The other half of the energy budget is subthreshold

What the work claims

This is a methods paper with benchmark evaluation, and its central quantitative claim is diagnostic rather than architectural. The authors took a ResNet-18 spiking network trained on ImageNet and plotted the cumulative distribution of membrane potentials across neurons. Roughly 26 percent of neurons sit in the active spiking region, about 30 percent are subthreshold but close enough to threshold to matter (membrane potential between 0 and threshold), and 44 percent are deeply inhibited, at or below zero, with a negligible probability of firing in the immediate future.¹ Every one of those deeply inhibited neurons still runs its leak and integration arithmetic on every timestep. The paper's argument is that this postsynaptic overhead is a large, chronic, mostly ignored cost, and that instance-level, state-driven gating attacks it directly.

The proposed Dormant Spiking Neuron (DSN) adds a gate to the standard dynamics: when a neuron's membrane potential drops below a negative dormancy threshold, the neuron switches off for the remainder of that input sample, implemented as a 1-bit mask that skips both weight-memory reads and membrane updates. Crucially, the gate is not a post-hoc trick applied to a finished network. In the main variant the dormancy decision is trained end to end with backpropagation through time and surrogate gradients, so the network co-adapts its weights to the constraint and learns a dormancy policy rather than inheriting the accidents of an untrained potential distribution. A training-free plug-and-play variant is also reported, which delivers most of the savings at a modest accuracy cost.

How it works

Spiking networks already exploit presynaptic input sparsity: if no spike arrives, nothing happens. The DSN targets a different axis, postsynaptic state sparsity, which composes multiplicatively with the first. A neuron can be bombarded by presynaptic spikes that hold it deep in inhibition; conventionally it must keep integrating them. The DSN recognizes that a membrane potential parked at or below zero, relative to a threshold at roughly one, is a strong predictor that the neuron has nothing to say about this particular input, and it acts on that prediction by freezing the neuron for the sample. The decision is deliberately conservative in one direction only: once dormant, a neuron cannot wake within the same sample, so the training has to arrange dynamics where deep inhibition really does mean irrelevance.

The headline numbers are strong. On ImageNet with ResNet-18, the trained DSN reaches 64.44 percent top-1 accuracy at an estimated 1.17 mJ per inference, against 4.29 mJ for the baseline, a 73 percent reduction, and it does this while improving accuracy by 1.34 points over the baseline rather than degrading it. On the event-based DVS-Gesture benchmark it matches the baseline's 96.88 percent accuracy with 42 percent less estimated energy; on DVS-CIFAR10 it saves over 52 percent of energy for a 0.2 point accuracy drop, and when the simulation is extended to 24 timesteps the accuracy deviation from baseline grows only to 0.4 points. Across all benchmarks the reported savings run from 40 percent to over 70 percent. The authors also position the method against static pruning, which removes structures that are redundant at the dataset level: at comparable reported cost reductions on ImageNet, the static methods they list land below the baseline accuracy, while the DSN lands above it.

The price is paid in training, not inference. Training the DSN on ImageNet reduces throughput from 742.4 to 601.6 images per second on a single GPU (a 19 percent hit) and raises GPU memory from 15.3 to 20.5 GB (a 34 percent increase), because the surrogate-gradient machinery must track the gating decisions through time. A threshold ablation on CIFAR-100 across three seeds finds a plateau: energy stays low and flat once the dormancy threshold is at least mildly aggressive, then climbs sharply as it is relaxed, with accuracy stable across the tested range; the authors select a threshold of -0.05 for static datasets.

Where a skeptic should push

The load-bearing assumption is that the energy numbers mean what they say, and here the paper's own caveat matters: every energy figure is an analytical estimate from a hardware cost model in the appendix, not a measurement on a chip. The model counts operations in the idiom of a digital accelerator, where a bypassed 32-bit weight read is a genuine saving. That idiom does not transfer cleanly to either analog neuromorphic hardware or to biology. On an analog substrate, a neuron's leak is a physical process that happens whether or not you compute it; in living tissue the analogue of a dormant neuron, a hyperpolarized cell, still pumps ions and still consumes ATP. The DSN is a digital accounting fix for a digital architecture. It tells you where the accounting gap is, not how big the gap is on any substrate that does not charge per operation.

