My PhD asked a simple question with a surprising answer: when we pay attention, do we take in the visual world continuously, or in rapid snapshots? Across several experiments, attention turned out to sample: about 8 times a second for a single object, and about 4 times a second for each of two things that compete for it, whether they compete in space, in features or between the two eyes.
PhD in Cognitive Neuroscience, Hebrew University of Jerusalem (2017–2023), supervised by Prof. Ayelet N. Landau. Current Biology 2019 · J. Cognitive Neuroscience 2023 · Trends in Cognitive Sciences 2026.
Visual information is organised in stages. Early on, the two eyes are still separate channels. Then objects are mapped by location in the visual field. Higher up, neurons care about features such as colour or motion. If attention samples rhythmically, does it do so at one stage only, or wherever two representations compete?
Three levels at which two representations can compete for attention: between the eyes, between locations, and between features.
A grating appears around fixation. A brief, salient reset (a flash of four dots) pulls attention to it and aligns everyone's ongoing rhythm. Then a faint target (a small contrast dip on the grating) appears at a random delay between 0.25 and 0.85 s, at one of six difficulty levels, and the participant reports whether they saw it. Plotting accuracy against delay gives the accuracy time course; its spectrum shows whether, and how fast, performance waxes and wanes.
A simulated observer, for illustration: real detection is noisy, so the rhythm only emerges after hundreds of trials. One object: a single grating; accuracy rises and falls about 8 times a second. Two competing objects: the paradigm of Landau & Fries (2012): two gratings, the reset flashes around one of them, and the target appears on either the same or the opposite grating. Each location is sampled about 4 times a second, in antiphase. The real data are below.
Why the reset matters. Faint lines: the rhythm on individual trials. Before the reset their phases are random and the average (bright line) is flat; the reset aligns them, so good and poor moments line up across trials and the rhythm becomes measurable.
One object: ~8 Hz. Group spectrum of the accuracy time course (n = 32), with a significant peak at 8 Hz (bar). Re et al. 2019, 2023; redrawn from the published figure.
The starting point was Landau & Fries (2012, Current Biology): with two gratings on either side of fixation and a reset on one of them, detection at each location fluctuated at about 4 Hz, and the two locations were sampled in alternation. Two competing objects are each sampled at half the single-object rate, as if a ~8 Hz attentional rhythm were shared between them.
Does the same happen when the two objects are not in different places? I overlapped two dot clouds of different colours in the same location, so they could only be told apart by their features, and reset attention to one of them. Detection in each cloud fluctuated rhythmically, peaking around 4 Hz: feature-based attention samples too.
Feature-based sampling (n = 25). Left: accuracy time courses for the two clouds. Right: their spectra, with significant power around 4 Hz (bars, p < 0.05). Re, Inbar, Richter & Landau, Current Biology 2019. Redrawn from the published figure.
The earliest level where two inputs compete is between the eyes: each eye feeds its own monocular channel before the signals merge in V1. To test it, I had to flash the reset and the target into one eye only, while the participant still saw one ordinary grating with both eyes.
A PROPixx projector runs at 480 frames per second, and consecutive frames take turns between the eyes. A circular polariser in front of the lens gives each frame the polarisation handedness of its eye; the screen preserves it, and polarised glasses let each lens pass only its own frames. Each eye therefore receives its own image stream, far too fast to see the alternation: the grating looks completely normal, but a flash or a target can be drawn into one eye's frames only.
Orange frames are polarised for the left eye, blue frames for the right. Panels on the right show what each eye actually receives.
A, binocular: reset and target in both eyes. B, opposite: reset in the right eye, target in the left. C, same: reset and target both in the right eye. Shown with a right-eye reset; the reset was also given to the left eye. The target, a small contrast dip at one of six levels, came 0.25 to 0.85 s after the reset.
Eye channels. Top: accuracy over time; bottom: spectra relative to the aperiodic background. Left: condition A (binocular), sampled near 7 Hz. Middle and right: the monocular conditions, one panel per reset eye, comparing a target in the same eye (C) with the opposite eye (B). With one reset eye, same-eye targets were sampled near 4 Hz (middle, teal), while the opposite eye dipped near 8 Hz (bar below zero). Significance against shuffled surrogates, corrected for multiple comparisons. Re, Karvat & Landau, J. Cognitive Neuroscience 2023. Redrawn from the published figure.
| Where | What competes | Sampling | Source |
|---|---|---|---|
| Space | nothing: a single object | ~8 Hz | Re et al. 2019, 2023 |
| Space | two locations | ~4 Hz | Landau & Fries 2012 |
| Features | two overlapping clouds | ~4 Hz | Re et al. 2019, Curr Biol |
| Eyes | the two eye channels | ~4 Hz | Re et al. 2023, JoCN |
Wherever two representations compete, at the level of the eyes, of locations or of features, each is sampled at about 4 Hz, in alternation. A single object is sampled at about 8 Hz. Our 2026 Trends in Cognitive Sciences paper builds this into a framework: attentional sampling is how the brain resolves competition along the visual hierarchy.
On rigour. Rhythms in behaviour are easy to over-claim, so every spectrum was tested against shuffled surrogate data with multiple-comparison correction, and we published a methods reply on how periodicity in behaviour should be tested (Re, Tosato, Fries & Landau, 2022).