Derin öğrenme yöntemlerini kullanarak çok denekli beyin okuma.

In this study, a new method is proposed for analyzing and classifying images obtained by functional magnetic resonance imaging (fMRI) from multiple subjects. Considering the multi level structure of the brain and success of deep learning architectures on extracting hierarchical representations from raw data, these architectures are used in this thesis. Initially, the S500 data set collected in the scope of Human Connectome Project (HCP) is used to train formed deep neural networks in an unsupervised fashion. Then, pre-trained networks are utilized for two different multi-subject brain decoding tasks. Goal of these tasks is discriminating the cognitive state of a subject, using the fMRI data of other subjects. In the first task, brain decoding is performed by fine-tuning the pre-trained neural networks with the label information included in the S500 data set. In the second task, pre-trained networks are used to obtain hierarchical representation for object recognition data set to transfer the information between fMRI experiments. Obtained results show us that deep neural networks are more successful than traditional machine learning algorithms on multi-subject brain decoding tasks and experiment-independent representations can be obtained with deep neural networks better than factor models used in the literature. Besides, regions activated for different cognitive states of S500 data set are visualized by implementing a saliency analysis over trained deep neural networks.M.S. - Master of Scienc