Dense-2-Type CNN for Stereo Visual Odometry – In the context of deep learning, deep neural networks (DNN’s) have recently gained popularity due to their impressive performance on most tasks, such as object classification, language understanding and object recognition. However, for the most part, DNN’s are not fully-connected. In order to handle multiple layers, these layers are used to perform the classification of the input image, which has been a challenging task for deep neural networks. In this work, we propose a novel hierarchical LSTM architecture, which is capable of being stacked to provide a higher level learning capability. Unlike the previous hierarchical architectures, the learned LSTM structures are connected to the learned models by a novel set of hidden layers, which can be easily updated via a back-propagation algorithm. Moreover, we show that the learnt LSTM can directly be used for segmentation, which is a highly desirable task for neural networks in this context. Experimental analysis using a simulated human benchmark dataset demonstrates that the proposed architecture is significantly better for the proposed task.
We investigate the problem of visualizing the temporal dynamics of a user interacting with a user from a natural perspective. We propose a novel architecture that achieves state-of-the-art performance on several benchmark datasets, and propose that it can be used to learn a state-of-the-art representation from the user’s observed actions. This means that our network-based models offer state-of-the-art performance even in datasets that lack user interaction. Experimental results show that the proposed representation can be used for modeling of the user’s action and the user’s behavior.
Fluorescence: a novel method for dynamic time-image classification from fMRI-data using CNNs
Learning the Structure of a Low-Rank Tensor with Partially-Latent Variables
Dense-2-Type CNN for Stereo Visual Odometry
Determining Pointwise Gradients for Linear-valued Functions with Spectral Penalties
Sketching for Linear Models of Indirect SupervisionWe investigate the problem of visualizing the temporal dynamics of a user interacting with a user from a natural perspective. We propose a novel architecture that achieves state-of-the-art performance on several benchmark datasets, and propose that it can be used to learn a state-of-the-art representation from the user’s observed actions. This means that our network-based models offer state-of-the-art performance even in datasets that lack user interaction. Experimental results show that the proposed representation can be used for modeling of the user’s action and the user’s behavior.