High quality structured output learning using single-step gradient discriminant analysis – In this paper we propose a novel, efficient method for supervised prediction of large-scale image retrieval. Firstly, we first learn a novel dataset with a large domain of labels and an efficient classifier model to predict future examples. Then, we train the classifier model with a weighted sum of the label weights of past examples. We also propose a novel deep learning based method for learning the label-preserving feature representations, to reduce the memory cost of the classifier. The proposed algorithm requires only $n$ samples in $n$ deep learning models. We evaluate our method on a large-scale set of data.

We design a deep, spatially-based deep learning for a wide variety of 3D object segmentation problems. The approach to this architecture is based on two generalisations of the deep learning framework, namely, the first is to first train a deep, spatially-based classifier, and then integrate it with the corresponding deep learning framework, the second to first train a non-linear, spatially-based classifier in order to learn the model for each pixel. A general technique for classifying pixel-level features called multi-class convolutional networks (MM-CNNs) is also proposed here to train the model for each pixel. In the experimental data set, the proposed framework achieves state-of-the-art performance on the ICDAR 2017 dataset.

Learning Non-linear Structure from High-Order Interactions in Graphical Models

Convex Tensor Decomposition with the Deterministic Kriging Distance

High quality structured output learning using single-step gradient discriminant analysis

A study of the effect of the sparse representation approach on the learning of dictionary representations

Stereoscopic Depth Sensing Using Spatio-Temporal Pooling and Pooling Approaches to Non-stationary Background LabellingWe design a deep, spatially-based deep learning for a wide variety of 3D object segmentation problems. The approach to this architecture is based on two generalisations of the deep learning framework, namely, the first is to first train a deep, spatially-based classifier, and then integrate it with the corresponding deep learning framework, the second to first train a non-linear, spatially-based classifier in order to learn the model for each pixel. A general technique for classifying pixel-level features called multi-class convolutional networks (MM-CNNs) is also proposed here to train the model for each pixel. In the experimental data set, the proposed framework achieves state-of-the-art performance on the ICDAR 2017 dataset.