Photonic Spiking Neurons

The photonic AI module ended on an awkward note: optics does linear algebra beautifully and nonlinearity badly. This lesson opens a module of frontier ideas with a delicious reversal of that verdict. There is one photonic device that is violently, usefully nonlinear, and this course has been holding it in its hand the whole time: the laser itself. Biased just below threshold, a semiconductor laser does something startling. Small inputs produce small, fading responses — but push it past a critical point and it erupts, firing a giant, stereotyped optical pulse whose shape has nothing to do with the input that triggered it, then going briefly dead while it recovers. Integrate, threshold, fire, rest. Any neuroscientist would recognise the behaviour instantly: that is a spiking neuron — the leaky-integrate-and-fire dynamics of the cells in your cortex — except that where your neurons spike a few hundred times per second, an excitable laser spikes billions of times per second. The field that grew from this observation calls itself neuromorphic photonics: build the brain's algorithm out of physics that runs ten million times faster.

Excitability: the all-or-none response

An excitable system is one that sits at a rest state guarded by a threshold. The classic photonic version is a two-section laser: a pumped gain section that stores energy as an inverted carrier population, and a saturable absorber section that blocks lasing — a gate that is opaque to weak light but bleaches transparent under strong light. The pump continuously loads the gain section, like current charging a neuron's membrane. A perturbation — an input optical pulse — briefly adds to the stored energy:

Below threshold, the perturbation produces a small, graded ripple in the output that decays away as the carriers relax. The response is proportional to the input, and then it is forgotten.
Above threshold, the intracavity light grows enough to bleach the absorber, the gate flies open, and the entire stored energy of the gain section dumps out as one tall, narrow pulse — tens of picoseconds wide, with an amplitude and shape set by the laser, not by the input. This is the all-or-none property: a spike is a spike.
Immediately after firing, the gain section is empty. Until the pump refills it, no input of any strength can trigger another spike: a refractory period set by the carrier lifetime, of order a nanosecond.

Play with the threshold below. Three input pulses of different heights arrive; the laser answers each one either with a token sub-threshold ripple or with its full, identical spike (followed by the shallow refractory dip while the gain refills):

Notice what the spike's height does not do: it does not scale with the input. Crossing the threshold by a whisker or by a mile yields the same pulse. That regeneration is precisely what cascaded analog computing was missing — a spike entering the next neuron is as clean as a spike leaving the first, so noise does not compound stage by stage. Information lives in the timing and rate of spikes, which are far more robust currencies than analog amplitude.

The leaky-integrate-and-fire dictionary

Computational neuroscience compresses a biological neuron into three lines of dynamics, and the laser obeys the same three lines with different constants:

a state variable V(t) integrates its input while leaking back toward rest: \frac{\mathrm{d}V}{\mathrm{d}t} \;=\; -\frac{V}{\tau} \;+\; I(t)\,;
when V reaches a threshold \theta, the neuron emits a stereotyped spike and V resets;
for a refractory period after each spike, further input is ignored.

In the biological cell, V is the membrane voltage, the leak is ion channels, and \tau \sim 10\ \mathrm{ms}. In the excitable laser, V is the carrier density in the gain section, the leak is spontaneous carrier recombination, and \tau \sim 1\ \mathrm{ns}. The leak term matters: two weak pulses that arrive close together sum and can jointly cross the threshold, while the same two pulses spaced far apart do not — the first has leaked away before the second arrives. The neuron is a coincidence detector with a fading memory, and the fading is a feature.

