When the stimulator, not the controller, sets the energy budget

What the work claims

The paper argues that once a neural-network controller is made efficient, the energy it spends on inference stops being the thing worth minimising, because in any physical closed loop the actuator, the part that acts on the world, can cost as much as or more than the controller. It demonstrates this in adaptive deep brain stimulation for Parkinson's disease, where the actuator is the neurostimulator delivering electrical charge into brain tissue.¹ The central move is to write the actuator's energy directly into the reinforcement learning reward, so the controller optimises what it can actually reduce rather than only its own compute.

The reported numbers are a 45.2 percent reduction in pathological alpha-beta band power and an 80.0 percent reduction in stimulation charge relative to continuous stimulation, both in a biophysical circuit simulation, averaged across ten training seeds. The learned policy is then compressed by sparsity-constrained knowledge distillation and run on a SynSense XyloAudio 3 neuromorphic processor at 0.52 milliwatts, which the authors report as 28.1 times lower energy per inference than an equivalent conventional network on an edge GPU. This is a methods-and-simulation result, not a clinical one, and the weighting should follow: the conceptual contribution, put the actuator in the objective, is strong and general, while the specific percentages are properties of a model.

How it works

The controlled system is a biophysically detailed model of the cortico-basal-ganglia-thalamic circuit, with Hodgkin-Huxley neuron dynamics, tuned to the dopamine-depleted state of the standard 6-hydroxydopamine lesioned rat used to model Parkinson's disease. The pathological signature it must suppress is an exaggerated oscillation in the alpha-beta band, roughly 7 to 35 hertz, a well-established biomarker of Parkinsonian basal-ganglia activity. The controller is a Deep Spiking Q-Network, a reinforcement learner whose value estimates are computed by a spiking network, and it reads the circuit in its native spiking form rather than collapsing the signal into time-averaged spectral windows. The authors make a specific point of this: averaging spikes into per-window mean rates discards the fine inter-burst timing that actually tracks pathological synchrony, and an event-driven spiking controller preserves it.

At each decision step the network chooses adjustments to three stimulation parameters, frequency, pulse width and amplitude, each nudged up, held, or nudged down, within clinically bounded limits such as a frequency ceiling of 180 hertz. The key design choice is the reward. Rather than rewarding only oscillation suppression, the objective jointly credits therapeutic effect and penalises the physical charge delivered, so a policy that achieves the same suppression with less charge scores higher. That is what drives the charge down by 80 percent: the controller learns to stimulate when it helps and to stay quiet otherwise, which a sensory-ablation check confirmed by showing the policy falls silent when deprived of meaningful input rather than stimulating blindly. Finally, deployment tackles the other half of the budget. Implantable devices sit under hard physical limits, a sub-milliwatt power envelope and a tissue-heating constraint of about 2 degrees Celsius, and the distilled spiking policy fits that envelope on neuromorphic hardware, trading synaptic operations for accuracy until it is sparse enough to run at 0.52 milliwatts.¹

Where a skeptic should push

The load-bearing claim is that these results transfer beyond the simulation they were produced in, and that is exactly what is not shown. The entire study lives inside a rat-tuned biophysical model; there is no in vivo validation, no living tissue, and no human. The 45.2 percent and 80.0 percent figures are relative to continuous stimulation inside that model, and a model that both defines the disease signal and scores its suppression can flatter a controller in ways a real basal ganglia will not. The biomarker itself is a simplification: real Parkinsonian pathophysiology is more than one band of one signal, and a policy tuned to suppress that band may not map onto symptom relief.

