Alessandro Gifford, Kendrick Kay

OOD generalization tests reveal differences between brain models that are not detected in-distribution. The degree of OOD (quantified as the test data distance from the training data) is predictive of the magnitude of model failures.

James V. Haxby, Peter J. Ramadge

Response-tuning functions for visual population codes are common across individuals. 35 response basis functions capture fine-grained distinctions among representations. The common model space greatly improves between-subject classification of fMRI data.

Alexander G. Huth, Jack L. Gallant

Ventral visual cortex encoded a graded biological taxonomy, separating primates, birds, and insects. This neural similarity structure matched human judgments of biological similarity. Early visual cortex mainly reflected low-level visual features, not taxonomy.

David M. Watson, Tom Hartley

Clustered scene images based on visual features, ignoring semantics. Each cluster elicited distinct response patterns in PPA. Similarity of PPA responses across clusters was predicted by image properties, not semantic content. PPA responses were also predicted by human perceptual similarity judgments.

Meenakshi Khosla, Nancy Kanwisher

Finds food-selective component in the ventral visual cortex. Not explained by color, shape, or texture.

Colin Conwell, Talia Konkle

Many models achieve similarly high brain predictivity, despite clear variation in their underlying representations.

J. Swaroop Guntupalli, James V. Haxby

A shared high-dimensional representational space could be learned across participants from ventral temporal responses. Projecting individual brains into this common space improved between-subject pattern classification. The common-space model generalized to new response vectors not used during model fitting.

Talia Konkle, Aude Oliva

Find large-scale organization of big and small object responses across the cortex. Regions are tolerant to retinal size changes and activate during mental imagery.

Margret Henderson, Leila Wehbe

Encoding model based on texture statistics shows how feature selectivity evolves from early to higher cortex.

Eshed Margalit, Daniel Yamins

Proposes a topographic deep network (TDANN) balancing self-supervised objectives with spatial smoothness to match cortical maps. Predicts function and spatial structure in early and higher visual cortex.

Thomas Naselaris, Jack Gallant

Demonstrate a Bayesian decoder that uses fMRI signals from early and anterior visual areas to reconstruct complex natural images.

Aria Wang, Leila Wehbe

Models trained with language supervision (vision–language alignment) have improved general-purpose visual representations with respect to predicting brain responses.

Andrew Luo, Leila Wehbe

Introduces BrainDiVE, which synthesizes images to maximally activate a brain region. Can generate preferred images with appropriate semantic specificity for well-known category-selective regions. Also identified novel functional subdivisions within ROIs.

Ghislain St-Yves, Thomas Naselaris

Hierarchical representations in DNNs are not necessary for accurate prediction of V1–V4 responses.

Diego Garcia Cerdas, Iris Groen

Diffusion-based image manipulations can selectively increase or decrease predicted activity in specific visual cortical regions from the same reference image. The resulting changes reveal which visuo-semantic features drive regional selectivity and align with known category preferences.

Marieke Mur, Nikolaus Kriegeskorte

FFA and PPA show almost perfect categorical ranking: every preferred-category image elicits higher response than any nonpreferred. Responses are graded within and outside the preferred category. FFA shows more graded tuning within-category; PPA shows a stronger category-step discontinuity.

Ian Pennock, Jenny Bosten

Two ventral food streams begin in V4 and diverge medially and laterally of the FFA. Color-biased regions in the ventral visual pathway are food selective.

Daniel D. Leeds, Michael J. Tarr

All tested models significantly explained some neural data. SIFT features (scale-invariant interest points) best accounted for representations in intermediate ventral areas.

Apurva Ratan Murty, Jim DiCarlo

Developed ANN-based encoding models that accurately predict responses in FFA, PPA, EBA. Provides precise, computable evidence for domain-specificity of each area. Models outperform descriptive models.

Marius V. Peelen, Alfonso Caramazza

Anterior temporal cortex patterns generalized across visually different exemplars of the same object concept. Posterior ventral visual regions were more sensitive to visual form differences. Cross-exemplar object decoding was strongest in anterior temporal cortex.

Jacob Prince, Talia Konkle

Category-selective tuning naturally emerges for faces, bodies, scenes, and words in models trained with contrastive self-supervised objectives.

Atlas Kazemian, Michael F. Bonner

Without pretraining, CNN architectures have more cortex-aligned representations in the visual cortex, compared to other architectures. Alignment emerges through compression in the spatial domain and expansion in the feature domain.

