Convex Hulloo: Minimally Supervised Learning with Extreme Hulloo Search – This paper addresses the problem of high-dimensional high-resolution images. In this work, we propose a new deep nonlinear generative model to learn high-dimensional shape images by considering their temporal dynamics. We train the deep model via convolutional layers for predicting the shape features of the image by minimizing the reconstruction error. Our experiments show that our model provides high-resolution shape images with a rich temporal structure and can learn accurate predictions that outperform previous methods.

We propose a new novel approach to the problem of learning natural dialogues via deep encoder-decoder neural networks (DCDNNs). First, we augment the deep convolutional network with two layers of DNNs: one for dialogue, and one for language. Next, we train the convolutional DCDNNs to learn a convolutional dictionary with various convolutional sub-problems. Our approach leverages both hand-crafted and annotated dialogues. We propose two models that outperform DCDNN-based models (both train the DCNNs to learn the dictionary). Finally, we experiment our approach on a large collection of 10,000 human and 100,000 character short dialogues. To evaluate our approach, we conduct a trial on an audience sample for the SemEval 2017 evaluation of a class of short dialogues with 2.2 million dialogues.

Learning Multiple Tasks with Semantic Similarity

Towards a Deep Multitask Understanding of Task Dynamics

Convex Hulloo: Minimally Supervised Learning with Extreme Hulloo Search

Multispectral Image Fusion using Conditional Density Estimation

Probabilistic Learning and Sparse Visual Saliency in Handwritten CharactersWe propose a new novel approach to the problem of learning natural dialogues via deep encoder-decoder neural networks (DCDNNs). First, we augment the deep convolutional network with two layers of DNNs: one for dialogue, and one for language. Next, we train the convolutional DCDNNs to learn a convolutional dictionary with various convolutional sub-problems. Our approach leverages both hand-crafted and annotated dialogues. We propose two models that outperform DCDNN-based models (both train the DCNNs to learn the dictionary). Finally, we experiment our approach on a large collection of 10,000 human and 100,000 character short dialogues. To evaluate our approach, we conduct a trial on an audience sample for the SemEval 2017 evaluation of a class of short dialogues with 2.2 million dialogues.