A spiking classifier that decides before the record ends

What the work claims

Dillon, Javed, Harkin, Dąbkowski and Lienhard replace the conventional matched-filter classifier in superconducting qubit readout with a small spiking neural network (SNN) that processes the measurement window in successive chunks and maintains a running, time-resolved state assignment¹. This is a primary result on real experimental data: the five-qubit, frequency-multiplexed readout dataset published by Lienhard and colleagues, 1.12 million labelled test traces acquired at 2 nanosecond sampling after preparing all 32 computational-basis states of a transmon processor.

The headline numbers: the best SNN reaches a geometric-mean assignment fidelity of 0.9084, against 0.8960 for a carefully built matched-filter benchmark and 0.9100 for a full-trace ANN with the same input features. Quantisation-aware training shows an 8-bit fixed-point implementation loses almost nothing (0.9066). Synthesis with hls4ml targeting a Zynq UltraScale+ RFSoC gives per-chunk inference latencies of 31 to 52 nanoseconds, comfortably inside the 100 nanosecond duration of the shortest chunk, with an initiation interval of one. Several configurations use zero DSP slices.

How it works

Matched filtering compresses the entire measurement record into one overlap score against a precomputed template. That is optimal for a known signal in noise, but it throws away timing: a transient event that occurs partway through the window, such as a qubit relaxing from excited to ground state mid-readout, gets smeared across the whole score and can be assigned to the wrong state entirely. Frequency multiplexing makes it worse, because five readout tones share one feedline, one amplifier chain and one digitiser, so channels carry correlated, state-dependent distortion that no per-channel template captures.

The SNN instead sees the record as a sequence. The 2 nanosecond samples are averaged into 5 or 10 chunks of 200 or 100 nanoseconds; each chunk contributes a local feature (the per-qubit I and Q averages within the chunk) and a cumulative feature (the running average over all chunks so far). A single leaky integrate-and-fire hidden layer of 32 to 128 neurons processes one chunk at a time. Its membrane potentials carry state across chunks, and a membrane-accumulation readout emits fresh class scores after every chunk. Because the network is recurrent in time but tiny, it can be trained with surrogate gradients and then synthesised to FPGA fabric with quantisation-aware training at 8 to 16 bit precision.

Two consequences fall out of the architecture. First, joint processing of all five channels lets the network learn cross-channel correlations: the largest off-diagonal entry of the cross-fidelity matrix, which measures readout crosstalk, drops from 0.041 for the matched filter to 0.013. Second, because a valid classification exists at every chunk boundary, the training loss can be aimed at intermediate times: adding an auxiliary loss term at 600 nanoseconds raises the 600 nanosecond fidelity from 0.8885 to 0.8939 while changing the final-window fidelity by only about 0.0007.

Where a skeptic should push

The gains over matched filtering are real but modest. Moving from 0.8960 to 0.9084 cuts the error rate from 10.4 percent to 9.2 percent, roughly a 12 percent relative improvement, and the full-trace ANN still beats the SNN. The hard part of the dataset does not yield: qubit 2 remains the dominant error source for every model, with single-qubit fidelity around 0.742, and its cross-fidelity improves from 0.479 under matched filtering only to about 0.48 to 0.49 under the learned models. Most of the headline gain comes from modest improvements spread across the easier qubits.

The most load-bearing assumption is that the FPGA numbers mean what they appear to mean. They are C-synthesis estimates from AMD Vitis via hls4ml, not place-and-route results, and they exclude the demodulation, feature construction and data-transfer stages of a complete pipeline. The authors say so plainly. Relatedly, the accumulator in the quantised design uses truncation and wrapping rather than rounding and saturation, accepted as a latency trade; wrapping in an accumulator is a bounded nonlinearity only if the designer bounds it, which is the kind of detail that separates a synthesis report from a shipped instrument. Finally, all reported model results come from a single training run with a fixed seed; the authors report seed-to-seed fidelity fluctuations at the 0.0001 level from repeated runs, but every table in the paper rests on one draw.

Streaming decisions and the MEA acquisition chain

Strip away the microwave hardware and the readout chain here is the MEA acquisition chain, channel for channel. Dozens to thousands of electrodes share digitiser bandwidth; channels are correlated through common-mode noise, culture-medium coupling and stimulation artifacts; and the events that matter, burst onsets and seizure-like cascades among them, are time-localised transients that a fixed-window integrator smears exactly the way a matched filter smears a relaxation event. The paper demonstrates, on a million-trace experimental dataset, that a streaming classifier with a persistent internal state beats the integrate-then-classify baseline and nearly matches a network that must see the whole window first.

The non-obvious implication is about latency, not accuracy. The paper's own mid-readout experiment is the tell: by adding one loss term at 600 nanoseconds, the network re-allocates fidelity toward the middle of the window, gaining decision time at a cost of 0.0007 in final fidelity. Mapped onto MEA hardware, that is a closed-loop controller that can commit to a stimulation decision at 60 percent of the integration window, or a spike-detection front end that stops integrating once its running confidence crosses threshold, shortening effective per-channel dwell time and with it the power and data-rate budget of the whole array. An integrator-then-classify chain cannot make that trade at all; the decision latency is structurally locked to the full window plus inference.

There is also a practical procurement signal in the resource tables. The small models hit their latency budget at 8 to 10 bit precision with zero DSP slices in several configurations, meaning a classifier of this class can live on the same FPGA as the acquisition state machines without fighting the signal-processing fabric for multipliers. For vendors integrating on-headstage inference, that is the difference between a cohabiting block and a second chip.

The genuine threat is twofold. First, the accuracy ceiling: a full-trace ANN still wins, and an MEA vendor shipping a streaming classifier will face recordings where the transient structure is exactly what a fixed template handles best, so the honest pitch is latency and power, not accuracy. Second, the estimate-versus-measurement gap: HLS latencies without place-and-route have a history of optimistic surprises, and a closed-loop instrument built on a 100 nanosecond budget that turns out to need 140 nanoseconds after routing is not a paper, it is a redesign.

The bottom line

Established: on a large, real, multiplexed dataset, a streaming spiking classifier beats matched filtering, suppresses correlated channel error, quantises to 8 bits nearly for free, and fits its per-chunk update inside the chunk budget at the synthesis-estimate level. Hypothesis: that these estimates survive place-and-route and a full pipeline, and that the mid-window fidelity trade transfers to biological signals, where the ground truth is messier than 32 prepared basis states. The claim breaks if a placed-and-routed build misses the chunk latency, or if on real MEA streams the streaming advantage evaporates once bursts, motion artifacts and stimulation transients replace well-characterised qubit relaxation.