When spiking features match brain activity better than static ones

What the work claims

The central claim is about target choice, not decoder power: spiking neural network features are a better regression target than conventional network features for decoding visual content from functional MRI. Most pipelines map brain activity onto features extracted by an artificial neural network, and the authors question whether those are the right target at all.¹ The experiment isolates that variable. The same fixed decoder, a linear ridge regression with L2 regularisation, maps functional MRI voxel responses onto the target features, and only the source of those target features changes. The baseline is a ResNet-18 artificial network; the alternatives are four spiking variants built on the same architecture, differing only in their neuron dynamics. Because the decoder and the architecture family are held constant, any difference is attributed to the representation.

Lead with the interpretable number. On the Generic Object Decoding dataset, switching to spiking features raises top-1 semantic decoding accuracy from 0.1800 to 0.4400. That is a real and sizeable gain against a chance level near 0.02 on the fifty-way test, and it is also a reminder that the absolute decoder is still modest, wrong more often than right. The paper also reports the feature-prediction error falling from 0.7707 to 0.0282 on the same dataset, a striking figure that should be read with care because it is a scale-dependent regression error, not a task score. Gains are reported across the Mini-Algonauts 2021 and Natural Scenes datasets too, at the voxel and semantic levels, and ablations attribute the advantage to the spiking neuron dynamics and to the number of temporal simulation steps. This is an empirical result about representations, and the authors are explicit that it is not evidence that functional MRI measures spikes.

How it works

The pipeline has three stages. A subject views natural images while functional MRI records blood-oxygen-level-dependent responses, an indirect and slow proxy for neural activity. The same images are passed through a feature extractor, either the standard network or a spiking one, to produce a target vector. A linear decoder is then trained to predict that target vector from the voxel responses, and its output is scored on downstream tasks: semantic classification, image retrieval and reconstruction. The comparison is fair because this decoder is identical across conditions, so it cannot be the reason one representation wins.

The spiking extractor is a spike-based version of ResNet-18 that keeps the residual-block structure but replaces the usual activation units with spiking neurons that integrate and fire over a set number of discrete time steps. Its features are not raw spike trains; they are the network's penultimate activations pooled over those steps, so what changes versus the standard network is the internal dynamics that shape the representation, not that the output is literally a train of events. The four variants, a leaky integrate-and-fire neuron and three richer temporal neuron models, vary the internal dynamics while holding the wiring constant. The authors read the result as spike-shaped features carrying a structure that a linear map can align with measured brain responses more readily, and the time-step ablation supports this by showing that more temporal resolution improves the alignment, which is why they credit the gain to spiking dynamics rather than to the network merely being different.

Where a skeptic should push

Start with what brain-decodable means here. The target is more linearly predictable from voxels, which is a statistical property of the feature geometry as much as a claim about biology. The headline error drop, from 0.7707 to 0.0282, is a regression error on target vectors whose magnitude depends on how those vectors are scaled, and there is a specific reason to discount it: a standard network's dense activations are higher-variance than a spiking network's temporally pooled ones, so a lower-variance target is mechanically easier to regress onto. A good part of the apparent improvement could be that variance difference rather than a truer picture of cortex, which is exactly why the scale-free task score, the rise from 0.18 to 0.44, is the number to trust, and even that leaves a decoder that is wrong more often than right.

The deeper caution is what the setup contains. This is intact human visual cortex, shaped by a lifetime of seeing and organised retinotopically, measured through a slow hemodynamic signal that is not spikes, viewing natural images, with an artificial spiking network standing in for the target. Separate the demonstrated from the asserted. Demonstrated: with the decoder fixed, spiking target features outperform static ones on three human imaging benchmarks, and more temporal steps help. Asserted or out of scope: that this reflects a spiking code in cortex rather than a convenient feature geometry, and that anything here transfers to reading out an untrained lump of neural tissue.

What decodable spikes mean for readout

Reading computation back out of living tissue is the perennial bottleneck in organoid intelligence, so it is tempting to read this paper as good news for the linear-readout premise: if spiking representations are more linearly decodable, maybe an organoid's spiking output is too, recoverable with a cheap linear map. That reading inverts what the study actually did, and the inversion matters. The decoder here was held fixed; what varied was the artificial target the decoder was asked to predict. The result is about which target representation aligns with brain activity, not about whether a biological signal is linearly readable, because the readability of the brain signal was the constant, not the finding. The paper offers no evidence that a substrate's own activity becomes more decodable, so it cannot underwrite the hope that organoid spikes are linearly decodable into anything.

Once the direction is fixed, the honest correction is sharper than a generic hedge. The reason fMRI is decodable at all lives on the input side: a trained, retinotopically organised human visual cortex, viewing natural images, paired with a spiking network trained on ImageNet. The alignment works because both sides sit in the same visual domain, one shaped by a lifetime of vision and the other by supervised training on natural images. A naive organoid has neither the training nor the topographic structure, and its computation is not vision-specific, so there is no matched-domain target model to play the role ImageNet features play here. That is not just a limit on the conclusion; it is a limit on the method, because the whole comparison depends on having a domain-matched artificial encoder, which for a general organoid substrate does not exist.

What still transfers is smaller, and it is a heuristic rather than a derived result. First, a design hint: if temporally structured, spike-shaped target features align with a spiking brain signal better than static ones, then organoid readout pipelines that regress onto static feature vectors may be discarding temporal structure the tissue actually produces, and matching the target to the substrate's temporal code is worth trying. This is an extrapolation from an artificial-target-to-brain result, so it should be tested, not assumed. Second, a method worth borrowing: fix the decoder and vary the target, as a controlled way to ask how decodable a given substrate's output is, separating the substrate's contribution from the readout's cleverness. The threat sits alongside and should be stated at its true weight: this brain-aligned representation was produced in silico, but the alignment is only to a coarse hemodynamic signal under a linear map, so it is evidence that spiking features track BOLD-predictable variance, not that silicon has captured biological neural coding. It is a niche silicon occupies loosely, not one it has won.

The bottom line

Established: with a fixed linear decoder, spiking network features are a better target than static ones for decoding human visual cortex from functional MRI, across three benchmarks, with spiking dynamics and temporal resolution both contributing. The trustworthy headline is the top-1 semantic gain from 0.18 to 0.44, a real but still error-prone decoder; the 0.7707 to 0.0282 error drop is partly a variance artifact of the target scaling. Not established, and not claimed by the authors: that this reveals a spiking code in cortex, or that it says anything about reading out untrained tissue. For organoid intelligence the durable value is a readout-design heuristic, match the target representation to the substrate's temporal code, and a benchmarking method, fix the decoder and vary the target, both to be tested rather than assumed, wrapped around a correction: the study varied an artificial target while holding brain-signal decodability constant, so it gives no licence to expect a naive organoid's spikes to be linearly readable, and its domain-matched design does not exist for a non-visual substrate. What would confirm the useful transfer is applying the fix-decoder, vary-target protocol to real organoid recordings with a domain-appropriate target; what would break the optimistic reading is finding organoid spikes decode no better with temporal targets than static ones.