The reservoir trick that needs no living cell

What the work claims

This is a primary experimental result on a physical optical machine, not a simulation. The authors report a deep reservoir computer, its computational core built from optical hardware wrapped in a digital feedback loop, that reaches classification accuracies above 95 percent on three different modalities while operating at gigabit-per-second information rates. The specific figures are about 96 percent on the KTH human-action video set, 95.2 percent on MNIST handwritten digits, and 99.4 percent on the TI-46 spoken-digit set, each reported as a mean over three independent runs of a five-layer architecture.¹

The claim that carries weight for anyone thinking about biological computing is not the accuracy as such; comparable numbers exist in many places. It is that the reservoir transform itself, the nonlinear projection at the heart of the machine, is fixed, random, and passive. The reservoir is not trained. It is a slab of disordered glass that scatters light, read out by a camera, with the recurrence and bookkeeping handled by a digital feedback loop around it. The only trainable part is a linear readout fitted by ridge regression. That is precisely the recipe by which living neural tissue is proposed as a computer, and here the projection at its heart is performed by inert matter, at throughputs well above the rates at which neurons signal.

How it works

Reservoir computing is a way of using a recurrent dynamical system without training its internals. You take some high-dimensional nonlinear system, the reservoir, drive it with your input, read out its state, and train only a simple linear map from that state to the answer. The bet is that a sufficiently rich reservoir projects the input into a space where a linear classifier suffices, so you get the expressive power of a recurrent network while only ever solving a linear regression. The reservoir itself can be almost anything with rich dynamics, which is why people have built reservoirs from electronics, from spintronic devices, from optical systems, and, in the wetware community, from cultured neurons.

The optical reservoir here works as follows. A 532 nanometre laser illuminates a digital micromirror device, a chip of tiny mirrors each of which flips on or off, so it can only display binary patterns. To pack real-valued data onto a binary display the authors use a scheme they call basket encoding, mapping each 8-bit value into a 10-bit binary vector. The patterned light passes through a ground-glass diffuser, which scatters it into a complex speckle field, and a fast monochrome camera records the result. Within the reservoir model, the single source of nonlinearity is the camera itself: a detector measures light intensity, which is the squared magnitude of the optical field, so the unavoidable modulus-square operation of photodetection supplies the nonlinear transform, while the surrounding digital pipeline contributes only fixed encoding steps. State is fed back digitally: selected groups of camera pixels become the reservoir state, are concatenated with the next input, re-encoded, and displayed again, which creates the recurrence.

Depth is achieved by time-multiplexing. Rather than build five separate optical stages, the authors reuse the one physical setup in sequence to emulate a five-layer stack, with only the first layer receiving external input and later layers processing progressively transformed states. Two design knobs matter. A layer-dependent leakage rate, set as a linear function of depth, tunes how much each layer holds its past state versus responds to new input, and the authors find that a decreasing leakage schedule lets early layers track fast features while deeper layers integrate slower structure. A layer-wise bias, implemented by holding a fraction of micromirrors permanently on, shapes the input-to-reservoir mapping per layer. The upshot is an engineered, tunable hierarchy of timescales, produced not by any intrinsic property of a physical medium but by a scheduling choice in software, and the authors quote a processing rate of roughly a thousand frames per second per layer.

Where a skeptic should push

The most load-bearing assumption is that the reservoir is doing the computation. In reservoir computing generally, and here in particular, the readout is a trained linear map and the reservoir is fixed, so it is always fair to ask how much of the performance is the reservoir versus the preprocessing and the readout. The paper is candid that each task carries a conventional preprocessing pipeline: histogram-of-oriented-gradients features and principal-component reduction for the video task, a Lyon cochleagram for the audio task. Those are not trivial front ends. A demanding reading is that the optical stage is a fast random projection stapled between hand-built feature extraction and a linear classifier, which is a useful thing to have but a narrower claim than a general-purpose computer.

