A paper that finally asks whether your network represents time

What the work claims

This is a formalization-plus-benchmark paper by Alban Colas-Gattepaille and Alexandre Muzy, both at Universite Cote d'Azur, CNRS, and the I3S laboratory in Nice.¹ Its central construct is the degree of temporal representation: count a network's four function types (encoding, activation, decoding, learning) as temporal only if time appears explicitly as a computational variable in them, for example as timestamps, delays, or intervals. Under this definition, conventional artificial networks and many spiking networks are equally non-temporal: an SNN that evolves over time while its functions only count spikes has the same degree of temporal representation as a feed-forward ANN.¹

The empirical claim: holding architecture and dataset fixed, raising the degree of temporal representation improves both accuracy and computational efficiency on event-based data. Their fully temporal construction, a delay neural network (DeNN) that encodes inputs as average inter-spike intervals, activates through learned spike-time delays, and decodes with a softmin over output spike times, reaches state-of-the-art on the 35-class Google Speech Commands dataset at 98.94% accuracy with 175,000 parameters, compared with 98.11% for an audio transformer, 97.51% for a 5.63-million-parameter transformer keyword spotter, 96.92% for a spiking transformer with 2.13 million parameters, and 95.69% for a strong adaptive-LIF spiking model with 610,000 parameters.¹

How it works

The comparison is deliberately controlled. Event streams (from the speech2spikes audio conversion and the Neuromorphic-MNIST video dataset) are preprocessed into arrays by the event2time algorithm, and then three networks with the same depth and training setup differ only in how time enters their functions. The ANN uses weighted-sum activations. The SNNs (integrate-and-fire, leaky, and adaptive leaky variants) use binary spiking activations with rate-style encoding and softmax decoding. The DeNN fills array cells with each neuron's average inter-spike interval, propagates activity through explicit delay parameters so a neuron's output spike time is the quantity of interest, and decodes by taking a softmin over output spike times, so earlier firing means stronger evidence.¹

On the visual N-MNIST benchmark the ANN and DeNN perform similarly; on the ten-word speech subset, the DeNN reaches strong accuracy from the first training epoch while the ANN catches up slowly and the spiking models trail (the adaptive-LIF is best among them within five epochs, and converges with more). The theoretical section shows why the mapping between a temporal and a spatial activation that exists after training does not imply equivalent learning: even where the two can be mapped onto each other, their parameter trajectories during training differ, the delay parameters absorb short-timescale structure while neuron gain absorbs long-timescale structure, and the temporal network's gradients depend explicitly on the history of event times rather than being dominated by spatial components as in backpropagation through time.¹

Two further results matter for practitioners. A trained DeNN converts to a plain integrate-and-fire network essentially without loss (mean accuracy change across ten trained models of -0.00375 plus or minus 0.0014), so the temporal machinery can be compiled away for deployment. And on efficiency accounting, the DeNN achieves better accuracy-per-chip-cycle ratios than the ANN on both conventional (Intel) and neuromorphic (SpiNNaker) hardware, better accuracy-per-joule on the evaluated GPUs, and slightly better resilience to pruning away low-activity synapses.¹

Where a skeptic should push

First, the state-of-the-art claim. It is a preprint result on a well-studied benchmark, and the comparison table is curated rather than exhaustive; independent replication, or a submission record, would firm it up. The DeNN lineage is the authors' own prior work, so the yardstick and the winning entry come from the same shop. That does not invalidate the result, but it raises the bar for external confirmation.

Second, the definition, while crisp, is stipulative. Counting functions as temporal only when time appears as an explicit variable is one principled choice; a rate-coding population with rich temporal filtering on synapses does represent temporal information in a physically meaningful sense that the four-function audit would score as non-temporal. The paper's own framing, that many SNNs share the ANN's degree of temporal representation, is provocative and partly a consequence of that choice. Readers should treat the degree as a useful design audit, not a measure of temporal depth in general.

