Cancellation as Inhibition: A Streaming Spiking Receiver for Imperfect Successive Interference Cancellation

When several users share the same radio channel at different power levels, a receiver separates them by successive interference cancellation (SIC). It decodes the loudest user, rebuilds that signal, subtracts it, and then decodes the quieter user from what remains. In real hardware the subtraction is never exact. Channel-estimate errors, quantization and amplifier distortion leave a residue proportional to the loud user's power. That residue caps the quiet user's signal quality, so outage stops improving at high SNR. Learned receivers such as SICNet handle this mismatch well, but they run a stack of dense neural networks for every user. That cost is hard to justify on battery-powered terminals, which are exactly where cheap many-user access matters most.

This proposal builds a spiking neural network that does SIC natively. A spiking network is one whose neurons exchange binary events and whose hardware only adds a weight when an event arrives, so energy tracks activity. Two ideas carry the design. First, the loud user's decoded estimate is not converted back to a waveform and subtracted. It is fed as inhibitory spikes into the quiet user's neurons, so cancellation becomes a learned, confidence-weighted reduction of membrane voltage. Second, the user stages run concurrently, with each stage streaming partial estimates to the next, instead of waiting one after another. If this works, a receiver keeps the robustness of learned SIC under imperfect cancellation, uses mostly additions, and pays a latency close to one stage rather than one stage per user.

Core Claim

A spiking SIC receiver can match the robustness of dense learned SIC under imperfect cancellation, measured as SNR loss at a fixed symbol error rate, at lower energy than an equally accurate low-bit artificial network. It does this by turning subtraction into learned inhibitory synapses and by pipelining stages in time. The claim has three separable parts, and each can fail on its own:

  1. Learned inhibition reduces the effective residual fraction $\beta$ below that of hard decode-and-subtract whenever the residual comes from structured error, such as CSI mismatch or a quantized front end.
  2. Streaming cuts decoding latency from $K\cdot T$ timesteps to roughly $T + (K-1)d$ for $K$ users, $T$ integration steps and an inter-stage spike delay $d$.
  3. The energy advantage over a quantized artificial network comes from these two structural features, not merely from replacing multiplies with additions.

Prior Work and Its Roles

The motivating problem. The analysis in Outage Performance of A Unified Non-Orthogonal Multiple Access Framework shows that under imperfect SIC, diversity order collapses to zero for both power-domain and code-domain multiple access. It also shows that the stronger "near" user can fall below orthogonal access. The proposal takes this floor as the quantity to move.

The accuracy reference, and a method we reuse. Deep Learning Based Successive Interference Cancellation for the Non-Orthogonal Downlink introduces SICNet, which cascades dense networks that pass soft estimates forward. It loses under 0.5 dB at a symbol error rate of $10^{-3}$ under imperfect CSI. We keep its cascade-with-soft-information structure and its training-from-decoder-feedback idea. We change two things: the cancellation operator becomes inhibitory synapses, and stage execution becomes streamed rather than sequential.

The neuromorphic method we borrow. Neuromorphic In-Context Learning for Energy-Efficient MIMO Symbol Detection shows that a spiking detector with Bernoulli spike encoding and leaky integrate-and-fire neurons can approach linear MMSE detection. It reports 5.4× to 26.8× lower power than an artificial-network counterpart. We borrow its encoding and its per-operation energy accounting. That detector solves a joint linear inversion with no unequal powers, no peeling, and no residual floor. Whether such a joint detector already suffices for superimposed users is a question we test directly rather than assume away.

Training method. Threshold-dependent batch normalization from Going Deeper With Directly-Trained Larger Spiking Neural Networks stabilizes deep spiking training, and we use it for the stage networks.

Front-end option. The resonate-and-fire neurons in Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons act as in-neuron band-pass filters on raw radio samples. We consider them as an optional replacement for encoding, but that paper classifies signals and does not separate them.

Off-platform alternatives that constrain the scope. Structured Masked Diffusion for Joint Multiuser Decoding avoids SIC entirely through GPU-parallel diffusion, which is out of reach on a terminal's power budget. Hybrid SIC Aided Hybrid NOMA shows that switching decoding order removes fixed-order floors. That paper constrains our design: a fixed-order spiking cascade will inherit fixed-order floors in some channel-gain regimes. We therefore treat order selection as an explicit extension rather than claim it away.

Research Program

Encoding. The receiver encodes in-phase and quadrature samples into spike trains, using Bernoulli rate encoding as the default and delta encoding as a sparsity variant.

Stage 1. This leaky integrate-and-fire population reads the input spikes. Its output layer has one neuron group per constellation point, so firing rates act as a soft posterior over the strong user's symbol.

Stage 2 and cancellation. Stage 2 reads the same input spikes and also receives stage 1's output spikes through learned inhibitory weights. A stage-2 neuron $i$ updates its membrane voltage as

$$ v_i[t+1] = \lambda v_i[t] + \sum_j w_{ij}\, s^{y}_j[t] - \sum_k u_{ik}\, s^{(1)}_k[t], $$

where $\lambda$ is the leak, $s^{y}_j[t]$ are input spikes, $s^{(1)}_k[t]$ are stage-1 output spikes, $w_{ij}$ are ordinary input weights, and $u_{ik} \ge 0$ are inhibitory cancellation weights. Inhibition is proportional to how strongly stage 1 believes each symbol. This makes the cancellation soft, so a low-confidence first decision subtracts little. That is the spiking analogue of avoiding the error propagation in which subtracting a wrong decision adds interference.