Thomas Naselaris, Jack L. Gallant

The top principal component of voxel tuning corresponds to the animate–inanimate axis (50–60% of tuning variance). Anatomical dissociation: voxels preferring animate objects are anterior to retinotopic areas, flanked by inanimate-preferring voxels.

Xiaomin Yue, Leslie Ungerleider

A network of cortical patches with curvature response preferences was observed in the human ventral visual stream.

Guohua Shen, Yukiyasu Kamitani

Reconstructed perceptual and subjective images from fMRI data.

Umut Güçlü, Marcel van Gerven

CNN layers map onto visual hierarchy.

Soojin Park, Aude Oliva

PPA and LOC process scenes in complementary ways: PPA codes spatial layout (scene geometry), LOC codes content (objects).

Iris Groen, Chris Baker

Similarity of cortical responses to scene images in scene-selective areas was uniquely explained by mid- and high-level DNN features only. An object label model did not contribute uniquely.

Oliver Contier, Martin Hebart

Found dimensional coding in the visual cortex, aligning with behaviourally-derived dimensions.

Hans P. Op de Beeck, Johan Wagemans

Face and object information could be decoded from broad distributed patterns across ventral temporal cortex, not only from peak-selective voxels. Removing the most selective voxels reduced but did not abolish category decoding.

Yue Wang, Shuo Wang

Show semantic coding throughout the MTL.

Matteo Ferrante, Rufin VanRullen

Compositional processes in neural representations lead to predictable perceptual outcomes, as interpreted by decoding models.

Mark D. Lescroart, Jack Gallant

Fourier and distance models each predicted most of the variance that the category model did. Much variance is shared. No single feature type uniquely dominates scene representations but multiple factors contribute.

Navve Wasserman, Michal Irani

Compute functional brain-to-brain transformations without any shared data (stimuli). Improve image-to-fMRI encoding of subjects scanned on older low-resolution 3T fMRI datasets, by using a new high-resolution 7T fMRI dataset.

Tolga Çukur, Jack Gallant

PPA response similarity was best predicted by high-level scene properties, especially real-world size and spatial layout. OPA similarity was more strongly related to low- and mid-level visual features. CNN feature spaces explained variance in all three scene-selective regions.

Ana Torralbo, Diane M. Beck

Humans categorize good exemplars of scene categories faster/more accurately than bad exemplars. MVPA decoding more accurate for good vs bad exemplars. Good exemplars produce clearer neural patterns, not stronger overall activation.

Adrien Doerig, Ian Charest

Strong alignment between visual cortex and LLM derived representations.

Yu Takagi, Shinji Nishimoto

Latent diffusion model can reconstruct high-resolution images with high fidelity in straightforward fashion, without the need for any additional training and fine-tuning of complex deep-learning models.

Zijin Gu, Amy Kuceyeski

Images predicted to achieve maximal activations evoke higher responses than average activation images. Anterior temporal lobe face area and fusiform body area 1 had higher activation in response to maximal synthetic images. Synthetic images from a personalized encoding model elicited higher responses than group-level models.

Marieke Mur, Nikolaus Kriegeskorte

Objects with similar fMRI-IT patterns are judged similar by humans. Human IT explains judgments better than early visual areas or computational models. Judgments include IT-like clusters but add distinctions: human vs. non-human, natural vs. man-made. hIT is more similar to monkey IT than to human judgments.

Mark D. Lescroart, Jack Gallant

3D-structure encoding model explains unique response variance in scene areas. Voxels encode distances and orientations of large 3D surfaces. Main dimensions are distance and openness. The model can reconstruct 3D scene backgrounds from brain activity.

Changde Du, Huiguang He

Strong alignment between model object embeddings and neural activity patterns in brain regions such as the extrastriate body area, parahippocampal place area, retrosplenial cortex and fusiform face area.

Kendrick Kay, Jack Gallant

Decoded natural images from fMRI using a decoding method based on quantitative receptive-field models.

Hojin Jang, Frank Tong

CNNs trained with both clear and blurry images better predict human and neural responses across varied viewing conditions than standard CNNs. Blur-trained CNNs show increased sensitivity to global shape and robustness to noise and degradation.

Alessandro Gifford, Radoslaw Cichy

Develop a method to generate in silico fMRI responses using encoding models. Identify images that align or disentangle responses between pairs of visual areas. They find a network-level configuration of representational relationships across visual cortex.

