Research analysis · Neuromorphic benchmarking

The millijoule transformer: what 1.02 mJ per image means for tissue

Spiking neural networks sell themselves on energy efficiency, but spiking vision transformers have mostly inherited the transformer habit of spending compute uniformly everywhere. AdaS-ViT, from Peking University and the Beijing Academy of Artificial Intelligence, breaks that habit: it reads each token's attention score at test time and routes salient tokens to fine-grained firing patterns while everything else gets an energy-frugal coarse treatment. The result on ImageNet-1k is 74 to 79 percent accuracy at 1.02 to 2.45 millijoules per image, estimated. The number matters far beyond computer vision, because it sets the energy bar that biological computing keeps claiming it can clear.

Source: AdaS-ViT: Test-Time Adaptive Routing for Spiking Vision Transformers with Ultra-High Energy Efficiency, Hao, Huang, Xie, Zhang, Yu and Huang, arXiv:2611.23728v1 [cs.NE], submitted 18 November 2026. Primary source. Read: full author HTML, with the main comparison table and the energy-estimation convention verified inline. Preprint, not peer reviewed.

What the work claims

This is a methods-and-benchmarks paper: a new architecture, evaluated on standard vision benchmarks, with energy accounted by a stated convention. Its central observation is that existing spiking vision transformers lock every token into the same firing scheme along the temporal, spatial, and channel dimensions, which spends spikes on tokens that do not need them. AdaS-ViT introduces an Adaptive Efficient Integrate-and-Fire unit that consults the attention score of each spiking token and routes salient regions to fine-grained firing patterns while assigning coarse patterns to the rest.1 Two further components, Spiking Gated Self-Attention and a Spiking Local-Gated Unit, enforce selective spike emission at token and channel granularity, and a progressive spatio-temporal curriculum keeps the co-adapting firing precision and routing masks stable during training.

The headline table, verified against the full text, gives three model sizes on ImageNet-1k, all at 4 inference timesteps: the Small model with 5.36 million parameters consumes 1.02 millijoules per image at 74.03 percent top-1 accuracy and an accuracy-per-energy ratio of 72.58; the Base model with 9.40 million parameters consumes 1.66 millijoules at 77.25 percent, ratio 46.54; the Large model with 14.56 million parameters consumes 2.45 millijoules at 78.94 percent, ratio 32.22. Against representative spiking baselines at comparable accuracy, converted ResNet-34 models consume 59.30 millijoules per image for 71.61 percent (ratio 1.21), and the strongest Spikingformer variant needs 66.34 million parameters and 16.30 millijoules for 77.64 percent (ratio 4.76). The paper's own summary phrases the ratio gain as 32x to 72x.

How it works

The routing idea is simple to state and hard to make stable. In a conventional spiking transformer, every token is encoded with the same number of timesteps and the same channel budget, because the architecture has no way to know which tokens matter before it has read them. AdaE-IF sidesteps that by using the attention score, which the network computes anyway, as a free saliency signal: tokens whose attention scores exceed a threshold are promoted to fine-grained temporal and spatial firing patterns, and the remainder run on coarse patterns that cost a fraction of the spikes. Because the routing decision is made at test time from the input itself, the energy budget adapts per image rather than being fixed at design time.

The gating modules handle the two places where uniform firing is most wasteful. Spiking Gated Self-Attention suppresses spike emission from low-value tokens inside the attention operation itself, and the Spiking Local-Gated Unit does the same at the channel level inside the feed-forward path. Training uses a curriculum that first stabilizes firing precision and only then lets the routing masks co-adapt, which addresses a real failure mode: if masks and firing patterns train from scratch simultaneously, each destabilizes the other.

One accounting detail deserves emphasis, because it governs how the numbers should be read. The energy column in the paper's table is an estimated inference energy per image, computed following the convention of a prior reference work, not a measurement taken on neuromorphic hardware. That convention is standard in this literature and makes cross-paper comparisons possible, but it is a model of operation counts and per-operation costs, not a power measurement from a chip.

Where a skeptic should push

The load-bearing assumption is that estimated energy is the right currency. Every striking figure in this paper, the 1.02 to 2.45 millijoule range, the 32x to 72x ratios, comes from a theoretical accounting convention. Nobody plugged a chip into a power meter. The estimate also covers inference only. Training a transformer on ImageNet, even a spiking one with a curriculum, ran on GPUs for a very large number of steps, and that energy does not appear anywhere in the efficiency claim. For deployment-shaped questions, which is what a fair comparison against biological substrates would require, inference-only accounting flatters silicon.

The accuracy-per-energy ratio is also a ratio of two engineered quantities, and the paper's own table shows the trade it encodes: Small achieves the best ratio (72.58) precisely because it is the least accurate model, while Large achieves the best accuracy (78.94) with the worst ratio in the family (32.22). Ranking architectures by this ratio rewards moving along an accuracy-energy frontier rather than shifting it. Whether AdaS-ViT shifts the frontier or merely navigates it better than prior spiking transformers is harder to tell from a single benchmark table, especially one where all three variants sit on the same curve.

And the comparison class deserves a note. Non-spiking vision transformers at similar parameter counts still lead on raw accuracy, and quantized and pruned conventional networks, which are the actual competition in edge deployment, are barely in the table. The claim "ultra-high energy efficiency" is a claim within the spiking literature, and within that literature it is well supported; the claim that this ends the debate with conventional edge AI is not made by the data shown.