The efficiency headline needs the same scrutiny. The 28.1 times figure is one specific comparison, a distilled spiking network on a SynSense chip against one conventional network on one edge GPU, so it is a favourable hardware-to-hardware anecdote rather than a paradigm-level law, and energy per inference is not the same as total system energy over a treatment session. It is worth separating what is genuinely demonstrated from what is asserted. Demonstrated in simulation: writing actuator charge into the reward reduces delivered charge substantially while retaining suppression, and the resulting policy is small enough for a sub-milliwatt neuromorphic part. Asserted or untested: that the effect sizes survive in vivo, that the biomarker tracks clinical benefit, and that the energy ratio generalises past the single hardware pairing. There is also a consistency point to keep in view: if the actuator is the dominant term, then the controller's own efficiency numbers, the 0.52 milliwatt draw and the 28.1 times edge, are by the paper's own logic the minor part of the budget, and they matter less than their prominence suggests.

What the write channel costs a living computer

Organoid intelligence sells itself on energy: the brain does extraordinary computation on roughly twenty watts, so a living substrate should undercut silicon by orders of magnitude. This paper attacks the accounting behind that pitch without ever mentioning organoids. Its thesis is that in a closed-loop physical system the inference is not the only cost that matters: the actuator, the part that acts on the world, becomes the cost worth reducing once the controller is efficient. Any deployed organoid computer is a closed-loop physical system too, so its write channel, the stimulation that acts on the tissue, is a cost the twenty-watt story omits, and its life support is a further and separate cost this paper does not speak to.

Make the mapping concrete. A useful organoid does not compute in a vacuum: it needs perfusion, temperature regulation, gas exchange and fresh media to stay alive, and it needs a read channel, the microelectrode or optical recording, and a write channel, the stimulation that injects information back in. The tissue's own metabolic draw is the part that looks cheap. Of these, the write channel is the direct analogue of this paper's actuator, and that is the part the paper actually grounds: the stimulation charge, not the neurons' spiking, is a first-order energy term. Perfusion, thermal control and gas exchange are a separate and probably larger cost that this source does not address and should not be made to vouch for, though they point the same way. The honest implication is narrower than a slogan: an efficiency claim that counts only the tissue is measuring the wrong boundary rather than the relevant figure of merit, and at minimum the write channel belongs in the ledger.

There is a real opportunity in the same result, and it is a blueprint rather than a caution. The system here is a controller trained in a biophysical model of neural tissue, then compressed onto a sub-milliwatt neuromorphic chip that stimulates that tissue in a closed loop. Reverse the roles and it is a template for interfacing silicon with a living organoid: the read-compute-write loop, the reward that charges the write channel for the energy it spends, and the neuromorphic front end that fits the same thermal and power envelope living tissue also imposes. The reward-shaping idea in particular ports naturally. Writing to an organoid is not free either; charge injection degrades electrodes and can damage tissue, so a training protocol that penalises stimulation charge, as this one penalises actuator energy, is a way to keep a living substrate healthy while it learns. The genuine threat is one of positioning. If a half-milliwatt spiking chip can run adaptive neuromodulation inside the very thermal and power limits an implant demands, then the case for growing tissue to do implantable control is weak, because silicon already owns that niche on energy, size and manufacturability. The defensible role for the living substrate is not to be the low-power controller; it is to be the thing worth controlling, or the higher-fidelity environment in which such controllers are trained.

The bottom line

Established in simulation: putting actuator charge into a reinforcement learning reward lets a spiking controller cut stimulation charge by about 80 percent while still reducing the pathological rhythm by about 45 percent, and the policy compresses onto a sub-milliwatt neuromorphic processor. Not established: that any of this holds in living tissue, that the single biomarker tracks clinical benefit, or that the 28.1 times energy figure generalises beyond one hardware pairing. For organoid intelligence the durable takeaway is an accounting one: a living computer is a closed loop, its write channel is a real energy term a tissue-only account omits, its life support is a further cost that needs its own measurement, and an efficiency claim that ignores them measures the wrong boundary rather than the relevant figure of merit. What would confirm the transfer is an in vivo closed-loop demonstration with the same reward design; what would undercut the broader OI reading is a full-system energy audit showing tissue upkeep is negligible against compute, which no current platform has shown.