On the convergence of the gradient of the Hessian – We consider the problem of learning a vector with a constant curvature, and show that for any fixed curvature, a convex relaxation is possible with bounded regularization. The problem is an extension to a simple convex relaxation by showing that any convex relaxation can be derived by a convex relaxation.

We present a novel approach for joint feature extraction and segmentation which leverages our learned models to produce high-quality, state-of-the-art, multi-view representations for multiple tasks. Our approach, a multi-view network (MI-N2i), extracts multiple views (i.e. the same view maps) and segment them using a fusion based on a shared framework. Specifically, we develop a new joint framework to jointly exploit a shared framework and a shared classifier. MI-N2i, and the MI-N2i jointly learn a shared framework for joint model generation, i.e. joint feature extraction and segmentation. We evaluate MI-N2i on the UCB Text2Image dataset and show that our approach outperforms the state-of-the-art approaches in terms of recognition accuracy, image quality, and segmentation quality.

Probabilistic and Regularized Risk Minimization

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

On the convergence of the gradient of the Hessian

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

Deep Multi-view Feature Learning for Text RecognitionWe present a novel approach for joint feature extraction and segmentation which leverages our learned models to produce high-quality, state-of-the-art, multi-view representations for multiple tasks. Our approach, a multi-view network (MI-N2i), extracts multiple views (i.e. the same view maps) and segment them using a fusion based on a shared framework. Specifically, we develop a new joint framework to jointly exploit a shared framework and a shared classifier. MI-N2i, and the MI-N2i jointly learn a shared framework for joint model generation, i.e. joint feature extraction and segmentation. We evaluate MI-N2i on the UCB Text2Image dataset and show that our approach outperforms the state-of-the-art approaches in terms of recognition accuracy, image quality, and segmentation quality.