Haibao Wang, Yukiyasu Kamitani

Develop an inter-individual functional alignment technique that does not rely on shared stimuli between subjects. The converted brain activity can be accurately decoded using the target's pre-trained decoders, producing high-quality visual image reconstructions that rival within-individual decoding, even with data across different sites and limited training samples.

Emilie L. Josephs, Talia Konkle

Dissociate objects, reachspaces, and scenes in the brain.

Margaret Henderson, Leila Wehbe

Voxels in category-selective visual regions exhibit systematic biases in their feature and spatial selectivity, which are consistent with their hypothesized roles in category processing.

Dirk B. Walther, Diane M. Beck

Decoded scene category from PPA, RSC, LOC, and V1.

Furkan Ozcelik, Rufin VanRullen

Trained a model to reconstruct both low-level layout and more semantic/high-level aspects of complex natural scenes.

Zvi Roth, Elisha Merriam

Find reliable coarse-scale orientation tuning signatures in V1 in response to scene images.

Dustin Stansbury, Jack Gallant

Brain activity encodes scene categories that reflect real-world object statistics. Scene categories and individual objects can be decoded from measured brain activity.

Hongru Zhu, Daniel Kersten

Maps 2D vs 3D body-pose feature along visual pathways.

Marius V. Peelen, Sabine Kastner

fMRI activity patterns in object-selective cortex distinguished scenes containing people from scenes containing cars. This category discrimination was present even when the objects appeared outside the attended location. Early visual cortex showed weaker category discrimination than object-selective cortex.

Aria Wang, Leila Wehbe

Mapped representations from CNNs trained with different computer vision tasks to the brain. Features from 3D tasks explain greater variance compared to 2D tasks across the whole brain. Tasks with higher transferability make similar predictions for brain responses from different regions.

Zijin Gu, Amy Kuceyeski

Detect and amplify differences in regional and individual human brain response patterns to visual stimuli. Create synthetic images predicted to achieve regional response patterns not achievable by the best-matching natural images.

Zijin Gu, Amy Kuceyeski

Ensemble approach to create encoding models for novel individuals with relatively little data by modeling each subject's predicted response vector as a linear combination of the other subjects' predicted response vectors. The ensemble encoding models preserve patterns of inter-individual differences in the image-response relationship.

Dwight J. Kravitz, Chris Baker

PPA activity patterns grouped scenes by openness (open vs closed). Early visual cortex grouped scenes more strongly by relative distance (near vs far). Scene content was weakly decoded compared with spatial properties.

Nicholas J. Sexton, Bradley C. Love

Found that all regions along the ventral visual stream best corresponded with later model layers. Tested by replacing model activity with brain activity to test DCNN's object recognition decision.

Paul Scotti, Tanishq Abraham

MindEye can retrieve the exact original image even among highly similar candidates, indicating that its brain embeddings retain fine-grained image-specific information.

Zvi Roth, Elisha Merriam

Reports evidence consistent with representational drift in some measures while emphasizing that relative structure/similarity can remain more stable than raw patterns (supporting relational-coding views).

Michael J. Arcaro, Sabine Kastner

Multiple retinotopic visual field maps in ventral temporal cortex. Category-selective regions were systematically positioned relative to these maps. Scene-selective cortex was biased toward peripheral representations, whereas face-selective cortex was more foveally biased.

Matteo Ferrante, Nicola Toschi

Semantic classification and image retrieval: correct evaluation in over 80% of the test set.

Nidhi Jain, Leila Wehbe

Identifies two food-selective regions in ventral visual cortex.

Guy Gaziv, Michal Irani

1000-way semantic classification of fMRI data using cycle-consistent DNN approach. Combines reconstruction with perceptual losses, enabling classification and reconstruction of never-before-seen images.

Pablo Marcos-Manchón, Lluís Fuentemilla

Applied RSA to fMRI data and compared with hierarchical vision and language models. Identify two distinct processing routes: ventromedial (scene layout) and lateral occipitotemporal (animate content). Vision models correspond to both routes, while language models align mostly with the lateral pathway.

Marcie L. King, Chris Baker

fMRI (vTC) representation shows clear separation of faces/bodies from other categories. Behavior RDM best matches top network layer; fMRI RDM best matches mid-level layers.