The millijoule bar living tissue must clear

Organoid intelligence's founding pitch is energy. A brain runs on roughly twenty watts, biological synapses are claimed to operate at femtojoule scales, and the argument follows that living tissue must eventually beat silicon on joules per operation. This paper puts a number on what "beat silicon" now means for a real, high-value workload: on the order of one millijoule to classify one ImageNet image at 74 percent accuracy, with the accounting done end to end for inference and the silicon needing no incubator, no perfusion rig, no ethics review, and no well plate of living human cells.

The non-obvious implication cuts in the useful direction for once. Look at what AdaS-ViT's mechanism actually is: a saliency signal, computed by the network itself, that gates where expensive representation gets spent. That is not an engineering novelty so much as a formalization of what cortex does with attention, predictive routing, and gain modulation, processes that living tissue performs natively and that organoid experiments already probe with optogenetic and electrical stimulation. The opportunity is a benchmarking convergence: spiking networks are converging on resource-routed computation, which is the one computation style that a plastic, self-wired substrate is plausibly good at. A closed-loop organoid experiment that gates stimulus precision by measured internal state would be implementing a biological AdaE-IF, and the efficiency accounting in this paper gives exactly the experimental template for claiming the win: joules per decision, end to end, including readout, conversion, and control hardware.

The threat is that almost no one in the biological computing literature reports that number. Dish-scale demonstrations routinely quote spike rates or task accuracy while the energy ledger quietly excludes the amplifiers, digitizers, incubators, and pumps that the substrate cannot function without. Against a published, reproducible, one-millijoule silicon baseline, that accounting habit stops being an omission and starts being a disqualification. If a future organoid vision pipeline wants to claim the efficiency crown, it will need to present a full system bill on the same convention this paper uses, and the margin it needs to show is shrinking with every spiking-transformer release cycle.

A calibrated bottom line for the field: this paper does not threaten biological computing's existence case, which can rest on adaptation, sensing, and closed-world integration. It threatens the energy-only case, and it does so with a mechanism, attention-gated routing, that tissue is unambiguously capable of. That combination, a rising silicon bar built from a cortical trick, is what should be on every organoid computing roadmap this quarter.

The bottom line

Established: within the standard energy-accounting convention of the spiking literature, test-time adaptive routing delivers ImageNet-1k accuracy of 74.03 to 78.94 percent at 1.02 to 2.45 estimated millijoules per image with 5.36 to 14.56 million parameters and 4 timesteps, beating prior spiking transformers on the same accounting by large factors. The architecture contribution, attention-informed firing routing with gated attention and gated channel units, is coherent and mechanistically transparent.

Not established: any measured-hardware energy figure, any training-energy accounting, superiority over non-spiking edge vision systems at deployment, or behavior under distribution shift, since the routing thresholds are learned on ImageNet statistics. What would confirm the practical claim is a Loihi or SpiNNaker2 implementation with a wall-power measurement, and a comparison against a quantized conventional ViT of matched accuracy at the same technology node. What would weaken it is evidence that the routing masks collapse under domain shift, sending every token down the coarse path exactly where fine-grained firing was needed.

For organoid intelligence, the paper's lasting value is the metric it normalizes: joules per decision, stated with an explicit accounting convention. The next biological computing result that wants to be taken seriously on efficiency should be prepared to be scored on exactly this sheet, full system, no exclusions.

Frequently asked questions

Is the 1.02 millijoule figure a hardware measurement?

No. It is an estimated inference energy per image, computed following the convention of a prior reference work in the spiking literature, which models operation counts and per-operation costs. No neuromorphic chip was plugged into a power meter for these numbers, and training energy is not included anywhere in the accounting.

How accurate is AdaS-ViT on ImageNet?

The paper reports three variants, all at 4 inference timesteps: Small reaches 74.03 percent top-1 at 1.02 millijoules, Base reaches 77.25 percent at 1.66 millijoules, and Large reaches 78.94 percent at 2.45 millijoules. Parameter counts are 5.36, 9.40, and 14.56 million respectively.

What does test-time adaptive routing mean?

The network computes attention scores for its tokens anyway, and AdaS-ViT reuses those scores as a saliency signal at inference time. Tokens judged important are routed to fine-grained spiking patterns that cost more energy but preserve detail; the rest run on coarse patterns that emit far fewer spikes. The energy budget therefore adapts per input image instead of being fixed at design time.

What is the accuracy-per-energy ratio?

A compact figure of merit used in this literature: classification accuracy divided by the estimated energy per inference, so higher is better. Note the inversion in this paper's own results: the least accurate model (Small) has the best ratio (72.58) and the most accurate (Large) the worst in the family (32.22), so the ratio partly tracks where a model sits on an accuracy-energy tradeoff rather than a pure efficiency breakthrough.

Does this make spiking networks better than conventional vision models?

Within the spiking transformer literature, clearly yes on the stated accounting. Against the models actually deployed at the edge, quantized and pruned conventional networks, the paper does not settle the question: they are largely absent from the comparison table, and raw accuracy still favors non-spiking transformers of similar size.

Why does an energy-efficiency paper matter for organoid computing?

Energy is the core historical argument for computing on living tissue. This paper sets a concrete, published silicon reference point, on the order of one millijoule per ImageNet classification, achieved with a mechanism, saliency-gated routing, that biological tissue performs natively. Any future biological computing claim based on efficiency will need to beat this class of number with full end-to-end accounting, including incubation, readout, and control hardware.

References

  1. Hao, Z., Huang, Z., Xie, S., Zhang, J., Yu, Z. and Huang, T. AdaS-ViT: Test-Time Adaptive Routing for Spiking Vision Transformers with Ultra-High Energy Efficiency. arXiv preprint arXiv:2611.23728v1 [cs.NE]. 2026. https://arxiv.org/abs/2611.23728. Accessed 2026-11-22.