Using Deep Learning and Deep Convolutional Neural Networks For Deformable Object Recognition – We present a method to automatically identify unlabeled and labeled objects from video that are likely to be labeled with a particular label. The identification of such instances is a challenging task in computer vision, which has an interesting dynamic due to multiple factors. To tackle the problem, we propose a joint model framework called K-CNN and N-CNN. Extensive evaluation on a challenging dataset, CIFAR-10 and CIFAR-100, shows that N-CNN outperforms CNN based approaches by a large margin, with near-optimal classification performance.
Neural networks provide a powerful representation of abstract thought patterns and can be used to model biological systems, as has been observed by many other researchers. However, the network representation suffers from overfitting, which leads to the lack of discriminative representations given the input data. We propose a novel approach to perform neural network representation learning by leveraging sparse representations and a recently proposed learning algorithm to learn a sparse representation from a single input. Through a novel deep learning mechanism that explicitly incorporates the dimensionality of the input data, the network learns a classification objective to capture the learned model structure. Importantly, we demonstrate that the proposed approach outperforms some state-of-the-art classifiers in the task of human visual recognition.
Predicting behavior of a child by exploiting context information in reinforcement learning networks
Neural Embeddings for Sentiment Classification
Using Deep Learning and Deep Convolutional Neural Networks For Deformable Object Recognition
Object Category Recognition in Video by Joint Stereo Matching and Sparse Coding
Density Characterization of Human Poses In The Presence of Fisher Vectors and One-Class ClassifiersNeural networks provide a powerful representation of abstract thought patterns and can be used to model biological systems, as has been observed by many other researchers. However, the network representation suffers from overfitting, which leads to the lack of discriminative representations given the input data. We propose a novel approach to perform neural network representation learning by leveraging sparse representations and a recently proposed learning algorithm to learn a sparse representation from a single input. Through a novel deep learning mechanism that explicitly incorporates the dimensionality of the input data, the network learns a classification objective to capture the learned model structure. Importantly, we demonstrate that the proposed approach outperforms some state-of-the-art classifiers in the task of human visual recognition.