Streaming. Stage 1 starts emitting output spikes after a short delay $d \ll T$. Stage 2 therefore integrates in parallel, and the final user's estimate is ready after about $T + (K-1)d$ steps.

Training. All weights are trained end to end with surrogate gradients and threshold-dependent normalization, using a sum of per-user cross-entropy losses. The outputs are read as log-likelihood ratios for a downstream LDPC decoder.

Measurement. Energy is computed as counted additions and multiplications times per-operation and memory-access energies, using the same accounting convention as the neuromorphic MIMO paper. Physical neuromorphic silicon is not assumed, and the result is reported as an operation-level estimate.

Baselines

Classical SIC (the do-nothing reference). This is LMMSE channel estimation followed by fixed-order decode, reconstruct and subtract. It needs no training and has the lowest compute of any option. Beating it means achieving lower SNR loss under CSI error and a lower or later error floor as the injected $\beta$ grows. It is not expected to lose on energy, so the spiking receiver must justify its energy over classical SIC by accuracy alone.

SICNet (the strongest accuracy competitor). It is the only method in the file built for learned SIC under imperfect CSI and non-linear quantized channels. Its reported figure is an SNR loss under 0.5 dB at a symbol error rate of $10^{-3}$ under imperfect CSI (SICNet). We reimplement it on identical simulated data with the same training-set sizes (SICNet's reported regime is 1,000–5,000 samples) and the same hyperparameter search budget. Beating it means staying within 0.5 dB of its symbol error rate curve while using less energy per decoded frame.

SICNet quantized to 4-bit and binary weights (the strongest energy competitor). A low-bit artificial network also avoids most expensive multiplies, so it is the fair test of whether spiking adds anything. It gets the same quantization-aware training budget. Beating it means lower energy at matched symbol error rate.

A joint spiking detector. This adapts the neuromorphic in-context MIMO detector (Neuromorphic In-Context Learning) to output both users at once, with no cascade. It tests whether the SIC structure is needed at all.

Two ablations. The first is a spiking cascade that decodes stage 1 to floating point and subtracts a waveform. It isolates the inhibition ingredient. The second is the same cascade run strictly sequentially. It isolates streaming.

No reported result was found on Lacuna for spiking SIC on NOMA, so the spiking numbers have no prior reference point.

Why not just…?

Why not just quantize SICNet? This is the real threat. If a 4-bit SICNet matches our energy at equal accuracy, then the energy story reduces to "use fewer bits," and spiking execution is unnecessary for this task. The quantized-SICNet baseline settles this directly. We will report energy-versus-SNR-loss curves for both, and a loss here would be published as a negative result for neuromorphic SIC. Spiking can win only through activity sparsity and streamed latency, which a quantized dense network does not share. Both are measured.

Why not just keep classical SIC with better channel estimates? Classical SIC is cheapest. But the unified outage analysis shows that any nonzero $\beta$ leaves a floor (Outage Performance), and SICNet's comparison shows larger CSI-induced loss for analytic SIC. Our $\beta$ sweep shows where the crossover lies. In regimes where classical SIC is within a fraction of a dB, the right recommendation is classical SIC, and we will say so.

Why not just use a joint spiking detector with no cascade? Joint detection avoids error propagation and sequential latency. However, the authors of the joint detector evaluated it only on equal-power linear channels, and joint networks risk collapsing onto the strong user. The joint-detector baseline tests this. If it matches the cascade, the inhibition idea is not needed.

Proposed Experiments and Decisive Checks

Data. The data is simulated two- and three-user downlink power-domain NOMA following SICNet's scenarios, with Gaussian channels, block fading, CSI error and a low-resolution quantized front end. On top of these we add an explicit residual-fraction sweep, so that $\beta$ is controlled. A harder setting uses 3GPP CDL-B channels at 30 km/h, following A Universal Neural Receiver, with users superimposed.

Metrics. We report per-user symbol and bit error rate versus SNR and the high-SNR floor level. We also report the effective residual fraction $\beta$, inferred from the weak user's achieved SINR, and the SNR loss at a symbol error rate of $10^{-3}$. Energy per frame is broken into computation and memory access, alongside latency in timesteps and spike counts per stage.

Check 1: inhibition. Inhibitory cancellation should lower the effective $\beta$ relative to the float-subtract ablation when the residual comes from CSI error or quantization. We predict no gain when $\beta$ is injected as unstructured random residual, since nothing learnable remains. Seeing that pattern would show the network learns the structure of the impairment rather than merely regularizing.

Check 2: energy. At matched accuracy, within 0.5 dB of full SICNet, the spiking receiver's energy should be below 4-bit SICNet. Failure on this check refutes the energy claim.

Check 3: latency. Final-stage latency should track $T + (K-1)d$ as $K$ grows from 2 to 3. Total error should rise by no more than a small, reported margin relative to sequential execution.

Check 4: regime coverage. We sweep channel-gain ratios to expose any fixed-order floor of the kind hybrid SIC removes. Where one appears, we test a light extension that runs both orderings as competing populations, with mutual inhibition selecting one, and report whether it closes the gap.