Spiking neural networks (SNNs) are networks whose neurons communicate with binary pulses, or "spikes," instead of real-valued activations. They save energy because a spike replaces each multiply with an addition, and a silent neuron costs almost nothing. Recent spiking 3D classifiers report accuracy close to conventional networks at a small fraction of the estimated energy. One example is a spike-driven transformer reaching 85.5% on real scanned objects at about 1.7 mJ per inference (SpikeNet). That energy figure, however, is measured on clean or fixed test sets. Real 3D sensors corrupt data in ways that act directly on how often spikes occur. Occlusion and beam dropout remove points, so neurons go silent. Background clutter and reflections add floating points, so neurons fire too much. A spiking network's accuracy and its energy cost therefore both drift when it moves from synthetic training data to a real sensor, and nobody has measured either drift on standard 3D corruption benchmarks.
We propose to measure this drift and correct it with a single mechanism. Each spiking layer keeps the per-channel firing rates it showed during training. At test time it nudges its own firing thresholds toward those reference rates, using only forward-pass spike counts. The update is normalized by how many points are actually present, so that a sparser scan is not mistaken for a shifted one. It needs no backpropagation and no retraining, and it works on existing checkpoints. It runs both across a stream of inputs and across the simulation time steps of a single input. The goal is a spiking 3D classifier that keeps its accuracy under real-world corruption and stays within a fixed energy budget, with no gradient computation on the device.
Core Claim
Sensor corruption on point clouds shows up inside a spiking network as a predictable, corruption-specific shift in per-channel firing rates. Density corruptions push rates down and noise corruptions push them up. Restoring the training-time rates through feed-forward threshold adjustment recovers a large share of the lost accuracy and also prevents energy blow-up under clutter. The new ingredients relative to prior SNN test-time adaptation are three. The firing-rate statistics are normalized by point count. The mechanism retrofits to 3D spiking architectures without special pretraining. Energy is treated as a quantity that corruption changes and that must be reported per corruption.
Relation to Prior Work
Two papers motivate the mechanism and also constrain it. Threshold Modulation for Online Test-Time Adaptation of Spiking Neural Networks identified "firing rate shift" under 2D image corruption. Its backprop-free variant, TM-NORM, cuts CIFAR-10-C error on a ResNet-20 from 46.7% to 25.2% at about 3% energy overhead. TM-NORM folds normalization statistics into the threshold, but it requires pretraining with membrane-potential batch normalization and was only tested on images. Images keep a fixed number of pixels no matter the corruption. Point clouds do not: an occluded scan feeds fewer active inputs into every pooled or neighborhood-aggregated layer. Image-style statistics would read this reduced density as a feature shift.
SPACE adapts SNNs from a single instance by making spike counts consistent across augmented views. It reaches 71.03% on CIFAR-10-C at severity 5, against 66.57% without adaptation, but it updates weights by backpropagation. Inference-only neuromorphic chips with fixed weights cannot run that update. Our method borrows SPACE's observation that time-aggregated spike counts are a stable signal. We use that signal to set thresholds rather than to compute gradients.
The 3D spiking backbones supply the testbed: 3DSMT, SpikeNet, Spiking Point Transformer, and Spiking Discrepancy Transformer. All four report ScanObjectNN accuracy and theoretical energy. None of them reports results on ModelNet40-C, which defines 15 corruptions in density, noise, and transformation families. None reports results on ScanObjectNN-C from AdaptPoint either. The Spiking Discrepancy Transformer deserves particular attention. It deliberately amplifies differences between a point and its neighbors to sharpen edges, and impulse and background-noise corruptions are exactly high-difference outliers. Its robustness to those corruptions is therefore an open question that our diagnosis answers directly.
Noise-Injected Spiking Graph Convolution injects noise into membrane potentials during training and evaluates only on synthetic Gaussian denoising. We include its noise injection as a training-time comparison, but it is not the core of this proposal. The energy model follows the accumulate-only argument of Wu et al.: about 0.9 pJ per addition against 4.6 pJ per multiply-accumulate in 45nm. As a consequence, energy scales with spike count, which is why corruption-induced rate changes translate directly into energy changes.
