Event-Gated Virtual Pilots: A Spiking Receiver for Joint Channel Estimation and Detection

Wireless receivers usually spend a slice of every transmission on pilots, known symbols used only to measure the radio channel, and then decode the payload with that fixed measurement. Data-aided receivers do better. They iterate: a rough channel estimate yields tentative data decisions, and those decisions are fed back as extra "virtual pilots" that sharpen the channel estimate. This loop recovers several decibels and lets the receiver send fewer pilots. The cost is that every iteration repeats dense floating-point matrix work over every antenna and every symbol, and that is too expensive for massive-antenna or high-bandwidth radios.

This proposal starts from an observation about that loop. After the first few iterations, almost all tentative symbol decisions stop changing. Only symbols near a decision boundary keep flipping. The iterative receiver is therefore naturally sparse in time, which is exactly the regime where spiking neural networks save energy. A spiking neural network is a network whose neurons communicate with binary events and consume dynamic energy only when an event occurs. We propose to build the data-aided loop so that a symbol emits an event only when its decision changes or becomes confident enough to serve as a virtual pilot. The channel is then updated only from those events.

Core Claim

Our claim is that a joint channel estimation and data detection receiver (JED) can be unfolded into a spiking network whose work scales with the number of decision changes per iteration rather than with array size times block length. The claim rests on two structural facts. First, data symbols come from a small discrete alphabet: QPSK points are ±1±j, and 16-QAM points take values in {±1, ±3} on each axis. Multiplying by a symbol decision is therefore a sign flip, a shift, and an addition. The channel side of the loop can then run on accumulate-only arithmetic, the multiplier-free operation that Wu et al. and Sengupta et al. identify as the source of neuromorphic energy savings. Second, the residual can be maintained incrementally. Let $Y$ be the received block, $\hat H_t$ the channel estimate at iteration $t$, $\hat X_t$ the current symbol decisions, $R_t = Y - \hat H_t \hat X_t$ the residual, and $\Delta X_t = \hat X_{t+1} - \hat X_t$ and $\Delta H_t = \hat H_{t+1} - \hat H_t$ the changes between iterations. Then

$$R_{t+1} = R_t - \hat H_{t+1}\,\Delta X_t - \Delta H_t\,\hat X_t .$$

Three consequences follow. The first correction term costs work proportional to the number of nonzero entries in $\Delta X_t$, and those entries are small constellation differences, so the term needs no multiplications. The second term is again a product with constellation points. The only genuinely multiplicative step is the data-side gradient $\hat H^H R$. We make it event-driven by sending changes in $R$ through send-on-delta neurons: a neuron fires when its accumulated change crosses a threshold, and changes below the threshold are held back rather than transmitted.

A per-layer confidence threshold decides which symbols may act as virtual pilots. It is learned end to end with surrogate gradients, which replace the non-differentiable spike threshold with a smooth approximation during backpropagation (Zheng et al.). Only symbols whose soft value crosses the threshold fire into the channel update. The same gate therefore does two jobs: it saves energy, and it guards against the known failure mode in which wrong low-SNR decisions become corrupted virtual pilots and push the estimate to a bad solution.

Relation to Prior Work

The method is built on S-JED. S-JED unrolls forward-backward splitting (a gradient-plus-projection solver for the joint channel and data problem) into about ten learned layers. Each layer produces soft outputs for an LDPC decoder, and a hyper-network predicts per-layer step sizes. We keep its unrolled structure and its soft-output estimator. We replace its dense floating-point iterations with event-gated incremental updates, and its hyper-network step sizes with learned firing thresholds. DPST unrolls complex-valued gradient detection and reaches near-maximum-likelihood BER on 8×4 4-QAM, but it needs 100 layers and does not estimate the channel. It informs how we handle complex arithmetic natively, as real and imaginary parts processed with shared weights. Classical message-passing JED, such as Schniter and PBiGAMP, motivates the bilinear formulation. It also shows that the loop converges in few iterations, which supports our bet that decisions stabilize early.

On the neuromorphic side, the closest work is Song, Simeone and Rajendran. They run a spiking transformer that infers the detection mapping from pilots presented in its input sequence, and they report 5.4x–26.8x lower system power. The design is purely feedforward from pilots and never refines a channel estimate from decoded data. Their finding that small spiking models hit BER floors from the randomness of their binary input encoding constrains our design. It is the main reason we use deterministic send-on-delta events rather than rate or Bernoulli encoding. Zhang et al. call their task spiking "channel estimation", but it is binary LOS/NLOS classification on eWINE and recovers no channel coefficients. NeuroJSCC studies spike-based transmission over simplified channels without multi-antenna channel state information. Neither addresses the problem posed here, which is continuous complex channel estimation inside a feedback loop.

Research Program

We will build the receiver in three stages.

Stage 1: floating-point event-gated S-JED. We first implement an event-gated version of S-JED in ordinary floating point. Symbols update only when their soft decision crosses the learned threshold, and the residual is maintained through the incremental equation above. This version measures, before any spiking hardware question arises, how sparse $\Delta X_t$ actually is across iterations, SNR values, and modulation orders. Sparsity is the load-bearing assumption. If more than roughly a third of symbols still change after iteration three at operating SNR, the idea loses most of its value.

Stage 2: spiking port. We then convert the gated loop into a spiking network. Symbol decisions and confidence gates become integrate-and-fire neurons. Residual changes are carried by send-on-delta neurons. Constellation products are compiled to shift-and-accumulate operations. Training uses backpropagation through the unrolled layers with surrogate gradients.

Stage 3: training against energy. Finally, we train with a loss that adds a penalty on total event count to the soft-output decoding loss. Because dynamic energy scales with spike count (Tavanaei et al., Jang et al.), sweeping the penalty weight traces an accuracy–energy frontier.

