Spike timing deserves better than a bin: log-time encoding at the MEA front end

What the work claims

Event cameras output asynchronous, timestamped brightness-change events at microsecond resolution, and object detectors built on them usually handle time with recurrent hidden state that accumulates events over successive windows. Lundell, Forssén, Wadenbäck and Lundmark instead encode temporal history directly into the input representation: each event is projected onto four overlapping logarithmically spaced B-spline functions of its age, multiplied by a local spatial confidence, and fed to a purely feed-forward detector. Their claim is that carefully designed representations capture much of the temporal information that recurrent models learn, without any sequential state.¹

This is a methods and benchmarking paper, not a biological result, and its metrics are detection benchmarks (PEDRo and Gen1 automotive datasets), not neural signals. The claim is bounded: they do not say recurrence is obsolete. They show representation design recovers a substantial, measurable share of what recurrence buys, on fixed feed-forward hardware.

How it works

The baseline they beat is CSTR, a compact three-channel surface of recent events. Their representation assigns each incoming event a temporal age and evaluates that age against four second-order B-splines spanning 1.2 to 35 ms. The logarithmic spacing is the point: resolution near zero age (recent events) is fine, while old history is compressed into few channels. Positive and negative event polarities get separate channels, and a normalized local confidence, computed with a 3x3 binomial kernel over temporally weighted event counts, marks how well supported each temporal response is.¹

The hardware-relevant trick is the last stage. Analytical B-splines are expensive to recompute per event, so the authors fit, by least squares, a learned polynomial mapping from a small set of recursively decaying exponential states onto the spline responses. A second-order fit uses 9 coefficients per spline at an approximation error of 1.44 x 10^-5; a fourth-order fit uses 34 coefficients and reaches 1.03 x 10^-6. Because the exponentials can be updated recursively, each event costs constant work, and the whole representation is maintained event by event without ever recomputing a window.¹

The numbers: on PEDRo (40 ms windows), spline encoding with local confidence lifts AP50:95 from 0.566 (CSTR) to 0.630, and against exponential decay it raises small-object AP from 0.015 to 0.062 and small-object recall from 0.142 to 0.219. On Gen1, the spline representation reaches 0.370 AP50:95 versus 0.355 for CSTR, and a window study pushes that to 0.406 with 150 ms windows and finer near-term resolution. The recursive polynomial replacement costs almost nothing: 0.626 AP50:95 versus 0.636 for the analytical spline. Against recurrent ReYOLO variants, the feed-forward detector posts the best AP50 on PEDRo (0.918) but the recurrent models still win AP50:95, especially on Gen1 where they exploit 550 ms of context versus 150 ms.¹

Where a skeptic should push

The single most load-bearing assumption is that benchmark gains in automotive event vision transfer to other asynchronous event domains. That is an analogy, not a result. The datasets are driving and pedestrian scenes; nothing in the paper touches neural spike trains, whose statistics (burst structure, refractory dead time, drifting firing rates over days of recording) differ from photon-driven event generation. Treat every number above as evidence about representations in general, not as a measured property of neural signals.

Second, the margins are uneven. The Gen1 gains are small (0.355 to 0.370 AP50:95), local confidence helps on PEDRo and not on Gen1, and small-object AP in absolute terms remains terrible (0.062 at best). Much of the improvement comes from simply lengthening the temporal window, which any pipeline can do. Third, the recurrent baselines still win where long context matters, and the authors say so plainly. And the detector itself is a 25M-parameter convolutional network evaluated on GPU, so nothing here has been demonstrated on the kind of edge silicon an MEA front end would actually use.

Log-time encoding at the MEA front end

Here is why this matters to anyone building microelectrode array instrumentation. An event-camera stream and a thresholded multiunit spike stream are the same kind of object: asynchronous, sparse, timestamped, polarity-bearing, with useful information at microsecond to second timescales. Yet most MEA analysis pipelines quantize spikes into fixed raster bins of 10, 20, or 50 ms before any learning happens. That binning is the frame-camera move: it throws away exactly the fine timing this paper shows carries outsized discriminative weight. Their small-object AP moved more than fourfold when the finest temporal scale was properly placed, and their window study showed the shortest timescale parameter, not the window length, was the sensitive knob. If a detector on vision events is that sensitive to temporal placement, a decoder on spike timing, where information is known to live at millisecond precision, has no excuse for blind binning.

The opportunity is architectural. The recursive exponential-polynomial update is constant work per event with no window recompute, which maps cleanly onto per-channel digital logic or analog leaky integrators placed at the array edge. Feed-forward decoding then removes the sequential hidden-state dependency that makes recurrent pipelines awkward to parallelize and hard to bound in latency. For closed-loop stimulation, where the whole point is a decision within a known millisecond budget, a representation that carries its own temporal context and never touches a window buffer is a genuine gift: latency becomes a throughput question, not a recurrence-depth question.

The threat is quieter. A logarithmic basis spends its channels on recent history and compresses the tail, which is where slow biological change lives. Electrode drift, seal degradation, gradual shifts in firing rate as a culture matures are all slow-moving signals that a log-time front end represents at very coarse resolution. The paper's own confidence channels were designed for spatial support of recent events, not for flagging slow degradation, so a pipeline built this way could decode today's spikes beautifully while staying blind to the instrument failing over days. The second threat is calibration: their best temporal parameters did not transfer between PEDRo and Gen1, so a representation tuned on one preparation or one recording regime should be assumed wrong for another until revalidated.

The bottom line

Established: on event-vision benchmarks, multi-timescale log-time basis representations with per-event recursive updates recover a substantial share of recurrent temporal modeling, at near-zero approximation cost, in a fully feed-forward pipeline. Hypothesis, untested here: the same construction preserves spike-timing information that binning destroys, at MEA front ends. What would confirm it is a direct comparison on real array data: binned rasters versus log-time basis features fed to matched decoders, with detection or decoding latency and accuracy both reported. What would break it is evidence that neural temporal statistics punish log-compression of the tail, where adaptation and drift live, more than vision statistics do. Either way, the bin should now have to justify its existence.