In future wireless networks with huge numbers of IoT devices, most devices are silent in any given frame. Each active device is also heard by only a handful of nearby antennas, because signals fade quickly with distance. Receivers can exploit this "double sparsity," few users and few antennas per user, to jointly work out who is transmitting, what their channels look like, and what data they sent. They do this from pilot sequences much shorter than the number of devices. The best current receivers for this problem are iterative solvers "unfolded" into neural networks: each iteration of the solver becomes one layer, and the layer's step sizes and thresholds are learned from data. Yet every layer of these networks still does dense arithmetic over every (device, antenna) pair, including the roughly 98% of pairs that the model itself says are zero.
This note proposes a receiver whose compute cost scales with how much its estimate changes, not with the size of the network it serves. Each unfolded layer is rewritten so that it passes only the change in the channel estimate to the next layer. That change is encoded as signed, fixed-size events ("spikes"), the way neuromorphic chips communicate. Inactive users and distant antennas produce no change and therefore no events. As the estimate settles in later layers, the changes shrink and so does the event traffic. The claim is that this turns the physical double sparsity of the channel directly into the event sparsity that event-driven hardware rewards. The resulting energy per frame should grow roughly linearly with the number of active (user, antenna) links while matching the accuracy of the best dense unfolded receiver.
Core Claim
The joint problem is called JACD: joint active user detection, channel estimation, and data detection. The strongest existing solution for cell-free JACD is the deep-unfolded receiver of Deep-Unfolded Massive Grant-Free Transmission in Cell-Free Wireless Communication Systems. It unfolds a box-constrained forward-backward splitting solver, a method that alternates a gradient step with a thresholding step. The solver comes from Joint Active User Detection, Channel Estimation, and Data Detection for Massive Grant-Free Transmission in Cell-Free Systems, and unfolding cuts it from about 200 iterations to 10 learned layers (DU-ABC and DU-POEM). That work establishes that double sparsity makes the problem solvable. It executes the layers densely.
Write the received pilot block as $Y \approx S H$. Here $S$ is the $L \times N$ matrix of pilot sequences for $N$ potential users of length $L$. $H$ is the $N \times M$ channel matrix to $M$ access points; it is row-sparse because few users are active and sparse within each row because few access points are nearby. One unfolded gradient layer computes $H^{(k)} = \mathrm{prox}_k\big(H^{(k-1)} + \mu_k S^H(Y - S H^{(k-1)})\big)$, where $\mu_k$ is a learned step size and $\mathrm{prox}_k$ is a learned shrinkage that zeroes small entries. Precompute the correlation $B = S^H Y$ once per frame and the Gram matrix $G = S^H S$ once per pilot design. Then the argument of the shrinkage can be carried forward incrementally:
$$Z^{(k)} = Z^{(k-1)} + \Delta^{(k-1)} - \mu_k\, G\, \Delta^{(k-1)}, \qquad \Delta^{(k)} = H^{(k)} - H^{(k-1)}.$$
Here $Z^{(k)}$ is the pre-shrinkage state of layer $k$, and the recursion is exact when the step size $\mu$ is shared across layers. Allowing $\mu_k$ to vary requires one extra scalar rescaling of a running residual correlation, still driven only by $\Delta$.
The only matrix product in this recursion is $G\Delta$. Its cost is proportional to the number of nonzero entries of $\Delta$, not to $NM$. The proposal quantizes $\Delta^{(k)}$ into signed events of a learned size $q_k$. Each event then triggers an accumulate of one column of $G$ into $Z$, with no multiplication. The total work of the receiver is therefore approximately $N \sum_k \|\Delta^{(k)}\|_1 / q_k$. This quantity is small exactly when the problem is doubly sparse and the layers converge smoothly.
This design is not rate coding, which represents a value by how often a neuron fires over $T$ time steps and is the source of the latency problem noted for spiking regressors. It is a delta, or sigma-delta, code across unfolded layers. The final channel estimate is read out the way Spiking neural networks for frame-based and event-based single object localization reads continuous bounding-box values: it integrates membrane potential in non-firing accumulator neurons.
