End-to-end deep image reconstruction from human brain activity

Shen, Dwivedi, Majima, Horikawa, and Kamitani (2019) End-to-end deep image reconstruction from human brain activity. Front. Comput. Neurosci. (bioRxiv preprint). Here we provide pre-trained deep artificial neural networks (DNN) for reconstructing images from human brain activity. The pre-trained DNNs are composed of generator networks and discriminator networks. The generator networks were trained to reconstruct the stimulus images from human brain fMRI activity. The discriminator networks were trained to distinguish between the reconstructed images and the original stimulus images. The pre-trained DNNs of 3 human subjects are shared here: - trainedmodel_sub-01.zip - trainedmodel_sub-02.zip - trainedmodel_sub-03.zip These pre-trained DNNs were trained for 500,000 mini-batch iterations, with batch size of 64. In each of the zip files, there are 4 files: - generator.caffemodel: the Caffe caffemodel file for the trained generator - generator_test.prototxt: the Caffe prototxt file for the generator - discriminator.caffemodel: the Caffe caffemodel file for the trained discriminator - discriminator.prototxt: the Caffe prototxt file for the discriminator Demo code: GitHub