Multi-level object recognition with distributed residual descriptors – This paper deals with a novel object-related neural network (NN) architecture for image-based collaborative filtering. In this design, the network includes the three elements of a discriminative model for each object category and a discriminative model for the image categories, which can generate a generic map which is the representation of each object category. To effectively incorporate the discriminative model in the current network, we propose a novel two-stage network architecture. First, the discriminative model has to be trained by incorporating the object categories into the network. Second, we use multiple discriminative models, which are trained independently and jointly under different tasks as a part of our supervised learning scheme. We demonstrate our framework and our use on several computer vision applications.
We present a new learning-driven method for predicting images based on feature vectors. The goal of the method is to determine if each pixel in a collection of images has a high probability of being a high-dimensional image. In this work, we use our learned model to predict a set of images. The model learns to distinguish objects from non-objects, and we combine this information with the observed feature vectors to predict their distribution. The method is highly time-efficient and has excellent generalization capabilities. We demonstrate how the proposed method can be used to enhance the prediction accuracy.
Fluency-based machine learning methods for the evaluation of legal texts
Robust 3D Registration via Deep Generative Models
Multi-level object recognition with distributed residual descriptors
Bayes-Ball and Fisher Discriminant Analysis
A new analysis of 3D street-view and local illumination using spatial and visual featuresWe present a new learning-driven method for predicting images based on feature vectors. The goal of the method is to determine if each pixel in a collection of images has a high probability of being a high-dimensional image. In this work, we use our learned model to predict a set of images. The model learns to distinguish objects from non-objects, and we combine this information with the observed feature vectors to predict their distribution. The method is highly time-efficient and has excellent generalization capabilities. We demonstrate how the proposed method can be used to enhance the prediction accuracy.