Research Program
The first stage is diagnosis. We take trained SpikeNet and Spiking Discrepancy Transformer models, plus Spiking Point Transformer for a backbone that uses multiple time steps. We run them on all 15 ModelNet40-C corruptions at 5 severities and on ScanObjectNN-C. For every layer and channel we log firing rates, and from those we compute total synaptic operations, which give energy at 0.9 pJ per accumulate. The test is whether rate shift sorts by corruption family and whether its size predicts the accuracy drop. We expect density corruptions to lower rates and noise and background corruptions to raise them. This stage alone fills a gap: no reported result on Lacuna exists for any spiking 3D model on ModelNet40-C or ScanObjectNN-C.
The second stage is the correction mechanism. During training we store a reference rate $\bar r_{l,c}$ for each layer $l$ and channel $c$. This is the average fraction of active point-slots that fire, so the denominator is the number of points present rather than a fixed tensor size. At test time we measure the same point-normalized rate $r_{l,c}$ and update the threshold multiplicatively:
$$V^{th}_{l,c} \leftarrow V^{th}_{l,c}\,\Big(\frac{r_{l,c}}{\bar r_{l,c}}\Big)^{\gamma},$$
where $\gamma \in (0,1]$ is a damping exponent and the threshold is clipped to a fixed band around its trained value. The update runs in one of two modes. The streaming mode uses an exponential moving average of $r_{l,c}$ over successive inputs, analogous to TM-NORM but needing no batch-norm pretraining. The within-instance mode updates thresholds between simulation time steps of a single input, using the rate measured up to step $t-1$. Adaptive thresholds are a native neuron feature on chips such as Loihi, so the update requires no gradient circuitry.
For layers that aggregate over k-nearest-neighbor groups, such as local offset attention and discrepancy attention, we compute rates per local group. A cluster of background outliers can then be damped without silencing the real surface elsewhere. We also add an optional energy cap that only raises thresholds. It bounds spike count at the training-time level, which tests whether energy can be held flat under clutter at an acceptable accuracy cost.
The third stage tests whether the correction composes with training-time robustness. We fine-tune the same spiking backbones with corruption augmentation and ask whether homeostasis still helps on top. The decisive version of this test holds out corruption families the augmentation never saw.
Baselines
The naive reference point is each spiking backbone evaluated with no adaptation. It gives the accuracy floor and, just as important, the uncontrolled energy under each corruption. No reported result on Lacuna exists for this on ModelNet40-C or ScanObjectNN-C, so we produce it ourselves.
The most direct competitor on the mechanism is TM-NORM from Threshold Modulation, ported to the same 3D backbones. The ported models are retrained with membrane-potential batch normalization as that method requires, on the same data, schedule, and tuning budget as our models. Its reported numbers are 2D only (25.2% CIFAR-10-C error with ResNet-20). No reported result on Lacuna exists for TM-NORM on point clouds. We count our method as beating it if it achieves lower mean corruption error on ModelNet40-C density and noise corruptions with equal or lower spike count, and if it does so without the batch-norm pretraining. Matching TM-NORM while needing no retraining would be a weaker but still useful result.
As a gradient-based upper reference among SNN adaptation methods, we adapt SPACE to point clouds. Its augmented views are generated with ModelNet40-C-style jitter and dropout, and we give it the same number of views and steps used in its paper's spirit. Our method should approach its accuracy. Its compute and its need for backward passes make it unusable on inference-only chips.
The strongest existing robustness numbers come from conventional ANNs. PointCutMix-R training combined with TENT test-time adaptation lowers average ModelNet40-C error from 26.1% to 15.4% (Sun et al.). AdaptPoint reports mean corruption error falling from 107.2% to 63.7% on ModelNet-C with PointNet++, and 78.3% on ScanObjectNN-C (AdaptPoint). We rerun PointCutMix-R with TENT on a PointNet++ or PointMLP backbone (pointMLP-pytorch) under our protocol, and we report AdaptPoint at its published numbers. We do not expect a spiking model to beat these on accuracy. Our success criterion is a Pareto claim: comparable error at an order of magnitude less estimated energy, with no backward pass at deployment.
