Photonic neuromorphic computing has a wiring problem that is easy to state and hard to solve. The world is analogue. A spiking neural network speaks in spikes — brief, discrete pulses whose rate and timing carry the information. Something has to translate between the two, and in photonics that translator has been the weak link. A paper from Fudan University and Zhangjiang Laboratory, published in Opto-Electronic Science on 19 October 2026 and newly publicised this week, offers the most convincing translator yet: an all-optical biomimetic spike encoder built on an optically injected Fabry–Pérot laser neuron that fires above 2.0 GHz, encodes with approximately 5-bit linear precision, spends about 0.5 pJ per spike, and operates at 1,550 nm — the wavelength of the fibre network already in the ground.
Earlier optical spiking devices could fire fast, but most produced only binary spikes: on or off, with no stable relationship between input strength and firing rate. Without that relationship, an encoder cannot represent a grey level in an image or the amplitude of a heartbeat. This device establishes exactly that mapping — stronger input intensity yields a higher firing rate, monotonically and linearly — and demonstrates it end to end, from fibre transmission to classification.
The input problem, stated plainly
Spiking networks are attractive for the reasons this desk has covered before: they are event-driven, so energy is spent only when a spike occurs; they are sparse; and their timing carries information a static activation cannot. But a photonic spiking network needs spikes to process, and real signals do not arrive as spikes. A chest X-ray is a grid of grey levels. An electrocardiogram is a continuous voltage waveform. Somewhere, each must be converted into a train of optical pulses.
Electrical spike generators can do the conversion, but they are limited to low spike rates and cannot support high-throughput computation, in the release’s words. Earlier optical spiking devices could produce fast pulses, but most generated only binary spikes, with no stable linear mapping between input strength and firing frequency — they could signal that something happened, but not how much of it there was. And few were designed to work with existing fibre-optic infrastructure, or had demonstrated an end-to-end system handling diverse real-life signals such as medical physiological data. The Fudan–Zhangjiang encoder is designed against all three failures at once: speed, linear precision, and fibre compatibility, with real signals as the test.
The biological analogy is more than marketing. In sensory neurons, stimulus intensity is encoded in firing rate: a brighter light produces more spikes per second. The paper’s device does the same thing in the optical domain — amplitude-to-rate encoding — which is why the authors call it biomimetic. Rate coding is the oldest and most robust neural code, and the easiest to read out: count the spikes in a window, and you recover the input level. The question is whether an optical device can implement that code faithfully at gigahertz rates.
How an injected laser becomes a neuron
The device is a Fabry–Pérot laser neuron under optical injection. A Fabry–Pérot laser is one of the simplest laser structures: a semiconductor gain medium between two mirrors, supporting a comb of longitudinal modes. Left alone, it lases steadily. Inject light from an external master laser into it, and the dynamics change — the injected light can lock the laser’s emission and, in the right regime, push it into a state where it emits a train of short pulses, spikes, rather than a continuous beam.
The encoder exploits that injection-locked dynamics. By tuning three knobs — the injection power, the frequency detuning between the injected light and the laser’s own modes, and the bias current driving the Fabry–Pérot laser — the team establishes a steady, monotonic correlation between input optical intensity and spike firing rate. Turn the input up, and the neuron fires faster; turn it down, and it fires more slowly. That is the entire encoding scheme, and its simplicity is part of the point: no electronic conversion step, no digital-to-analogue stage, no separate modulator.
The paper also reports the laser dynamical mechanism behind the encoding: an understood mapping can be tuned, stabilised, and reproduced. The three tuning knobs give the operator control over the encoding range and the firing-rate slope. Whether that control survives fabrication variation across many devices is not answered here, and it is the first thing to watch.
What the numbers actually show
The headline figures, all from the paper as reported in the release, are these. The spike rate exceeds 2.0 GHz. The encoding precision is approximately 5 bits — roughly 32 distinguishable input levels, mapped linearly onto firing rate. Each spike consumes about 0.5 pJ of energy. The device operates at 1,550 nm, the telecom C-band. And after transmission over 10 km of standard single-mode fibre, the spike encoding performance stays robust.
A spike rate above 2.0 GHz is roughly a million times faster than the fastest biological neurons, which fire at hundreds of hertz at most. That speed is the point of doing this in optics at all: the encoder can keep up with signals that arrive at communications rates, not biological ones. But speed alone is cheap in photonics; what is new is speed combined with linearity.
The 5-bit figure is the more consequential number. It means the firing rate tracks the input proportionally across about 32 distinguishable levels, rather than saturating or jumping. Binary spiking devices encode one bit: something or nothing. A 5-bit encoder can represent a grey level, a signal amplitude, a loudness — the graded quantities that real data is made of. It is not high precision by digital standards — but for a spiking front end, it is a genuine capability, and the number that separates this device from its binary predecessors.
The 0.5 pJ per spike is the energy of the spike itself, as reported in the release — not a system-level figure. It excludes the injection laser, the device bias, and the readout. That is normal for a device paper; the energy case for a system built from these encoders is still unmeasured.
The 1,550 nm wavelength and the 10 km fibre result are the numbers with the most practical significance. Most photonic neuromorphic demonstrations run at laboratory-convenient wavelengths; this one runs where the fibre already is. The 10 km test is a deliberate statement: the spike train survives a realistic fibre span without losing its encoding, so the encoder can sit at one end of a network and the processing at the other.
End to end: MNIST, then real signals
A device that encodes well but classifies nothing is a component looking for a system. The team paired the optical encoder with a spiking neural network and ran it on two kinds of task. On MNIST, the standard handwritten-digit benchmark, the system reached 94.1 percent classification accuracy — comparable to the 93.25 percent reported by the backpropagation-free photonic network this desk covered last month — showing the encoder’s spike trains carry enough information for a real task.
