Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. Here, to evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs indicated high confidence in their predictions, their responses were more likely to be correct, which presages a future where LLMs assist humans in making discoveries. Our approach is not neuroscience specific and is transferable to other knowledge-intensive endeavours.

Current models of how animals estimate motion involve correlations between pairs of points in space and time. Here the authors show that both fly and human visual systems can encode the direction and contrast polarity of moving edges using three-point correlations, and that this enhances motion estimation accuracy. Sighted animals extract motion information from visual scenes by processing spatiotemporal patterns of light falling on the retina. The dominant models for motion estimation exploit intensity correlations only between pairs of points in space and time. Moving natural scenes, however, contain more complex correlations. We found that fly and human visual systems encode the combined direction and contrast polarity of moving edges using triple correlations that enhance motion estimation in natural environments. Both species extracted triple correlations with neural substrates tuned for light or dark edges, and sensitivity to specific triple correlations was retained even as light and dark edge motion signals were combined. Thus, both species separately process light and dark image contrasts to capture motion signatures that can improve estimation accuracy. This convergence argues that statistical structures in natural scenes have greatly affected visual processing, driving a common computational strategy over 500 million years of evolution.

We used source imaging of visual evoked potentials to measure neural population responses over a wide range of horizontal disparities (0.5–64 arcmin). The stimulus was a central disk that moved back and forth across the fixation plane at 2 Hz, surrounded either by binocularly uncorrelated dots (disparity noise) or by correlated dots presented in the fixation plane. Both disk and surround were composed of dynamic random dots to remove coherent monocular information. Disparity tuning was measured in five visual regions of interest (ROIs) [V1, human middle temporal area (hMT+), V4, lateral occipital complex (LOC), and V3A], defined in separate functional magnetic resonance imaging scans. The disparity tuning functions peaked between 2 and 16 arcmin for both types of surround in each ROI. Disparity tuning in the V1 ROI was unaffected by the type of surround, but surround correlation altered both the amplitude and phase of the disparity responses in the other ROIs. Response amplitude increased when the disk was in front of the surround in the V3A and LOC ROIs, indicating that these areas encode figure–ground relationships and object convexity. The correlated surround produced a consistent phase lag at the second harmonic in the hMT+ and V4 ROIs without a change in amplitude, while in the V3A ROI, both phase and amplitude effects were observed. Sensitivity to disparity context is thus widespread in visual cortex, but the dynamics of these contextual interactions differ across regions.

Studying directional selectivity using neuroimaging in humans is difficult because the resolution is insufficient to directly access directionally selective activity. Here we used motion adaptation of the steady-state visual evoked potential (SSVEP) and source imaging in the frequency domain to detect brain areas that contain direction-selective cells. This study uses a definitive electrophysiological marker for direction-specific adaptation in the SSVEP to localize cortical areas that are direction selective. It has been shown previously that an oscillating stimulus produces an SSVEP response that is dominated by even harmonics of the stimulus frequency. This pattern of response is consistent with equal population responses to each direction of motion. Prolonged exposure to unidirectional motion induces an asymmetry in the population response that is consistent with adaptation of direction-selective cells. This asymmetry manifests itself in the presence of odd harmonic components after adaptation Critically, the feature that indicates the direction used for adaptation is the phase of the odd-harmonic responses. We recorded this signature of direction selectivity in a group of observers whose retinotopic visual areas had been defined from fMRI mapping. We find direction-specific responses throughout retinotopic cortex, with the largest effect in areas V1 (occipital pole) and V3/V3a (dorsal).