Classification errors (%) from the (a) softmax and (b) SVM classifiers using the hidden layer output of the DBN without DNN fine-tuning (white box), DNN without DBN pretraining (light gray box), and DNN with DBN pretraining (dark gray box). The performance of the hidden activation function using tanh and ReLU is also compared. The dashed lines indicate the classification performance using fMRI volume (from the input layer) as an input to each classifier. tanh, hyperbolic tangent; ReLU, rectified linear unit; SVM, support vector machine; DBN, deep belief network; DNN, deep neural network.