Extensive sampling of neural activity during rich cognitive phenomena is critical for robust understanding of brain function. We present the Natural Scenes Dataset (NSD), in which high-resolution fMRI responses to tens of thousands of richly annotated natural scenes are measured while participants perform a continuous recognition task. To optimize data quality, we develop and apply novel estimation and denoising techniques. Simple visual inspections of the NSD data reveal clear representational transformations along the ventral visual pathway. Further exemplifying the inferential power of the dataset, we use NSD to build and train deep neural network models that predict brain activity more accurately than state-of-the-art models from computer vision. NSD also includes substantial resting-state and diffusion data, enabling network neuroscience perspectives to constrain and enhance models of perception and memory. Given its unprecedented scale, quality, and breadth, NSD opens new avenues of inquiry in cognitive and computational neuroscience..

Advances in artificial intelligence have inspired a paradigm shift in human neuroscience, yielding large-scale functional magnetic resonance imaging (fMRI) datasets that provide high-resolution brain responses to thousands of naturalistic visual stimuli. Because such experiments necessarily involve brief stimulus durations and few repetitions of each stimulus, achieving sufficient signal-to-noise ratio can be a major challenge. We address this challenge by introducing GLMsingle, a scalable, user-friendly toolbox available in MATLAB and Python that enables accurate estimation of single-trial fMRI responses (glmsingle.org). Requiring only fMRI time-series data and a design matrix as inputs, GLMsingle integrates three techniques for improving the accuracy of trial-wise general linear model (GLM) beta estimates. First, for each voxel, a custom hemodynamic response function (HRF) is identified from a library of candidate functions. Second, cross-validation is used to derive a set of noise regressors from voxels unrelated to the experiment. Third, to improve the stability of beta estimates for closely spaced trials, betas are regularized on a voxel-wise basis using ridge regression. Applying GLMsingle to the Natural Scenes Dataset and BOLD5000, we find that GLMsingle substantially improves the reliability of beta estimates across visually-responsive cortex in all subjects. Comparable improvements in reliability are also observed in a smaller-scale auditory dataset from the StudyForrest experiment. These improvements translate into tangible benefits for higher-level analyses relevant to systems and cognitive neuroscience. We demonstrate that GLMsingle: (i) helps decorrelate response estimates between trials nearby in time; (ii) enhances representational similarity between subjects within and across datasets; and (iii) boosts one-versus-many decoding of visual stimuli. GLMsingle is a publicly available tool that can significantly improve the quality of past, present, and future neuroimaging datasets sampling brain activity across many experimental conditions.

The full neural circuits of conscious perception remain unknown. Using a visual perception task, we directly recorded a subcortical thalamic awareness potential (TAP). We also developed a unique paradigm to classify perceived versus not perceived stimuli using eye measurements to remove confounding signals related to reporting on conscious experiences. Using fMRI, we discovered three major brain networks driving conscious visual perception independent of report: first, increases in signal detection regions in visual, fusiform cortex, and frontal eye fields; and in arousal/salience networks involving midbrain, thalamus, nucleus accumbens, anterior cingulate, and anterior insula; second, increases in frontoparietal attention and executive control networks and in the cerebellum; finally, decreases in the default mode network. These results were largely maintained after excluding eye movement-based fMRI changes. Our findings provide evidence that the neurophysiology of consciousness is complex even without overt report, involving multiple cortical and subcortical networks overlapping in space and time.

Rapid simultaneous advances in machine vision and cognitive neuroimaging present an unparalleled opportunity to (re)assess the current state of artificial models of the human visual system. Here, we perform a large-scale benchmarking analysis of 85 modern deep neural network models (e.g. CLIP, BarlowTwins, Mask-RCNN) to characterize with robust statistical power how differences in architecture and training task contribute to the prediction of human fMRI activity across 16 distinct regions of the human visual system. We find: one, that even stark architectural differences (e.g. the absence of convolution in transformers and MLP-mixers) have very little consequence in emergent fits to brain data; two, that differences in task have clear effects--with categorization and self-supervised models showing relatively stronger brain predictivity across the board; three, that feature reweighting leads to substantial improvements in brain predictivity, without overfitting -- yielding model-to-brain regression weights that generalize at the same level of predictivity to brain responses over 1000s of new images. Broadly, this work presents a lay-of-the-land for the emergent correspondences between the feature spaces of modern deep neural network models and the representational structure inherent to the human visual system.