Third, the pre-processing does quiet work. The event2time algorithm collapses streams of events into a small number of arrays holding average inter-spike intervals per neuron, which throws away within-window fine structure before any network sees the data. The DeNN's advantage may partly live in this compression doing temporal integration upstream, and the comparison's fairness hinges on all models receiving equally good preprocessing, which the paper claims but which a skeptical reader will want to verify directly.¹

The most load-bearing assumption is that explicit temporal variables, rather than temporal depth acquired implicitly, are what drive the gain. The five-epoch comparison window on speech is short; the authors show the spiking models converge further with more epochs (their supplementary figure), so part of the gap is convergence speed, not asymptotic capacity. And the efficiency analysis leans on cycle counts and GPU joule estimates rather than measured device energy. None of this sinks the thesis, but it calibrates it: demonstrated on event-based benchmarks, under one formalism, pending replication.

Does your organoid compute time or just pass it?

For organoid intelligence this paper is less a result than a yardstick, and the measurement it enables is uncomfortable. Most of the field's readout of living neural tissue is rate-based: spike counts, firing rates, correlation structure over windows. Under this paper's audit, that is the non-temporal degree: the substrate evolves over time, but the interface never lets time appear as a computational variable. The efficiency result suggests the field may be paying for that in the worst way, throwing away precisely the variable that carries the compute, while simultaneously marketing itself as temporal computing. A cynic could read this paper as proving that a large share of what is called neuromorphic or biological computing fails its own central claim.

The constructive reading is better. The DeNN stack maps unusually well onto what microelectrode arrays already measure: inter-spike intervals are exactly the statistic an MEA extracts cheaply, delay parameters are exactly what patterned stimulation can impose, and spike-time readout is what high-bandwidth acquisition supports. The paper's timescale argument, that speech carries millisecond, tens-of-milliseconds, and hundreds-of-milliseconds structure that different parameters absorb, transfers directly to organoids, whose intrinsic time constants run slow, on the order of hundreds of milliseconds to seconds. A temporal-coding organoid interface would set its window structure to the tissue's native timescales rather than forcing silicon-rate coding onto it, which is what most stimulation protocols currently do. The zero-loss compilation result adds a deployment blueprint: train against rich temporal variables, then port to the simplest readout hardware that preserves spike times.

The threats are specific. If the efficiency advantage of explicit temporal representation holds, rate-only organoid computers are structurally capped, and the cap is architectural, not a matter of scaling culture size or electrode count. Second, the same audit applies to hype: a vendor claiming temporal computing from an organoid should be asked which of the four functions manipulate time, with the answer inspected rather than taken from the brochure. Third, the event2time caveat cuts both ways for wetware: compressing event structure into interval statistics upstream, as the DeNN preprocessing does, may discard the very richness that makes living tissue worth using; an organoid pipeline that collapses its readout to average ISIs before analysis would repeat that error with a substrate that is expensive precisely because of its temporal texture. The honest research program this paper points to: audit your organoid interface's degree of temporal representation, then raise it where the tissue's native dynamics, not the silicon's clock, define what time means.

The bottom line

Established: a formal, function-level definition of temporal representation that cleanly separates networks which process events over time from networks whose functions compute on time; under fixed architecture, raising that degree improved accuracy and efficiency on event-based benchmarks, with a 98.94% 35-class speech result at 175,000 parameters and a near-lossless conversion to plain integrate-and-fire hardware. Hypothesis: that explicit temporal variables, rather than temporal richness acquired implicitly, are the source of the gain, and that the result replicates under independent evaluation. What would confirm it: third-party replication of the speech benchmark and measured-device energy. What would break it: evidence that matched preprocessing and matched convergence budgets erase the gap between spiking and delay models. For organoid intelligence, the paper's durable contribution is the audit itself: apply it to any living-tissue interface and you immediately see whether you are computing with time or merely near it.