One qualification must be stated explicitly: the spiking neural network in these experiments is digital, not photonic. The optical encoder converts the input into spikes; a digital spiking network, running in software, does the classifying. The paper’s abstract says so plainly, and the distinction matters. What has been demonstrated is an optical front end feeding a digital network, not a fully photonic spiking computer. That is still a meaningful system, but it is not the same claim as an all-photonic network, and it should not be reported as one.
Beyond static images, the team translated continuous physiological and audio waveforms into spike trains and ran proof-of-concept tests on three medical tasks: chest X-ray screening, electrocardiogram arrhythmia identification, and heart-sound classification. The paper further describes a scalable architecture compatible with time-division and wavelength-division multiplexing, supporting multi-user access — many patients’ signals on different time slots or wavelengths feeding the same diagnostic pipeline.
Here the reporting thins out in a way readers should notice. The public summary gives an accuracy figure for MNIST and none for the three medical tasks — no sensitivity or specificity for the X-ray screening, no arrhythmia accuracy, no heart-sound figure, only that the tests were run. That is consistent with proof-of-concept status: the point is that the encoder can ingest very different signal types, not that it has beaten a clinical baseline. The medical results are a demonstration of generality, not evidence of diagnostic performance. Likewise, the release does not say how many multiplexed channels ran simultaneously; the scalable architecture is a proposal, demonstrated in principle, and the number of concurrent users it can serve is still an open question.
The honest ledger
Several limitations deserve equal weight, because they define what this paper does not establish. First, the network is digital. The optical contribution is the encoder; the classification runs in software. A fully photonic spiking network is a different and harder goal, and this paper does not claim it.
Second, the demonstration is a single encoder: one laser neuron, carefully tuned, with three knobs adjusted to find the encoding regime. Scaling to an array — tens or hundreds of encoders on a chip, each holding its calibration — will require fabrication consistency and control electronics that single-device papers do not test. The multiplexing architecture is the team’s answer to scaling, but it is not a substitute for demonstrating many encoders operating together.
Third, the precision ceiling matters. Approximately 5 bits is enough for the tasks shown, and spiking networks are famously tolerant of low precision. But if an application needs finer gradations — and many signal-processing tasks do — the encoder’s linear range becomes the bottleneck. Whether the precision can be pushed higher by refining the injection regime is not addressed in the public summary.
Fourth, there is no system-level energy or latency measurement. The 0.5 pJ per spike is a device figure; the full pipeline has not been measured as a system, and until it is, the “low-power pathway” claim is a direction of travel rather than a result.
Fifth, a note on timing and provenance. The paper was published on 19 October 2026 and made available online on 21 October; it surfaced in the wider press through an EIN Presswire release on 6 November. Its results are three weeks old, not announced yesterday. Every number here comes from the paper itself, via its abstract, and from the journal’s own release — not from secondary coverage. But the full paper text was not independently re-checked for the two figures that appear only in the release — the 0.5 pJ per spike and the 10 km fibre distance — so those are attributed to the release.
Where this fits, and what to watch
This encoder lands in a field converging on the same bottleneck from several directions. The backpropagation-free photonic network covered on this desk last month attacked the training problem: how to teach an optical network without a backward pass. The Koç University work covered earlier in October attacked the simulation-to-reality gap from the hardware side. The Fudan–Zhangjiang encoder attacks the input problem: how to get the real world into the network. These are complementary pieces of the same stack, and the field needs all three.
The fibre compatibility is what distinguishes this result from the other two pieces. Training algorithms and hardware-in-the-loop calibration are laboratory concerns; fibre compatibility is a deployment concern. An encoder that plugs into standard single-mode fibre at 1,550 nm can be imagined in a real system in a way that a bench-top device at an exotic wavelength cannot — which is why the 10 km test, modest as it sounds, is arguably the paper’s most important experiment: it is the one that leaves the laboratory behind.
Three developments would indicate this encoder is becoming more than a laboratory result. The first is an array: several encoders operating simultaneously, on separate wavelength channels, with the multiplexed architecture actually carrying multi-user signals. The second is a system-level energy and latency figure — the full pipeline, measured against a digital baseline. The third is a photonic network on the other end: the encoder feeding an optical spiking network rather than a digital one, closing the loop the paper leaves open. Any one of those would move the result from a component to a system; all three together would make it a platform.
Until then, the right way to read this paper is as the field’s best answer so far to its most neglected question. Photonic neuromorphic computing has spent years perfecting neurons and synapses and comparatively little time asking how the signal gets in. A laser neuron that speaks in spikes at 2 GHz, linearly, over ordinary fibre, is a serious answer.
Sources and further reading
This article is based on the peer-reviewed paper: “Ultrafast all-optical biomimetic encoder for spiking neural networks,” Opto-Electronic Science, Volume 5, published 19 October 2026 (available online 21 October 2026), DOI 10.29026/oes.2026.260051. The research group is led by Prof. Junwen Zhang and Prof. Nan Chi (Fudan University) and Prof. Jianyang Shi (Zhangjiang Laboratory). Spike rate above 2.0 GHz, approximately 5-bit linear encoding precision, 1,550 nm operation, MNIST accuracy of 94.1% with a digital spiking neural network, the multiplexing architecture, and the X-ray, ECG, and heart-sound proof-of-concept tests are from the paper’s abstract; the 0.5 pJ per-spike energy and the 10 km fibre transmission result are as reported in the journal’s EIN Presswire release of 6 November 2026. For the complementary training problem, see this desk’s coverage of the backpropagation-free photonic network; for background, see this desk’s explainer on neuromorphic photonics. No accuracy figures for the three medical tasks were given in the abstract or the release, and none are asserted here.