Radio receivers can double as radars. A base station that sends ordinary multi-antenna data signals can process the echoes that bounce back to map which parts of a room are occupied, and the data link itself is untouched. This is integrated sensing and communication (ISAC). The best-known neural version of this idea has two costs that limit deployment. It runs a full convolutional network on every radio frame, even though an indoor scene is almost entirely static from one frame to the next. It also needs an offline scan of the empty room so that walls and furniture can be subtracted out, and that scan becomes invalid as soon as a chair moves.
We propose that both costs come from the same missed opportunity: the receiver computes on the channel, not on changes in the channel. A spiking neural network is a network whose units emit discrete events only when their internal state crosses a threshold, and which stays silent otherwise. If such a network has slowly adapting thresholds, it learns to ignore whatever has been constant for a while. That one mechanism removes static clutter without calibration and also produces the event sparsity that cuts energy. We will build this receiver and measure it honestly. The main comparison is not against a dense network but against the obvious non-spiking version of the same idea: a quantized convolutional network that also subtracts a running background and skips unchanged frames. Whether spiking adds anything beyond "compute only on change" is the central empirical question.
Core Claim and Relation to Prior Work
The closest existing work is Neuromorphic Integrated Sensing and Communications. It pairs an impulse-radio pulse-position transmitter with a 6–10-neuron spiking receiver that decodes one bit and detects whether a target is present, trained with a weighted sum of communication and sensing losses. That paper motivates our direction: event-driven computation suits sparse radio signals. It does not tell us whether the idea survives the dense, complex-valued multi-carrier channels used by Wi-Fi and 5G/6G, or tasks harder than binary detection. Our setting is different. We use multi-antenna orthogonal frequency-division multiplexing (MIMO-OFDM) downlinks and dense occupancy maps, and in our case sparsity is not present in the input. It has to be created by temporal adaptation.
Our method and task come from Neural Integrated Sensing and Communication for the MIMO-OFDM Downlink. That work reconstructs 2D occupancy grids from downlink channel estimates in a ray-traced room. It turns them into beamspace-delay features (energy indexed by beam direction and echo delay), applies "subtraction fusion" against an empty-room scan, and runs a modified ResNet-18. It reports less than 5% accuracy loss at target speeds up to 30 m/s, and large gains in precision and recall from 20 to 320 MHz bandwidth. Its reliance on a static calibration scan is exactly the open problem we target.
Our claim is specific. Suppose a spiking network takes beamspace-delay frames as input, uses learned per-bin adaptation time constants, and uses 4-bit synaptic weights. We expect it to match the calibrated dense ResNet-18 on occupancy accuracy in static rooms, to recover automatically after the room's clutter changes, and to use substantially fewer operations than any per-frame dense network.
There is one structural reason to expect radio to suit spiking networks better than images. Image spiking networks repeat each input over $T$ simulated time steps; Going Deeper With Directly-Trained Larger Spiking Neural Networks needs 6 steps to reach 67.05% ImageNet top-1. That repetition multiplies latency and training memory. In our receiver, each simulated time step is a real radio frame, so this overhead does not exist.
Research Program
The receiver keeps the front end of the neural ISAC pipeline: regularized least-squares channel estimation followed by beamspace-delay magnitudes. Each beamspace-delay bin feeds an adaptive leaky integrate-and-fire input unit. Such a unit accumulates input into a membrane potential that leaks over time and emits a spike when the potential crosses a threshold. In our version, the threshold itself rises after recent activity, with a time constant that is learned for each bin. Bins dominated by walls or furniture therefore fall silent within a few frames, and bins where something moves keep firing. A spiking convolutional backbone sits above this front end. It is directly trained with surrogate gradients and threshold-dependent batch normalization, following Going Deeper With Directly-Trained Larger Spiking Neural Networks. A readout layer integrates spikes into an occupancy map.
Spiking networks are known to struggle with precise continuous regression. Occupancy mapping is a per-cell classification problem, which avoids that weakness. We will report sub-cell localization error separately so this limitation is visible rather than hidden.
The cost-reduction direction enters in two places. First, synaptic weights are quantized to 4 bits using the clipping, per-channel bit allocation, and bias correction of Post training 4-bit quantization of convolutional networks for rapid-deployment. Because spiking weights have been reported to be fragile under perturbation, we will compare that post-training recipe against quantization-aware fine-tuning. Second, energy is accounted for at the level of operations. A spiking layer pays one accumulate per incoming spike per synapse, while a dense layer pays one multiply-accumulate per connection per frame. Both are costed with the same per-operation energy constants at the same bit width, plus a count of weight-memory reads.
The communication link is not modified, so communication performance is unchanged by construction. ISAC here defines the operating regime: sensing must come for free from data-bearing frames.
Data come from the NVIDIA Sionna ray-tracing setup used by the neural ISAC paper: a 5 m × 5 m × 3 m concrete room, an 8×8 MIMO transceiver, and two single-antenna user equipments. We extend it from single snapshots to frame sequences. The sequences cover moving people at speeds up to 30 m/s, idle periods, and clutter-change events: furniture relocated mid-sequence, doors opened, and objects that are added and then left in place. Clutter-change events are the regime where an empty-room scan goes stale. We will sweep bandwidth from 20 to 320 MHz to match the original study.