Second, these are closed-set classification benchmarks. KTH, MNIST and TI-46 are small, well-behaved, and heavily studied, and reservoir computing has long done well on exactly this kind of task. Nothing here demonstrates the harder capabilities, long-horizon prediction, compositional generalisation, online adaptation, that separate a benchmark win from a computing platform. Third, the accuracies, while strong, are not uniformly state of the art; the authors note that some competing systems score higher using larger trained models or heavier preprocessing, and position their contribution as competitive accuracy at very high speed and modest reservoir size rather than as a new accuracy record. And the parametric-sensitivity results carry their own caution: the benefit of leakage-rate scheduling partly saturates beyond three layers, so the depth story should not be over-read. None of this undermines the result. It bounds it to what it is: a fast, compact reservoir that classifies well on standard sets.

The multi-timescale claim tissue was banking on

For organoid intelligence the useful cut is not the tired observation that silicon can also make rich dynamics. It is more specific and more uncomfortable. The subtlest remaining argument for a living reservoir is that neural tissue offers a naturally occurring hierarchy of timescales for free: membrane time constants, synaptic decays, short-term plasticity, and network-level oscillations span milliseconds to seconds, and that spread of intrinsic dynamics is meant to be what lets tissue hold and mix temporal information without any engineering. This paper shows that the multi-timescale property, the good part, can be produced deliberately by depth plus a per-layer leakage schedule, at least up to the depths and tasks tested, where the authors themselves note the benefit of leakage scheduling partly saturates beyond three layers. Within those limits the temporal hierarchy that tissue was going to offer as an accident of biology is, in optics, a design choice you set. That erodes the timescale argument in particular, not just a generic dynamics argument. The paper does not measure a living reservoir, so this is an argument about what the reservoir primitive requires, not a head to head against tissue.

The genuine threat, then, is to the passive-projection framing of wetware. If the entire value proposition is a fixed nonlinear reservoir with a trained linear readout, tissue competes badly: it is slow, its dynamics drift, it must be kept alive, and it cannot be scheduled the way leakage rates can. A ground-glass diffuser with a digital loop around it will out-throughput a cortical organoid by orders of magnitude, and it does not need an incubator. Any honest pitch for a living reservoir has to concede the projection primitive to physics and move the claim elsewhere.

Where it can move is the one thing this optical system explicitly cannot do, and here the paper hands the field a clean line. The photonic reservoir is fixed and random; its internal transform never changes; all learning lives in the external linear readout. That is generic to reservoir computing and it is exactly the limitation living tissue might not share. Neurons rewire, potentiate, and prune; a culture can in principle modify its own internal weights online in response to its own activity. So the defensible claim for wetware is not that it is a good reservoir but that it need not be a reservoir at all: it could be an adaptive recurrent system that trains its own bulk, which is precisely the capability a diffuser lacks. The opportunity is to stop selling tissue on the projection property optics has taken over, and to design experiments that isolate and measure online self-modification, along with any real edge in energy, fault tolerance, or developmental self-organization, the properties a fixed diffuser does not have. The risk is that if the field keeps pitching living reservoirs on speed, efficiency, or richness of dynamics for this fixed-reservoir primitive, it is competing on ground it has already lost.

The bottom line

Established: an optical reservoir made of a micromirror array, a diffuser and a camera, driven by a digital feedback loop, reaches roughly 96, 95.2 and 99.4 percent on KTH, MNIST and TI-46 at gigabit-per-second information rates, with the reservoir nonlinearity supplied only by intensity detection and all training confined to a linear readout. That is a real, physical demonstration that the reservoir primitive, including a tunable multi-timescale hierarchy up to the depths tested, does not require biology and runs far faster in optics. Hypothesis, not shown: that this generalises past closed-set classification to the predictive and adaptive tasks that would make it a computing platform. For organoid intelligence the durable conclusion is a reallocation of the argument. Concede projection, richness, and even engineered timescales to physics; stake the living-tissue case on online self-modification of internal weights, which this system by construction does not have. What would sharpen the case is a wetware demonstration that internal adaptation, not readout training, carries the performance; what would break the reservoir pitch entirely is a photonic or memristive system that also learns its bulk online, at which point tissue has no reservoir claim left.