Energy is estimated with the 45nm/28nm CMOS operation-energy model used by Song et al. (2024), counting floating-point multiply-accumulates, accumulate-only operations, and memory accesses separately. If Loihi access is obtained, we will add measurements for a small configuration. The file reports Loihi and TrueNorth gains of two to three orders of magnitude on unrelated tasks (Tang, Martí et al.). We will not assume those gains transfer to this workload.

Baselines

Naive reference. Pilot-only linear MMSE channel estimation followed by linear MMSE detection, with the standard pilot budget. It anchors both the accuracy gap that data-aided processing should close and the cost floor any iterative method must justify.

Strongest accuracy competitor. Floating-point S-JED, reimplemented on the same 8×4 16-QAM Rayleigh setup with IEEE 802.11n LDPC codes. Song et al. report that it outperforms L-MMSE and max-log STS sphere decoding by 2 dB to 4 dB in coded PER and uncoded BER, and that it converges within 10 iterations. No absolute PER-versus-SNR operating points or energy figures appear in the file, so we will regenerate the curves ourselves. Fairness conditions:

  • same channel draws;
  • same pilot count;
  • same number of unrolled layers (10);
  • same training budget in gradient steps and hyperparameter trials.

Strongest efficiency competitor. S-JED quantized to INT8 and INT4 with quantization-aware training under the same budget. Its energy is costed with the same CMOS model, charging integer multiply-accumulates.

Neuromorphic pipeline competitor. A decoupled design in the spirit of Song et al. (2024): a pilot LMMSE channel estimate feeding a spiking detector. We realize it as our spiking network with the channel-update path disabled, so that the only difference is the data-aided loop.

Ablations. Two ablations isolate the key ingredients. The first is a rate-coded spiking port of S-JED without delta coding, which isolates the send-on-delta design. The second is our model with confidence gating removed, so every symbol acts as a virtual pilot at every layer, which isolates the gate's role against error propagation.

What counts as beating the competitors. At a coded PER of 10^-2, our receiver must be within 0.5 dB of floating-point S-JED. It must also use at least 5x less estimated energy per decoded bit than INT8 S-JED. Both conditions must hold with pilots reduced to the minimum of four for four users.

Why not just…?

Why not just quantize S-JED to INT4 and run it on an ordinary accelerator? This is a serious alternative. Low-bit integer networks still need multipliers, and they compute every activation every iteration whether or not anything changed (Sengupta et al.). Our gains come from skipping unchanged symbols and from multiplier-free constellation products, and quantization provides neither. Still, INT4 multipliers are cheap. The INT4 baseline, costed with the same energy model, will show directly whether the accumulate-only savings survive against low-bit multipliers rather than only against 32-bit floating-point ones.

Why not just add "skip converged symbols" logic to a digital S-JED? This is the biggest threat. The savings might come entirely from sparse updates, with nothing gained from spikes. We therefore treat Stage 1, the floating-point event-gated S-JED, as a baseline in its own right. If digital skipping captures nearly all the energy reduction, the honest conclusion is that the contribution is the incremental, constellation-structured loop and that it is hardware-agnostic. The spiking implementation would then have to justify itself on data-side send-on-delta coding and asynchronous idle power alone. We will report the energy gap between the two versions explicitly.

Why not keep the two-stage pipeline with a spiking detector, as Song et al. (2024) did? It avoids the harder feedback design, but it keeps the full pilot overhead and forgoes the 2–4 dB that S-JED gains from coupling estimation and detection. Our decoupled baseline tests this directly. If, at reduced pilot counts, it comes within about 1 dB of the data-aided spiking receiver, the feedback loop is not worth its cost.

Why not use PBiGAMP or precomputed GMM subspace filters? PBiGAMP's low operation count is achieved with floating-point multiplications and mixture-model updates at every iteration. The data-aided GMM estimator of Weißer et al. needs at least about 500 data symbols before its roughly 3 dB NMSE gain appears, which rules out short coherence blocks. We include PBiGAMP as an accuracy reference on short blocks. We do not treat it as an energy competitor unless its operation count, costed with the same energy model, comes within 3x of ours.

Proposed Experiments

Main benchmark. The S-JED 8×4 multi-user MIMO setup: 16-QAM, flat Rayleigh block fading, IEEE 802.11n LDPC codes, pilot counts from 4 to 16, and block lengths from 64 to 512.

Higher-order modulation. The same setup at 4-QAM, and also at 64-QAM. Unfolded solvers are rarely validated at 64-QAM, and denser constellations should produce more boundary flips, so this directly stresses the sparsity premise.

Measured channels. If the Nokia Campus Stuttgart 64-antenna measurements used by Weißer et al. are accessible, we will add them to check whether event sparsity survives spatially correlated real channels. They also let us compare against the reported 3 dB NMSE gain at large block lengths.

Extension. The cell-free benchmark of Sun et al. (2023): 400 potential users with 20% active and QPSK. Inactive users should produce no events at all, which tests whether user-activity sparsity compounds with iteration sparsity.

Metrics. Coded PER and uncoded BER against SNR, channel NMSE, and the fraction of symbols firing per layer. On the cost side, operation counts split into multiply-accumulates and accumulates, estimated energy per decoded bit, and the SNR penalty at matched energy.

Decisive checks.

  1. Sparsity: the fraction of changing symbols per layer at the SNR where PER reaches 10^-2. This check comes first and can end the project.
  2. Robustness gain: whether confidence gating narrows the low-SNR gap to S-JED relative to the ungated ablation, which would show that the gate earns its place for robustness and not only for efficiency.
  3. Energy gap: the ratio between the digital skip-logic baseline and the spiking version, which determines whether neuromorphic execution is essential or only one good implementation of an event-gated loop.