Multi-point shape recognition with spatial regularization – We present a novel method to generate a realistic visual representation of the scene. Our method consists of three steps: 1) segment (pixel-wise) images from the ground state and 2) annotate our images. We show that each pixel corresponds to a unique image image in the input image space. Our method can be seen as a way to generate realistic visual representations of the scene in a novel way, by applying a neural network to a visual field and then applying multiple feature learning methods on this image to learn its semantic domain. The method is applied to the MNIST dataset and was evaluated on different datasets such as the Dictionaries and ImageNet, showing promising results.
We develop a new, yet complementary method of learning an attention-based topic model (DASMM) for multi-dimensional representations. Traditionally, the task of representing a new topic is treated as a single- or multiple-dimensional representation task. We solve the multi-dimensional representation learning problem in an alternating direction method of learning the topic labels, and show that even a single word can significantly help in learning a multi-dimensional representation. We prove that the learning problem can be solved efficiently via the non-convex convex relaxation of the convex relaxation equation over multi-dimensional representations. In a large amount of experiments, we evaluate the effectiveness of our method in various domains and show the benefits of this method.
Learning from the Hindsight Plan: On Learning from Exact Time-series Data
Towards machine understanding of human behavior and the nature of reward motivation
Multi-point shape recognition with spatial regularization
Stochastic Optimization via Variational Nonconvexity
Multi-dimensional representation learning for word retrievalWe develop a new, yet complementary method of learning an attention-based topic model (DASMM) for multi-dimensional representations. Traditionally, the task of representing a new topic is treated as a single- or multiple-dimensional representation task. We solve the multi-dimensional representation learning problem in an alternating direction method of learning the topic labels, and show that even a single word can significantly help in learning a multi-dimensional representation. We prove that the learning problem can be solved efficiently via the non-convex convex relaxation of the convex relaxation equation over multi-dimensional representations. In a large amount of experiments, we evaluate the effectiveness of our method in various domains and show the benefits of this method.