Three ablations isolate the key ingredient. The first replaces point-normalized rates with raw tensor-normalized rates, which tests whether density awareness matters. The second replaces per-group rates with global per-channel rates. The third adjusts thresholds in one direction only, either lowering them only or raising them only.
Why not just…?
Why not just train the spiking network with corruption augmentation and skip test-time adaptation? This is the most serious threat. PointCutMix-R and AdaptPoint show that augmentation covers much of the gap on ANNs. If an augmented SNN matches homeostasis on unseen corruption families, the mechanism adds little. Stage three tests this directly: we augment with only some corruption families, evaluate on the held-out families, and measure the additional gain from homeostasis on top of augmentation. If the gain on held-out families falls within seed noise, we report that outcome. The contribution would then reduce to the diagnosis and the energy-under-corruption measurements.
Why not just filter the point cloud first, with outlier removal and resampling to a fixed count? A non-learned statistical outlier filter followed by resampling is cheap, and it probably handles background and impulse noise well. It cannot restore geometry removed by occlusion, and it adds a dense preprocessing stage that runs outside the spiking pipeline. We include it as a baseline both on its own and combined with our method. We expect it to be competitive on noise corruptions and weak on density corruptions, and that per-family split will be reported.
Why not just use a small conventional network like SimpleView, or an ANN with TENT? SimpleView reaches 79.5% on ScanObjectNN with 0.8M parameters. It is a fair accuracy competitor, but it runs dense multiply-accumulates and cannot be event-driven. TENT needs backward passes. The question this proposal asks is what accuracy is achievable under the constraints of neuromorphic deployment, so the ANNs appear on the accuracy-energy plot as references rather than as alternatives to the same goal. Our energy figures are theoretical, as are those of all cited spiking papers. We state that limitation openly instead of claiming chip-measured savings.
Proposed Experiments
All models train on ModelNet40 and are evaluated on ModelNet40-C (15 corruptions, 5 severities). Separately, models train on ScanObjectNN and are evaluated on ScanObjectNN-C. The metrics are mean corruption error, per-family error, total synaptic operations and estimated energy per corruption, and the ratio of energy under corruption to clean energy. Streaming experiments use batch size 1 with randomly ordered and mixed corruptions, so they reflect deployment.
The first decisive check is diagnostic. Per-corruption rate shift, measured as the log ratio of $r$ to $\bar r$ averaged over channels, should correlate with accuracy loss across the 75 corruption-severity cells. Its sign should split density corruptions from noise corruptions. If rates do not shift systematically, the premise fails and we stop at the benchmark contribution.
The second check is the main result. Homeostasis should reduce mean corruption error relative to no adaptation and relative to ported TM-NORM, while keeping energy under clutter within roughly the clean budget.
The third check is the held-out-family augmentation experiment described in the previous section.
The fourth check concerns the Spiking Discrepancy Transformer. It should degrade more than SpikeNet on impulse and background corruptions relative to its clean accuracy, and per-group homeostasis should close more of that gap than global homeostasis. Confirming or refuting this resolves the tension between amplifying neighbor differences and robustness to outliers.
References
- SpikeNet: Sparse Spike-Driven Mask Vector Transformer for Energy-Efficient and Stable Spiking Point Cloud Processing
- Threshold Modulation for Online Test-Time Adaptation of Spiking Neural Networks
- SPACE: SPike-Aware Consistency Enhancement for Test-Time Adaptation in Spiking Neural Networks
- 3DSMT: A Hybrid Spiking Mamba-Transformer for Point Cloud Analysis
- Spiking Point Transformer for Point Cloud Classification
- Spiking Discrepancy Transformer for Point Cloud Analysis
- Benchmarking Robustness of 3D Point Cloud Recognition Against Common Corruptions
- Sample-adaptive Augmentation for Point Cloud Recognition Against Real-world Corruptions
- Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising
- Deep Spiking Neural Networks for Large Vocabulary Automatic Speech Recognition
- Revisiting Point Cloud Shape Classification with a Simple and Effective Baseline