Martin Schrimpf

This study shows that diverse perceptual representations in visual pathways arise from a single objective of understanding world dynamics, using deep neural network models to predict human neural data.

The paper tests if topographic deep ANN models of the primate ventral stream can predict perceptual effects of direct IT cortical interventions using perturbation modules and benchmarking against experimental results.

The paper uses a GPT-based model to predict and control brain responses to language in humans using fMRI data.

The paper shows that language models trained on developmentally realistic data can predict human brain responses to language, achieving near-maximal performance with 100 million words of training.

The paper argues that mapping models in cognitive neuroscience should be chosen based on research goals rather than being strictly linear or nonlinear, and introduces complexity metrics for evaluation.

This study improves AI models by aligning their representations with macaque brain data, enhancing behavioral alignment and adversarial robustness.

This paper shows that powerful transformer models predict neural and behavioral responses during language processing across datasets and modalities, supporting predictive processing as a core mechanism.

The paper builds and evaluates hierarchical neural network models to approximate primate V1 neurons and shows their alignment with human object recognition behavior.

This paper shows that unsupervised neural networks can model primate ventral visual stream responses as well as supervised methods, with brain-like representations from limited data.

A fixed-weight, biologically parameterised V1 front-end gives standard image classifiers adversarial and corruption robustness at no extra training cost; its simulated neuronal noise drives most of the adversarial gain.

This paper reduces the number of supervised synaptic updates needed to model the primate ventral visual stream, achieving 80% of the adult match with significantly fewer updates.

Among dozens of ImageNet-trained networks scored against macaque V4/IT recordings and human behavior, DenseNet-169, CORnet-S and ResNet-101 rank highest, and brain similarity stops tracking accuracy above 70% top-1.

The paper evaluates hyperparameter configurations for predicting human brain responses using convolutional neural networks and identifies optimal settings for fMRI and MEG data.

The paper finds that model performance on monkey IT data predicts performance on human IT data, outperforming ImageNet-based predictors.

CORnet-S, a four-area recurrent network with areas pre-mapped to V1, V2, V4 and inferior temporal cortex, ranks among top Brain-Score models while having a shorter longest convolutional path than competitors.