Stimulus-specific neuronal oscillations in orientation columns of cat visual cortex.

TL;DR: In 1989 Gray and Singer showed that neurons in cat visual cortex fire rhythmically near 40 Hz in a stimulus-specific way, and proposed that synchrony between distant neurons is how the brain tags which features belong to the same object — temporal correlation as the brain's variable-binding mechanism. The hypothesis could never be properly tested with the handful of electrodes and zero causal tools of the era; today, Neuropixels-scale recording plus optogenetic entrainment of gamma rhythms make a decisive test feasible, just as the binding problem has independently resurfaced in AI through object-centric and synchrony-based networks.

The idea

Your visual system processes an object's color, orientation, motion, and depth in distributed, partly specialized populations. When two objects are in view, the red features must be bound to this object's shape and the blue to that one. A rate code alone struggles here: if a neuron for "red" and a neuron for "vertical" are both active, nothing in their firing rates says whether red-and-vertical belong to the same thing. This is the binding problem, and in symbolic terms it's a variable-binding problem — the same one that plagues connectionist networks.

Christoph von der Malsburg had proposed a solution in 1981 (the "correlation theory of brain function," circulated as an internal report): use time as the glue. Neurons representing features of the same object fire in synchrony; neurons representing different objects fire out of phase. Relations are encoded not in which neurons fire but in when they fire relative to each other. This buys a combinatorial binding mechanism for free — no grandmother cells for every conjunction, no wiring explosion.

Gray and Singer's PNAS paper supplied the first serious physiological ammunition. Recording in areas 17 and 18 of anesthetized cats, they found that responses to an optimally oriented moving light bar were not just elevated in rate — they oscillated, with spiking probability peaking near 40 Hz, phase-locked to a local field potential oscillation at the same electrode. Crucially, the oscillation was stimulus-specific: LFP oscillation amplitude was maximal when the stimulus matched the local column's orientation and direction preference. And it was cortical in origin — LGN recordings showed no 20–70 Hz oscillation, ruling out inherited thalamic rhythm.

The companion result that made the story famous came the same year (Gray, König, Engel & Singer, Nature 1989): neurons in separate columns, even with non-overlapping receptive fields, synchronized their oscillations with near-zero phase lag — and did so more strongly when their receptive fields were stimulated by a single long bar than by two independent bars. Synchrony tracked a global property of the stimulus (is this one object or two?), exactly what a binding tag should do.

Why it could not be settled in 1989

The hypothesis makes claims about populations, multiple areas, milliseconds, perception, and causality. The 1989 toolkit failed on every count.

Population scale. Gray and Singer recorded with a few microelectrodes yielding single-unit or small multi-unit clusters — realistically fewer than ~10 units at a time. Binding-by-synchrony is a statement about the joint phase structure of assemblies spanning thousands of neurons across areas; with pairwise recordings you can only measure cross-correlograms between a handful of cells, and you have essentially no power to distinguish "assembly-level phase coding" from "shared common input" or "generic network resonance."

Behavioral state. The recordings were in anesthetized cats. A binding code for perception requires an animal that is perceiving — ideally reporting what it sees, so that synchrony can be compared on correct versus binding-error trials. Chronic awake multi-area recording at millisecond precision was roughly a decade away even in crude form.

Causality. The killer requirement: to show synchrony is a mechanism rather than an epiphenomenon of shared drive or recurrent dynamics, you must set the phase yourself — entrain two populations in phase or in anti-phase and show perception changes accordingly. In 1989 the available manipulations were lesions, cooling, and pharmacology, all with timescales of minutes to permanent. Millisecond-precise, cell-type-specific control did not exist and was not on anyone's roadmap; optogenetics arrived around 2005.

Analysis. Even the statistics were hard. Distinguishing genuine oscillatory synchrony from broadband correlations, stimulus-locked transients, and spectral leakage requires methods (multitaper spectra, spike-field coherence corrections, generalized phase) that were developed largely in the 1990s–2010s, often in response to controversies this very literature generated.

What happened in between: a real fight

The 1990s saw an explosion of supporting studies (Engel, König, Kreiter, Singer on interhemispheric and cross-areal synchrony; Fries et al. 1997 linking synchrony to perceptual dominance in strabismic cats). Then came the backlash. Shadlen and Movshon's 1999 critique argued the evidence was correlational, the readout mechanism unspecified, and cortical dynamics too noisy for precise spike timing to carry a code. Thiele and Stoner (2003) found that perceptual binding of plaid components in MT did not track synchrony. Ray and Maunsell showed gamma frequency and power shift with contrast and stimulus size in ways awkward for a stable binding tag — gamma looked more like a signature of local excitation-inhibition dynamics than a clock.

