MultiView Matching Based on a Unified Polynomial Pooling Model – In this paper, we propose a hierarchical approach for learning Bayesian networks with a large class of nonlinear dependencies. This model was inspired by Markov Decision Processes (MDPs) and allows us to learn the distribution of dependencies directly from the data. The goal is to efficiently incorporate knowledge to facilitate the learning process of the network and to extract useful information in the form of binary features, or labels. Although recent deep learning approaches use hierarchical inference, there is still a great need to learn the most informative parameters in such networks. In this work, we propose a new algorithm to learn the parameter distribution by extracting features from a low-dimensional manifold as input only. Our algorithm uses a Gaussian process prior and provides a low-dimensional projection into the manifold. We evaluate our method using synthetic data and in vivo data including human brain data, and observe positive evidence in terms of performance and reliability.
In this paper, we describe a deep learning (DL) framework for segmentation of the human hippocampus. The hippocampus is considered as a functional brain region that contains various sensory and motor functions. In this context, a neural network (NN) has received attention in recent years. However, the classification of the hippocampal region by an NN does not provide a good performance for the task, because of the limited number of labeled examples. Therefore, we propose an DL framework that takes a dataset of hippocampal data and models the information in the hippocampus as an optimization problem, using Deep Belief Networks (DBNs). The proposed framework, DeepDNN, enables a DL paradigm by learning nonlinear models of the hippocampus. Experiments on both synthetic and real-world data, and experiments using human and a dataset from the International Brain Project (IB) and the NIH NeuroImage Retinal Descent (RAED) datasets, demonstrate the efficacy of our DL system over the standard DeepDNN models.
Learning Spatial and Sparse Generative Models with an Application to Machine Reading Comprehension
MultiView Matching Based on a Unified Polynomial Pooling Model
An Overview of Deep Learning Techniques and Applications
Deep End-to-End Neural StackingIn this paper, we describe a deep learning (DL) framework for segmentation of the human hippocampus. The hippocampus is considered as a functional brain region that contains various sensory and motor functions. In this context, a neural network (NN) has received attention in recent years. However, the classification of the hippocampal region by an NN does not provide a good performance for the task, because of the limited number of labeled examples. Therefore, we propose an DL framework that takes a dataset of hippocampal data and models the information in the hippocampus as an optimization problem, using Deep Belief Networks (DBNs). The proposed framework, DeepDNN, enables a DL paradigm by learning nonlinear models of the hippocampus. Experiments on both synthetic and real-world data, and experiments using human and a dataset from the International Brain Project (IB) and the NIH NeuroImage Retinal Descent (RAED) datasets, demonstrate the efficacy of our DL system over the standard DeepDNN models.