Second, the irreversibility assumption. A neuron that goes dormant cannot return within a sample, which means the network is betting that deep inhibition predicts sustained irrelevance under temporal shift. The authors acknowledge that static post-hoc thresholding with the same bet degrades on nonstationary streams, and their answer is end-to-end training. But all evaluated datasets are classification tasks with bounded, stationary inputs; the regime where this bet should fail, long nonstationary temporal streams, is precisely the regime relevant to embodied and closed-loop computing, and it is exactly where the training-free variant shows its accuracy penalty.

Third, small print on the ImageNet win: a 1.34 point accuracy improvement over the baseline while deleting 73 percent of estimated energy is a claim about one backbone family trained with one recipe. The transformer-based architecture tested also responds well, which supports generality within spike-driven models, but there is no measurement here of latency on hardware, no dynamic workload, and no adversarial or out-of-distribution input where a learned dormancy policy might learn to freeze exactly the wrong neurons.

An energy ledger for living computation

Organoid intelligence rests on an energy argument: a brain-class system runs on twenty watts, so computation on living tissue should be able to beat a datacenter. This paper is a peer-reviewed style correction to how that argument is usually made, because it shows the argument's silicon twin was silently counting wrong. Spike counts are the visible part of a spiking network's budget; the membrane updates of the silent majority are the invisible part, and they turned out to be nearly half the ledger. The same miscounting is endemic in biological computing. Culture energy, perfusion, temperature control and housekeeping metabolism are the tissue analogue of subthreshold updates, and they are routinely left out of joules-per-inference comparisons. If the field wants to claim an energy advantage over silicon, it needs an accounting in which the 44 percent that never fire still show up on the bill.

The opportunity is that biology already invented the dormant neuron, and organoid systems can learn to use it deliberately. Real neural tissue has hyperpolarized down-states, shunting inhibition, and modulatory tone that park large fractions of a network in low-consumption states; that is one reason biological efficiency claims are plausible at all. A closed-loop organoid computer that could gate its tissue into a dormant regime between tasks, and wake defined subpopulations for computation, would be implementing the DSN principle with neurochemistry instead of a mask bit. The paper supplies the quantitative template for evaluating such a scheme: measure the state distribution of your substrate, identify the deeply inhibited fraction, and price its maintenance separately from spiking.

The threat is the mirror image: silicon keeps copying biology's efficiency tricks while remaining reconfigurable. A mechanism whose biological inspiration is obvious, state-driven dormancy, just bought a digital accelerator a 40 to 73 percent energy cut with no accuracy loss, and nothing about the idea is out of reach for hardware vendors. Every such advance narrows the honest margin between tissue and silicon. A further, sharper constraint follows from the training method: the headline result depends on end-to-end backpropagation through time, which no known protocol can run on living tissue. The variant that transfers to organoids is the training-free gate, and it is exactly the variant with the accuracy penalty and the fragility on nonstationary streams. So the transferable lesson is double-edged: state-driven gating works, but the version of it that works best requires a training signal that biology cannot execute, and a field building on living substrates should be honest that it must import that part from silicon.

The bottom line

Established: in a trained spiking ResNet-18, 44 percent of neurons sit deeply inhibited at negligible firing probability while still drawing per-timestep updates, and a trained state-driven dormancy gate removes 40 to 73 percent of estimated computational energy across static and event-based benchmarks, in one case while improving accuracy. Not established: measured hardware savings, savings on analog or biological substrates, robustness of irreversible dormancy on long nonstationary streams, or any result beyond classification benchmarks. The claim is confirmed if device measurements reproduce the analytical estimates; it breaks if dormancy on streamed, drifting inputs either erodes accuracy sharply or forces conservative thresholds that surrender most of the savings. For organoid intelligence the durable contribution is diagnostic: silence is not free, and any honest energy comparison with silicon has to count it on both sides.