Constrained Multi-View Image Classification with Multi-temporal Deep CNN Regressions – In this paper, we propose to model the input and output characteristics of the images by utilizing a combination of 3D and depth features in a manner to extract meaningful semantic information from the image. Unlike conventional 2D CNN based supervised learning, we propose a novel 3D segmentation and 3D convolutional neural network based approach for the 3D segmentation task. This new CNN architecture is able to be adapted to handle the different aspects needed by 3D CNNs, i.e., the feature representation and the 3D depth information. We evaluated our method on two datasets, one with RGB-D data and one without RGB-D data, and compared our methods on both datasets. Experimental results show that the proposed approach significantly outperforms the state-of-the-art CNN methods and also achieves state-of-the-art results on both datasets.

We describe a method to learn a posterior function of a model conditioned on unseen data under the assumption that the data is annotated. We show that this method is a proper approximation of the posterior, which is not a direct prior of any model but a prior of the entire training set. We illustrate by showing an example on supervised learning.

Learning and Estimation from Unbalanced Data

Learning from Past Mistreatment

Constrained Multi-View Image Classification with Multi-temporal Deep CNN Regressions

Training of Deep Convolutional Neural Networks for Large-Scale Video Classification

Learning to Rank from Unlabeled Data with Conditional Rank InferenceWe describe a method to learn a posterior function of a model conditioned on unseen data under the assumption that the data is annotated. We show that this method is a proper approximation of the posterior, which is not a direct prior of any model but a prior of the entire training set. We illustrate by showing an example on supervised learning.