A Simple End-to-end Deep Reinforcement Learning Method for Situation Calculus – We propose a novel neural model for action recognition and action planning, in which information is encoded explicitly by the task-oriented context of the scene. The structure of the context is modeled by a pair of temporal- and spatial-dependent neural networks. As the task-oriented context of the scene plays a key role in the learned performance, this model achieves state-of-the-art results in both datasets.
In this paper, we propose a general framework to learn an objective function for action recognition and planning using convolutional neural networks. We show how to extract features of the objective function in the training set of such networks. We also show that the learning of feature maps of the objective function is a crucial step for a successful decision making problem. In doing so, we show how to make use of the available information in a supervised learning setting to learn a discriminative objective function. Our training data are shown to be rich in semantic information and we show how to use state-of-the-art image-level classification techniques to further improve the learning performance.
In this paper, the task of training a new classifier on image data is presented. Based on the notion of the ‘good old-fashioned’ classifier, there is defined a new classifier based on its ability to infer the class label that is associated with the data. We provide experimental tests that show that the new classifier produces similar results as the existing classifier. Finally, it provides for the first time the results obtained using the popular Convolutional Neural Network technique.
Deep Learning Semantic Part Segmentation
A Simple End-to-end Deep Reinforcement Learning Method for Situation Calculus
The Information Bottleneck Principle
Learning from Distributional Features in Graph Corpora with Applications to Medical Image AnalysisIn this paper, the task of training a new classifier on image data is presented. Based on the notion of the ‘good old-fashioned’ classifier, there is defined a new classifier based on its ability to infer the class label that is associated with the data. We provide experimental tests that show that the new classifier produces similar results as the existing classifier. Finally, it provides for the first time the results obtained using the popular Convolutional Neural Network technique.