Fast Convolutional Neural Networks via Nonconvex Kernel Normalization – In this work, we propose a new framework for learning deep CNNs from raw image patches. As a case study, we propose a novel and scalable method for learning deep CNNs using compressed convolutional neural networks (convNNs). We first show that constrained CNNs achieve state-of-the-art performance in many tasks, while using a compact representation of the image patches. We then show that conv nets can be trained to generalize to unseen patches easily. Our experiments show that our deep CNN approach is able to achieve state-of-the-art performance on several benchmark datasets, as compared to other state-of-the-art methods.

Neural networks have achieved good results in many domains. However, they have become more generic and difficult to apply in applications that require large scale training data.

In this paper, we propose a novel method that simultaneously uses multiple layers of pre-trained convolutional neural networks to learn classifier labels, in a non-convex problem. Our model is trained in a low-dimensional space, where both the dimension and the number of layers are fixed. We employ a deep learning approach that can be used as the basis for learning classifier label pairs for a small number of layers, in all-important settings. The trained model is then transferred to another low-dimensional space, where it is trained to learn the labeled labels for a subset of a small set of labels, for instance an image. We evaluate our approach on the MNIST dataset and demonstrate that our model outperforms a state-of-the-art model trained only on image labels.

A Comparative Study of Support Vector Machine Classifiers for Medical Records

Dynamic Network Models: Minimax Optimal Learning in the Presence of Multiple Generators

Fast Convolutional Neural Networks via Nonconvex Kernel Normalization

Learning the Latent Representation of Words in Speech Using Stochastic Regularized LSTM

Learning to see through the mask: Spatial selective sampling for image restorationNeural networks have achieved good results in many domains. However, they have become more generic and difficult to apply in applications that require large scale training data.

In this paper, we propose a novel method that simultaneously uses multiple layers of pre-trained convolutional neural networks to learn classifier labels, in a non-convex problem. Our model is trained in a low-dimensional space, where both the dimension and the number of layers are fixed. We employ a deep learning approach that can be used as the basis for learning classifier label pairs for a small number of layers, in all-important settings. The trained model is then transferred to another low-dimensional space, where it is trained to learn the labeled labels for a subset of a small set of labels, for instance an image. We evaluate our approach on the MNIST dataset and demonstrate that our model outperforms a state-of-the-art model trained only on image labels.