2026-07-27
A Neuron review argues that traveling waves sweep over cortical sensory maps, embedding recent history into ongoing activity for predictive computation that feedforward networks cannot do.
The dominant account of how visual cortex computes is feedforward feature extraction with feedback modulation. A stimulus comes in, receptive fields at each layer pull out features, and higher areas send top-down corrections. This is the story a convolutional net implements directly, and it is the default assumption behind most brain-inspired vision models.
It has a stubborn problem. In the cortex of awake animals, neural activity almost never sits still as a static feature map. It propagates across the cortical surface as waves, the neural traveling waves (nTWs). As a wave sweeps over a sensory map (such as primary visual cortex, V1, laid out by visual-field position), it brings distant neurons into the same state in sequence. Pure feedforward or feedback wiring does not produce that kind of spatiotemporal dependence on its own.
This Neuron review comes from researchers who have worked on traveling waves for years: Lyle Muller, Aiden Busch, Zachary Davis, and John Reynolds. Their question is whether these waves are just noise the network happens to make, or something the brain actually computes with. They argue for the second, and promote it into a computational framework.
This is a review with no new experiments. The authors organize the evidence along two lines.
The first is how the waves arise. The core mechanism is the long-range horizontal and patchy connections within a cortical area: neurons that share orientation preference but sit millimeters apart are linked by thin axons, and signals take time to travel along them. Davis et al.'s 2021 Nature Communications modeling showed that once these horizontal-fiber conduction delays sit inside a locally asynchronous-irregular network, traveling waves emerge on their own, with no external drive. Horizontal-fiber delay plus locally asynchronous firing is enough to make waves appear. On the local side, the paper leans on the stabilized supralinear network (SSN), a standard recurrent motif for sensory cortex in which supralinear excitation balanced by inhibition runs near a critical point, amplifying patterns while staying stable. It also discusses whether non-synaptic effects such as ephaptic coupling (neurons influencing each other through extracellular electric fields) help push the wave along.
The second line is what the waves can compute. The paper stresses a counterintuitive ability: because a wave travels across a sensory map with a direction and a speed, it naturally encodes "what was just stimulated" into "what is firing now." In Benigno et al.'s 2023 Nature Communications model, a wave sweeping over a map of visual space ignites a short-term prediction of the incoming stimulus. Generalizing this, the authors propose a framework: sensory cortex does not process input frame by frame. It uses an evolving wave pattern to do spatiotemporal, generative processing, predicting the next input and feeding prediction error back in.
This framework is a cousin of predictive coding, but the emphasis differs. Classical predictive coding is about how error flows between layers, between feedforward and feedback. Here it is about how a wave within a single cortical map unfolds sensory history across space.
As a review, the paper reports no new numbers. Its "results" are the strands of evidence it threads together from more than a decade of work. The hardest ones:
The judgment the authors draw: waves are ubiquitous in awake, behaviorally relevant cortex and track perception, attention, and working memory. Treating them as a computational mechanism fits the data better than treating them as noise.
For people building AI, the value is not a model you can run. It is an angle: the paper makes the mismatch between how the brain computes and how we build networks explicit.
The most direct connections are a clutch of recent works the paper cites. Muller, Churchland, and Sejnowski (2024, Trends in Neurosciences) argue that the self-attention in a Transformer can be read as the cortical-wave mechanism that pulls context from different times into one place. Keller and Welling's neural wave machines (ICML 2023) learn spatiotemporally structured representations with locally coupled oscillatory RNNs. Karuvally et al. (2024, arXiv) found that ordinary trained RNNs grow hidden traveling waves internally, and use them to bind working-memory variables.
The other side is a warning. Arcaro, Fel, Linsley et al. (NeurIPS 2023), and Linsley et al.'s 2025 Trends in Cognitive Sciences piece (better AI does not mean better models of biology) point the same way: convolutional and Transformer vision models that score higher and higher on ImageNet-style benchmarks are drifting further from the actual neural responses in inferotemporal cortex. Set that next to the wave framework, and the implication is that optimizing only feedforward recognition accuracy may be diverging from the real cortical algorithm.
For anyone working on neuro-inspired architectures, recurrent networks, or brain-like computation, this is a direction worth tracking: model the wave dynamics on a spatial map as a first-class citizen, not as something to explain away. For pure benchmark-chasing engineering, there is nothing here to ship this quarter.
The biggest limitation is the paper's nature. It is a perspective and review, so it offers a framework and a set of proposals, not a model that has been run and beaten a baseline. "Waves are a computational principle" is closer to a well-supported conjecture than a settled result.
The evidence has soft spots too. Much of the work supporting waves uses wide-field and mesoscale imaging (voltage-sensitive dyes, GCaMP calcium imaging), which measures the population signal of tens of thousands of neurons, not what an individual neuron or a downstream area actually reads. Orsher et al. (2024, eLife) warned that under limited spatiotemporal sampling, a sequence of discretely activated modules can look exactly like a continuous traveling wave. The most arresting claim here, that waves carry computation within a single trial, cannot yet be cleanly separated from imaging artifacts.
The link to Transformers and neural wave machines is analogy and family resemblance, not mathematical equivalence. Whether attention really simulates a cortical wave is a question both sides can keep arguing. Finally, the full text sits behind a paywall; the account of the argument above is based on the abstract, the keywords, and the complete reference list, and some details from the body could not be checked against the original.