Nikolaus Kriegeskorte

This study shows that lateral recurrence in neural networks enhances identity separability in trained domains through dynamic refinement of representational geometry.

This paper introduces blur and retinal masks as self-supervision signals for vision transformers, showing improved robustness and generalization in visual representation learning.

The paper demonstrates that attentional control improves energy efficiency in a primate visual processing model by dynamically allocating energy to task-relevant features and locations.

The paper develops a factorized model of temporal abstraction to study how abstract knowledge is learned and represented in the brain during a decision task, showing that hippocampus, entorhinal cortex, and orbitofrontal cortex support this process.

This paper introduces framed RSA, a method that evaluates neural models by incorporating both geometric patterns and population-mean response preferences, improving model selection accuracy in brain-region and network identification.

The paper introduces an open-source Python toolbox for representational similarity analysis, providing tools to evaluate neural representations and models.

The paper investigates when measured representational distances in neural data reflect true neural geometry and proposes corrections for distortions.

This paper shows that mirror-symmetric viewpoint tuning emerges in convolutional neural networks trained on object recognition tasks, even without face-specific training, and explains its mechanism through reflection-equivariant representations.

This paper investigates whether iterative convergent computation improves ResNets but finds it does not enhance classification accuracy on standard or challenging visual tasks.

This paper uses a transformer model to predict human brain activity from images, achieving a challenge score of 63.5229 and demonstrating the effectiveness of self-supervised transformer features.

TorchLens extracts and visualizes hidden activations of PyTorch models, providing full computational graph records and validation for analysis.

The paper shows that visuo-semantic features explain more variance in higher-level ventral stream neural dynamics than deep neural networks, suggesting complementary roles in object recognition.

The paper investigates five animacy-related object dimensions in brain and behavior using fMRI, EEG, and similarity judgments, finding they explain variance in both brain representations and behavior.

The paper shows that Euclidean distance in a 3D morphable face model predicts human face dissimilarity and identity judgements as well as deep neural networks.

The paper investigates how conceptual categories are represented in the brain by integrating sensorimotor and distributional semantics, finding that combined semantic vectors best explain activation patterns in specific brain regions.

The paper compares nine DNN architectures to explain human inferior temporal cortex representational geometry, finding trained models outperform untrained ones and all architectures achieve similar performance after training and fitting.

The paper shows that viewing ambiguous social interactions increases connectivity between frontal and temporal brain regions in monkeys.

The paper uses fMRI and representational similarity analysis to show that FFA encodes higher-level face identity information, while OFA encodes low-level image properties.

This study shows that small changes in initial conditions of deep neural networks lead to significant differences in their representations, even when classification performance is similar.

This paper shows recurrent neural networks outperform feedforward models in visual recognition tasks, enabling flexible speed-accuracy trade-offs and better prediction of human reaction times.

The paper evaluates statistical tests for detecting discriminability between brain response patterns to exemplars and recommends more powerful methods over the standard t test.

The paper investigates rER BOLD fMRI in NHP, finding that while individual elements work well, their combination is challenging due to noise and motion artifacts.

The paper corrects an error in noise ceiling calculations for human data from a previous study, showing the best model does not significantly fall below the noise ceiling lower bound.

This paper introduces metrics to analyze linear transformations between multivariate fMRI patterns in different brain regions, showing they explain significant variance in output regions.

The paper investigates whether the brain has modality-general person-identity representations and finds evidence for such representations in the right posterior superior temporal sulcus using fMRI and RSA.

The paper shows that the human occipital place area (OPA) encodes scene layout invariant to surface texture within 100 ms, contrasting with V1 and PPA.

This paper discusses how deep neural networks can model neural responses and cognitive functions, highlighting their accuracy and potential in computational neuroscience.

GLMdenoise improves replicability and classification accuracy in multivariate pattern analysis of fMRI data by reducing noise in GLM estimates.

This study shows that prospective motion correction improves multi-voxel pattern decoding in fMRI at higher resolutions.

The paper shows that emotional voice stimuli are initially categorized in the brain and later refined into continuous dimensions.