The idea did not die; it mutated. Fries's communication through coherence (2005, refined 2015) repurposed gamma phase alignment as a mechanism for gating routing between areas rather than binding per se — and accumulated stronger evidence, especially in attention (V1–V4 coherence). Meanwhile the PING mechanism (pyramidal-interneuron gamma) gave the rhythm a solid circuit basis, and Cardin et al. and Sohal et al. (2009) showed optogenetic drive of parvalbumin interneurons at ~40 Hz entrains cortical gamma and modulates sensory responses — proving the causal manipulation half of the needed experiment is now routine.

So the honest scoreboard: stimulus-dependent gamma and long-range synchrony are real and replicated; synchrony-as-the-binding-mechanism remains unproven and contested. It is a genuinely open 35-year question, not a settled one in either direction.

What changed, and the 2026 experiment

Three capabilities now exist that define a decisive test:

1. Scale: Multiple Neuropixels 2.0 probes record thousands of well-isolated units simultaneously across V1, higher visual areas, and thalamus, with sub-millisecond spike timing — enough to estimate assembly-level phase structure trial by trial rather than pairwise cross-correlograms averaged over hours.
2. Causal phase control: Closed-loop optogenetics can entrain gamma in a targeted population and, critically, lock it in phase or anti-phase with ongoing rhythms elsewhere, on the fly, guided by real-time LFP.
3. Behavior + models: Mice and marmosets can be trained on figure-ground and feature-conjunction tasks where binding errors (illusory conjunctions) are measurable, and modern latent-variable models can decode "which features were bound to which object" from population activity.

The revival experiment, concretely: an animal views two overlapping or adjacent objects with conflicting feature conjunctions and reports a conjunction judgment. Record populations representing both objects' features across two or more areas. First, the correlational prediction from Gray & Singer, now testable at scale: on correct trials, feature-coding assemblies should partition into phase-coherent clusters matching the true objects; on illusory-conjunction trials the phase partition should match the erroneous binding. Second, the causal test: use closed-loop stimulation to force two assemblies belonging to different objects into a common gamma phase. Binding-by-synchrony predicts a specific, novel behavioral effect — an induced illusory conjunction — that no rate-based account predicts. Anti-phase entrainment of same-object assemblies should degrade conjunction reports without changing detection.

Keep from the 1989 paper: the focus on stimulus-specific, intracortically generated gamma (their LGN control was elegant and holds up), and the insistence that the phase relations, not just power, carry the content. Replace: anesthesia with behavior, pairwise cross-correlograms with population phase-partition statistics, and observation with closed-loop causal control.

The AI side of the mirror

What makes this worth an essay in 2026 is that the binding problem re-emerged in machine learning, independently and with the same shape. Object-centric models (slot attention and descendants) bind features to objects via discrete slots — a spatial, not temporal, solution, and one that struggles with softness, part-whole hierarchy, and open-ended object counts. A parallel line takes von der Malsburg literally: Reichert and Serre (2013) showed complex-valued deep networks where phase carries grouping; Löwe, Locatello et al.'s complex autoencoders and "rotating features" (2022–2023) demonstrated unsupervised object discovery where each neuron's activation is a magnitude (feature presence) times a phase (object identity) — binding by synchrony in the exact algebraic sense, minus the spikes. Spiking neuromorphic hardware makes the temporal version cheap to implement natively, since coincidence detection is the free operation.

This creates a rare two-way bet. If the neuroscience experiment vindicates synchrony-binding, it supplies a biologically grounded prior for phase-based object-centric architectures and a reason to prefer them on neuromorphic substrates. If it refutes it — showing binding survives with scrambled phase — that too is informative: it would push both fields toward the alternatives (dedicated conjunction populations, attention-as-serial-binding, or high-dimensional vector binding à la Smolensky/holographic representations, which is arguably what transformers do with key-query products).

Status and open questions

Partially vindicated, never decisively tested. Stimulus-specific gamma: replicated hundreds of times, mechanism (PING) understood. Long-range zero-lag synchrony tracking Gestalt properties: replicated in the 1990s, then complicated by negative results. Causal role in perceptual binding: still untested at the required specificity — existing 40 Hz optogenetic and sensory-flicker entrainment studies (including the Alzheimer's-related GENUS work) manipulate gamma globally, not the phase relations between specific assemblies, which is where the hypothesis lives. That experiment is now feasible and, to my knowledge, has not been done.

The deepest open question is readout: even if phase carries binding information, what downstream circuit uses it, and how does a phase code interface with the rate codes that demonstrably drive behavior? Any serious revival has to answer this, not just demonstrate correlated oscillations one more time.

Where to read it

The paper is at doi.org/10.1073/pnas.86.5.1698 (bibliographic details verified). Read alongside: von der Malsburg's 1981 correlation theory report (the theoretical source), Gray, König, Engel & Singer, Nature 1989 (the inter-columnar result that carries the binding claim), Shadlen & Movshon 1999 in Neuron (the sharpest critique, paired with Singer's reply in the same issue), Fries 2015 on communication through coherence (the mutated descendant), and Löwe et al.'s rotating-features papers (the idea reborn in ML, apparently working).