One contrast should be stated plainly because it constrains the energy story. With phase-shift-keyed pilots the entries of $S$ are already sign patterns, so a dense product $S\cdot$ is itself nearly multiplier-free. The multiply-to-accumulate saving reported in SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation is 0.9 pJ versus 4.6 pJ at 45 nm. It applies here mainly to the continuous-valued Gram matrix $G$ and to the data-detection and activity-detection stages. The main gain claimed is therefore a reduction in operation count driven by events. The per-operation saving is secondary.
Research Program
The build has three stages.
The first is an exact, non-learned delta-form forward-backward splitting solver. It establishes how the change magnitudes $\|\Delta^{(k)}\|_1$ decay across iterations on the cell-free simulator, and how that decay depends on activity probability and on the number of access points.
The second quantizes these deltas into events and trains the unfolded receiver end to end, using surrogate gradients to pass derivatives through the quantizer. The loss combines three terms. The first is the task loss of the DU receivers, which covers activity detection and data errors. The second is a channel NMSE term (normalized mean squared error of the channel estimate), added because Deep-Unfolded Massive Grant-Free Transmission in Cell-Free Wireless Communication Systems reports that task-only training badly degrades channel NMSE. The third is a weighted penalty on total event count. Sweeping that weight traces an accuracy-versus-events curve rather than a single operating point. The learned quantum $q_k$ lets early layers take coarse, cheap steps and later layers take fine ones. The event penalty encourages the network to converge monotonically rather than oscillate, which dense unfolding has no reason to do.
The third stage extends the same delta coding to the bilinear data-detection updates and to the soft activity detector. That detector already takes estimated channel and data energy as inputs, so it becomes an accumulator neuron over per-user event totals.
The analysis side asks one question. Does total event count track the number of active (user, access point) links, and how does it change when the activity rate exceeds the design value? That overload regime is where fixed-allocation grant-free designs fail, as Rethinking Grant-Free Protocol in mMTC argues. A receiver whose cost degrades gracefully with load would be a useful property in its own right. Energy is reported in two ways. The first is operation counts converted with the 45 nm per-operation figures above. The second is hardware measurement on an event-driven chip if access is available. Measured results must include static power, since Energy-efficient neuromorphic classifiers found leakage to be about 95% of TrueNorth consumption when event rates were low. A sparse workload may therefore save much less than operation counts suggest.
Prior spiking work in wireless constrains how much to expect. Exploring the Potential of Spiking Neural Networks in UWB Channel Estimation uses untrained liquid reservoirs and an unsupervised self-organizing map to classify a single link as line-of-sight or not, at 80.2% accuracy on eWINE. It does no multi-user detection, estimates no complex channel matrix, and was not run on hardware. This proposal regresses the full $N \times M$ channel and decodes data. It also borrows the solver's structure rather than a random reservoir, which is what makes the regression feasible.
Baselines
The naive reference is dense box-constrained forward-backward splitting run for 200 iterations, as in Joint Active User Detection, Channel Estimation, and Data Detection for Massive Grant-Free Transmission in Cell-Free Systems. It is the accuracy ceiling for channel NMSE and the operation-count ceiling. It is run as published on identical frames.
The strongest existing method is DU-ABC and DU-POEM at 10 layers from Deep-Unfolded Massive Grant-Free Transmission in Cell-Free Wireless Communication Systems. That paper reports the lowest activity detection error and symbol error rate among all compared methods at 10 iterations. No reported result was found on Lacuna for its exact scalar NMSE or symbol error rate values, because the file notes they appear only in plots. It will therefore be retrained with the same training frames, layer count, optimizer and hyperparameter search budget as the proposal. Two versions will be trained: one with its original task-only loss and one with the added NMSE term.
The key ablation is a zero-skipping DU receiver. It is the same dense unfolded network, but sparse kernels skip products involving zero entries of $H^{(k)}$, using ordinary full-precision multiply-accumulates. It isolates whether any gain beyond exploiting the sparsity of the state comes from delta coding over layers. A second ablation is the delta-form receiver with unquantized floating-point deltas. It isolates the effect of event quantization and the learned quanta. A third reference is a rate-coded spiking conversion of DU-ABC with $T$ time steps. It shows what a naive "make it spike" approach costs.