Baselines
The naive reference is the published pipeline as reported: an FP32 ResNet-18 with empty-room subtraction fusion from Neural Integrated Sensing and Communication for the MIMO-OFDM Downlink, run on every frame. We pair it with the same network without subtraction, to show how much calibration is worth. That paper gives no absolute occupancy accuracy that we could find. Its reported results are relative: under a 5% drop at 30 m/s, and "consistently high accuracy" across map resolutions. We therefore retrain it on our sequences. The calibrated version will serve as the accuracy target, and its stale-calibration accuracy after a clutter change will serve as the failure we aim to fix.
The strongest competitor is a change-driven dense network built from parts already in the literature. It replaces subtraction fusion with a running exponential-average background. The averaging time constants are learned per bin, exactly as in our spiking front end, so the two arms share the clutter-removal mechanism. Its ResNet-18 is quantized to 4 bits with the method of Banner et al. It also skips computation on frames whose background-subtracted input falls below a learned change threshold, and computes the remaining frames only on changed regions.
The best transfer evidence for the quantization step is Banner et al.'s ImageNet ResNet-101 result: 75.0% top-1 at 4-bit weights and activations, compared with 77.2% FP32 and 62.6% for naive 4-bit. No reported result was found on Lacuna for 4-bit quantization on RF occupancy mapping.
Every arm gets the same training sequences, the same hyperparameter-search budget measured in trials and GPU-hours, and the same energy model. The spiking receiver beats this baseline only if two conditions hold. Its occupancy F1 must be within one point of the calibrated FP32 ResNet-18. At that matched accuracy, its estimated energy per second of operation must be at least twice lower than the change-driven dense network's, averaged over realistic activity profiles.
A classical reference covers the localization subtask: MUSIC angle estimation followed by extended-Kalman tracking. This follows the sensing-assisted tracking in Sensing-assisted Accurate and Fast Beam Management for Cellular-connected mmWave UAV Network, which reports 1.1 ms echo processing on a real testbed, and the angle-prior Kalman filtering of Sensing-aided Uplink Channel Estimation for Joint Communication and Sensing. We score this tracker on person positions, not maps.
Three ablations isolate the key ingredient. The first replaces adaptation with empty-room subtraction while keeping spiking. The second makes the adaptive thresholds fixed instead of learned. The third uses FP32 instead of 4-bit weights.
Why not just…?
Why not just use a Kalman or MUSIC tracker? These are cheap and near-optimal when the scene matches their model. The angle-prior Kalman filter runs in linear time and lands within 0.2 dB of MMSE channel estimation. However, they estimate a handful of point targets. They do not produce occupancy maps of extended or unmodeled objects, and they need a clutter model. Our person-localization comparison will show where the classical tracker is simply better. If it matches the spiking receiver on localization at lower cost, we will say so and limit our claim to dense mapping.
Why not just quantize ResNet-18 to 4 bits? Banner et al. show 4-bit models stay within a few points of full precision on images. A 4-bit dense network still pays multiplies on every connection of every frame, and it still needs calibration for clutter. The per-frame quantized arm, with the same learned background but no skipping, measures how much of the gain quantization alone delivers.
Why not just add a learned background average and frame skipping to a dense network? This is the real threat. Much of a spiking network's advantage on near-static scenes may reduce to "don't compute when nothing changes," and that is why it is our main baseline. The spiking receiver could still win in two ways. Its sparsity is per-unit and graded within every layer, not all-or-nothing per frame or region. And it needs no hand-set skip threshold. If the energy gap at matched accuracy is under 2×, we will report that change-driven computation, not spiking, is what matters. That would still be a useful and publishable answer.
Why not just use Learngene-style transfer to make dense receivers cheap? Efficient Deployment of Deep MIMO Detection Using Learngene cuts training and adaptation cost across 55 channel tasks. It does not reduce per-frame inference cost, which dominates for always-on sensing. It is complementary, and we leave it to future work.
Decisive Checks
We will run three checks. The first plots occupancy F1 over time across clutter-change events. Calibration-free arms should return to steady accuracy within a bounded number of frames, while the stale-calibration ResNet-18 should not. The second plots energy per second against scene activity, from empty to fully busy. It will show whether the spiking advantage is large only when the room is quiet, and where it crosses the change-driven dense baseline. The third is the matched-accuracy energy ratio between the spiking and change-driven dense arms, swept across bandwidth and quantization level.
Prior energy claims from vision and neuromorphic chips, such as the 147–154× of A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection, were measured against per-frame dense baselines on other tasks. Our ratio is measured against a change-driven dense baseline on this task, and it is the number this proposal exists to produce.
References
- Neuromorphic Integrated Sensing and Communications
- Neural Integrated Sensing and Communication for the MIMO-OFDM Downlink
- Going Deeper With Directly-Trained Larger Spiking Neural Networks
- Post training 4-bit quantization of convolutional networks for rapid-deployment
- Sensing-assisted accurate and fast beam management for cellular-connected mmWave UAV network
- Sensing-Aided Uplink Channel Estimation for Joint Communication and Sensing
- Efficient Deployment of Deep MIMO Detection Using Learngene
- A fully spiking hybrid neural network for energy-efficient object detection