Structural Matching through Reinforcement Learning – This paper addresses the problem of supervised learning of visual attention networks by applying deep reinforcement learning (DL) to reinforcement learning tasks. DL is an end-to-end learning algorithm that does not require the user to learn any specific visual scene. In particular, DL can learn to capture visual dependencies and to adapt to different visual cues of the scene at different levels of its complexity, in a global way. In this paper, we propose a novel DL model trained with the task-dependent visual cue to learn to predict the next action sequence over the entire network. As an example, we consider our attention-to-sequence learning algorithm which is trained from scratch and learns to predict the next sequence over every visual cue of an object at each level of the network (i.e. after training the supervised models only on the task-dependent visual cue). We demonstrate that our DL model outperforms the state-of-the-art attention based vision models in terms of accuracy, on an unstructured object detection task.
This paper shows that a simple linear discriminantial method, based on a non-overlapping filter, can be more robust to noisy input signals while keeping high discriminative power in image domain. In this paper, we describe the method and compare it with several other discriminantial approaches. The method is evaluated on four widely used benchmark datasets with different classification rates. The results showed that our method achieves better results than the previous state-of-the-art methods for classification accuracy in the class of noisy images of an unknown size.
Eliminating Dither in RGB-based 3D Face Recognition with Deep Learning: A Unified Approach
Convolutional Residual Learning for 3D Human Pose Estimation in the Wild
Structural Matching through Reinforcement Learning
Linking and Between Event Groups via Randomized Sparse Subspace
Deep Learning for ClassificationThis paper shows that a simple linear discriminantial method, based on a non-overlapping filter, can be more robust to noisy input signals while keeping high discriminative power in image domain. In this paper, we describe the method and compare it with several other discriminantial approaches. The method is evaluated on four widely used benchmark datasets with different classification rates. The results showed that our method achieves better results than the previous state-of-the-art methods for classification accuracy in the class of noisy images of an unknown size.