Beating the strongest baseline means two things together. Activity detection error and symbol error rate must stay within 0.5 dB SNR of retrained DU-ABC/DU-POEM, with channel NMSE no worse. Operation count per frame must be at least three times lower than the zero-skipping DU receiver. These targets are our own success criteria, not reported numbers.
Why not just…?
Why not just run the dense unfolded receiver with sparse kernels that skip zeros? This is the most serious threat. It captures the first-order benefit, since inactive users contribute nothing, with no new training. It falls short in one specific way. Zero-skipping still recomputes the full correlation over all $N$ users against the residual at every layer. The residual is dense even when the estimate is sparse, so that per-layer cost does not shrink as the estimate converges. The delta form removes this term, because $G\Delta$ vanishes wherever the estimate has stopped changing. The zero-skipping ablation settles the question directly. If the event-driven receiver is not substantially cheaper at equal accuracy, the honest conclusion is that sparse kernels suffice and spikes add nothing here.
Why not just quantize the dense DU network to low precision, or rely on sign-valued pilots? Quantization cuts energy per operation but not the number of operations, which stays at order $NML$ per layer. As noted above, sign-valued pilots already make the $S$ products cheap, so a quantized dense network would capture some of what spikes offer on those products. The proposal does not claim the multiply-to-accumulate saving as its main effect. The comparison is run in operation counts, where quantization and delta coding differ, and the two can also be combined.
Why not just shorten the problem at the protocol level, as in the two-stage count-then-detect scheme? Rethinking Grant-Free Protocol in mMTC reports a "3x speedup in convergence time compared to standard methods without sacrificing accuracy." It achieves this by estimating the active count from a 4-symbol common preamble and restricting coordinate descent to the top candidates. It assumes perfect power control and synchronization and does not estimate channels jointly with data. It is also complementary rather than competing: a count estimate could initialize the event receiver. The overload experiment tests whether the event receiver achieves load-proportional cost without that protocol change.
Why not just convert the trained DU network to a standard rate-coded spiking network? Rate coding needs many time steps to represent the precise continuous channel values that regression requires. The latency and dynamic-range limits of rate-coded SNNs are the reason. It also spends events re-encoding values that have not changed. The rate-coded baseline measures this cost directly against delta coding.
Proposed Experiments
All experiments use the cell-free JACD simulator from Deep-Unfolded Massive Grant-Free Transmission in Cell-Free Wireless Communication Systems. Its setting is a 500 m by 500 m area, $N = 400$ potential users, 20 to 100 access points with 60 as the main setting, 20% activity, QPSK, and three-slope path loss.
Metrics are activity detection error, symbol error rate, channel NMSE, events and operations per frame, and estimated energy per frame. Energy is computed from the 0.9/4.6 pJ figures and reported separately from memory access, which those figures do not cover.
Three checks decide the claim. First, events per frame should scale approximately linearly with the true number of active (user, access point) links, across activity rates from 10% to 40% and across 20 to 100 access points. A dense or zero-skipping receiver should instead scale with $N$ or $M$. Second, at matched accuracy, the event receiver must beat zero-skipping DU by the stated margin. Third, adding the NMSE term must bring channel NMSE close to the 200-iteration solver without losing the activity-detection and symbol-error gains of task-trained unfolding.
A secondary test on the mmWave activity detection setting of Device Activity Detection and Channel Estimation for Millimeter-Wave Massive MIMO ($N = 2000$, $M$ = 32 to 64) checks whether angular-domain sparsity yields the same event savings. That setting has off-grid leakage, which spreads energy across neighbouring angle bins and should raise event counts. This is a predicted failure mode worth measuring rather than hiding. If hardware is available, measured energy with and without static power, compared against a dense run on an embedded GPU, would show whether operation-count savings survive leakage.
References
- Deep-Unfolded Massive Grant-Free Transmission in Cell-Free Wireless Communication Systems
- Joint Active User Detection, Channel Estimation, and Data Detection for Massive Grant-Free Transmission in Cell-Free Systems
- Spiking neural networks for frame-based and event-based single object localization
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- Rethinking Grant-Free Protocol in mMTC
- Energy-Efficient Neuromorphic Classifiers
- Exploring the Potential of Spiking Neural Networks in UWB Channel Estimation
- Device Activity Detection and Channel Estimation for Millimeter-Wave Massive MIMO