linear-probe diagnostic. Without slow homeostasis, class information recoverable from hidden spikes should fall on later batches even when a refit readout is allowed. If it does not, thresholding is not discarding drift signal, and Gap 1's worry is unfounded for this data.
The third check is the gate. We will correlate AOT-style uncertainty with per-batch error. If uncertainty does not rise with drift, the gate is uninformative and should be dropped. If it does rise, it doubles as the label-free warning signal that clinical monitoring work asks for.
Throughout, results will be reported per batch with five seeds. Accuracy and balanced accuracy are the primary metrics, and expected calibration error is reported separately, since calibration and discrimination behave differently under drift.
References
- HoSNNs: Adversarially-Robust Homeostatic Spiking Neural Networks with Adaptive Firing Thresholds
- Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine
- A Bio-Inspired Spiking Neural Network with Few-Shot Class-Incremental Learning for Gas Recognition
- Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift
- Efficient Uncertainty Estimation in Spiking Neural Networks via MC-dropout
- Using context to make gas classifiers robust to sensor drift
- Zero-shot temporal resolution domain adaptation for spiking neural networks
- Pretraining with Re-parametrized Self-Attention: Unlocking Generalizationin SNN-Based Neural Decoding Across Time, Brains, and Tasks
- Wild-Time: A Benchmark of in-the-Wild Distribution Shift over Time
- Systematic Review of Approaches to Preserve Machine Learning Performance in the Presence of Temporal Dataset Shift in Clinical Medicine
- Temporal Domain Generalization with Drift-Aware Dynamic Neural Networks
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting
- Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare