Yukiyasu Kamitani

This paper decodes 3D object structure from human visual cortex activity using a pretrained 3D autoencoder and shows cross-cue generalization across depth cues and object categories.

This paper demonstrates that closed-loop brain-computer interfaces can retrieve intended images using a CLIP latent space, achieving over 80% accuracy, while failing with AlexNet.

This study reconstructs visual illusory experiences from brain activity using a decoding technique and deep neural networks, showing that reconstructions match illusory percepts.

This study decodes multiple navigational information types from DG neuron activity in mice, showing independent encoding of position, speed, motion direction, and future location.

This study develops a neural code converter to align brain activity patterns across individuals, enabling inter-individual visual image reconstruction while preserving hierarchical representations.

The paper introduces a brain hierarchy score to measure how brain-like deep neural networks are, finding that high-performance networks are not necessarily brain-like.

This study reconstructs visual images from brain activity during attention tasks, showing reconstructions align with attended images rather than unattended ones.

The paper identifies neural substrates of a high-dimensional emotion space using fMRI responses to 2185 videos, showing emotion categories better predict brain activity than affective dimensions.

This paper trains a deep neural network end-to-end to reconstruct images from human brain fMRI data, showing improved accuracy with more training data.

The paper decodes deep neural network features from human brain activity using fMRI data and presents a dataset showing correlations in decoding accuracy across subjects.

The paper investigates how top-down and bottom-up signals interact during visual object recognition, finding that top-down signals have limited impact on lower-level processing but stronger influence on higher cognitive areas.

This paper introduces a method to reconstruct images from human brain activity by optimizing pixel values to match deep neural network features across multiple layers, showing it generates new images rather than matching exemplars.

This study introduces sparse ordinal logistic regression (SOLR) for decoding ordinal variables from brain data, showing improved performance over existing methods.

This study uses DNN representations to show neural activity shifts toward non-blurred images when viewing blurred images, suggesting top-down sharpening of visual representations.

The paper decodes object position from fMRI data across visual areas, finding similar prediction accuracy despite differences in receptive field size.