	Biological neuron	Excitable laser neuron
state variable	membrane voltage	carrier density (gain)
integration time	~10 ms	~1 ns
spike width	~1 ms	~10–100 ps
refractory period	~1–5 ms	~0.1–1 ns
max sustained rate	~0.2–1 kHz	~1–10 GHz
signal carrier	ions in fluid	photons in a cavity

Run the headline number. A refractory period of 200\ \mathrm{ps} caps the spike rate at 1/(200\ \mathrm{ps}) = 5\ \mathrm{GHz}. A cortical neuron with a 4\ \mathrm{ms} refractory period manages 250\ \mathrm{Hz}. The ratio is 2 \times 10^{7}: the photonic neuron fires in one second what the biological one would need eight months to deliver. Nothing else in the neuromorphic world — analog CMOS, memristor arrays, digital simulation — operates within three orders of magnitude of that speed.

The neuromorphic photonics program

A neuron alone is not a brain; the substance of the program is wiring spiking lasers into networks, and here the field leans on machinery this course has already built. In the broadcast-and-weight architecture (developed at Princeton by Prucnal, Shastri and colleagues), each laser neuron emits its spikes on its own wavelength. All wavelengths are multiplexed onto a shared bus and broadcast to every neuron — the photonic version of an axon fanning out. At each receiving neuron, a bank of microring weight banks tunes how much of each wavelength gets through — the synaptic weights — and a photodetector sums the weighted spikes into a current that perturbs that neuron's gain section. Weighted sum, leaky integration, threshold, fire: a complete spiking layer, running at gigahertz.

Be precise about the frontier line, because this is a frontier lesson. Demonstrated: single excitable lasers reproducing the textbook neuron phenomena (thresholding, refractoriness, temporal integration, inhibition) at GHz rates; few-neuron circuits doing spike-based logic and pattern classification; weight banks steering tens of channels. Speculative: everything at scale — thousands of laser neurons on one chip (each needs gain, and lasers on silicon remain the platform's sore point), on-chip learning rules, and any demonstration that a large photonic spiking network beats a GPU on a task anyone cares about. The promise is enormous and honestly unproven; that is what makes it research.

The two-section laser's dynamics were written down by Yamada in 1993 as three coupled rate equations — gain, absorption, and light intensity — and the mathematics places it in the same family of dynamical systems as the FitzHugh–Nagumo model of the biological neuron: an excitable system near a bifurcation. The rest state is stable, but a saddle point sits close by; a kick that clears the saddle commits the system to a long excursion through phase space — the spike — before it can return home. That is why the laser's spike is stereotyped: the excursion follows the system's own geometry, and the input only supplies the shove. The same excitable mechanics has been demonstrated in half a dozen photonic guises — lasers with saturable absorbers, lasers destabilised by optical injection, quantum-dot lasers, resonant-tunnelling photodetector–laser pairs, and even micropillar lasers a few microns across. The neuron, it turns out, is not a biological invention but a dynamical archetype — biology found it first, photonics found it fastest.

The 10^{7} speed ratio invites a seductive fallacy: "a photonic brain would think ten million times faster than we do." Resist it, on three counts. First, the analogy between a laser and a neuron is dynamical, not architectural: your brain's power lies less in the neuron than in the 10^{15} synapses, their dense 3-D connectivity, and their ability to rewire — and photonic hardware, confined to a 2-D chip with centimetre-scale waveguides, has demonstrated networks of a handful of neurons, not 10^{11}. Second, cascadability is genuinely hard: a spike must carry enough energy to trigger the next neuron, every laser adds noise and consumes pump power, and fan-out divides signal — problems biology solved with active regeneration along every axon. Third, the energy story is subtler than the speed story: a brain runs on 20 W including its memory, learning and packaging; a fair photonic comparison must count the pumps, the thermal tuning and the control electronics, and at that level the advantage is projected, not measured. The honest claim is narrower and still remarkable: for small, fast, latency-critical spiking computations, excitable lasers are the fastest neurons ever built.

Where this goes next

A neuron computes, but a network also has to remember — and every weight in this lesson was held by a tuned microring that forgets its setting the moment the power goes off, burning static power to remember. The next lesson steals a trick from rewritable DVDs: a speck of phase-change glass on a waveguide that holds an optical weight for years with zero power — and multiplies by it every time light passes through.