The document discusses a deep learning framework for image set classification that does not assume images lie on a certain geometric surface like existing methods. It introduces a Template Deep Reconstruction Model (TDRM) initialized through unsupervised pre-training with Gaussian Restricted Boltzmann Machines (GRBMs). Class-specific DRMs are then separately trained for each class and used to classify images based on minimum reconstruction error, outperforming existing state-of-the-art methods.