Kevin S. Weiner, Kalanit Grill-Spector

Face- and limb-selective activations formed alternating, minimally overlapping clusters in lateral VTC. This spatial arrangement was consistent across sessions and experiments within subjects. Face information was strongest in lateral VTC voxels with category selectivity, but remained present in weakly selective voxels.

Ruogu Lin, Leila Wehbe

The stacking encoding model can predict held-out brain activity better than individual encoding models. Weights are readily interpretable, showing the importance of each feature space for predicting a voxel. Stacked variance partitioning allows targeted questions about feature space similarity even in the presence of correlations.

Ian Pennock, Jenny Bosten

Co-occurrences between color and other properties of natural scenes allow to identify and visualize functionally distinct brain regions.

Caterina Magri, Alfonso Caramazza

Occipitotemporal cortex shows large-scale organization by real-world size for both motor-relevant and irrelevant objects. Contrasts consistent with natural covariation yield stronger topographies. Neural responses to object size also reflect action-related properties.

Rebecca F. Schwarzlose, Nancy Kanwisher

Object category and stimulus position could both be decoded from object-selective cortex patterns. Category information was stronger in more anterior regions, whereas position information was stronger in posterior regions.

Frederik S. Kamps, Daniel D. Dilks

Face and scene category information was strongest in classically selective ventral temporal regions. Decoding remained above chance in surrounding weakly selective voxels. Distributed ventral temporal patterns supported category discrimination beyond ROI peaks.

Andrew C. Connolly, James V. Haxby

Ventral visual patterns encode biological class: activity patterns cluster by species class (insects, birds, primates). The neural RDM correlates with behavioral similarity judgments. Representational continuum: primates vs. insects ends of animate hierarchy.

Markus Badwal, Martin Hebart

Ventral cortex patterns are dominated by image-specific effects; category-level effects are weaker. Weaker but reliable effects across different exemplars of the same object remain beyond low/mid-level visual features. Some of these residual effects align with high-level semantic similarity.

Maya Yablonski, Jason Yeatman

VWFA-1 is primarily correlated with bilateral visual regions. VWFA-2 is more strongly correlated with language regions in the frontal and lateral parietal lobes, particularly the bilateral inferior frontal gyrus. These patterns do not generalize to adjacent face-selective regions.

Matteo Ferrante, Nicola Toschi

Found that cross-subject brain decoding is possible, even with a small subset of the dataset. Ridge regression emerged as the best method for functional alignment in fine-grained information decoding.

Talia Konkle, Alfonso Caramazza

Animacy and real-world size organize ventral visual cortex.

Zirui Chen, Michael Bonner

Found that diverse networks learn to represent natural images using a shared set of latent dimensions, despite having highly distinct designs. The most brain-aligned representations are those that are universal and independent of a network's specific characteristics.

Raj Magesh Gauthaman, Michael F. Bonner

Neural population activity in human visual cortex follows a scale-free (power-law) spectrum across a large range of latent dimensions. Consistent across multiple visual regions and shared across participants.

Wei Huang, Jingpeng Li

Compare semantic decoding from ventral vs. dorsal stream. Dorsal stream shows higher decoding for verbs and nouns with motion attributes.

Umut Güçlü, Marcel van Gerven

Learned hierarchical sparse features (unsupervised) from natural image patches. These features better predict human early visual cortex (V1-V2) fMRI responses than standard Gabor models. Approach identifies meaningful visual features without label supervision.

Kshitij Dwivedi, Gemma Roig

Found a structured mapping between DNN tasks and brain regions. Low-level visual tasks mapped onto early brain regions, 3D scene perception tasks onto dorsal stream, semantic tasks onto ventral stream.

Ghislain St-Yves, Thomas Naselaris

Variation in classification accuracy across human visual cortex is driven by the distance between manifold centers. Variation across early and mid DNN layers is driven by an increase in the effective number of manifold dimensions. Signal and Dimensionality are strongly, negatively correlated.

Yoichi Miyawaki, Yukiyasu Kamitani

Multiscale local decoders to reconstruct 10×10 checkerboard images from fMRI. Achieved accurate single-trial reconstruction without assuming image priors. Reconstruction allowed identification among millions of potential images.

Linda Henriksson, Nikolaus Kriegeskorte

Intrinsic cortical dynamics make representational geometry appear similar across visual areas. Sharing the same trial artificially increased inter-area similarity. When accounting for intrinsic fluctuations, early visual areas fit a Gabor model and higher areas fit animate-inanimate distinctions.