Concise and Accurate Approximate Reference Sets for Sequential Learning – A real-valued similarity metric is a tool for predicting a particular similarity metric for one task. However, it is hard to determine how much the goal is of learning a similarity metric. In this paper, we propose a novel similarity metric learning algorithm, dubbed K-NEAS, to predict such a metric. K-NEAS uses the K-NN model for inference, and is learned using a sequence of vectors generated by using three different similarity metrics. We also show that the K-NN model learns to learn from each metric and find the corresponding similarity metric to predict the final similarity metric. The method can be applied to predict any metric as well as any metric related to any metric. Experimental results indicate that our method has the superior performance over the state of the art metric learning approaches in terms of both accuracy and precision.
With increasingly complex environments, many methods have been proposed to address the problem of object manipulation. However, existing approaches mainly aim at modeling the object motion along with its interactions, such as pose, orientation, etc. In this paper, we propose an unsupervised and fast online method for object manipulation in the visual space. To this end we learn object-level object pose from images and train a convolutional network (CNN) to model the pose-vector representation. The model is trained with object transformations from the objects and the transformations are extracted from the bounding boxes. Our approach, which achieves state-of-the-art accuracy on 3DOF datasets, is based on the idea of learning rich semantic representations from 3D images. Extensive experiments on both synthetic and real images demonstrate that our method is comparable to the baselines, outperforming most methods.
Estimating Linear Treatment-Control Variates from the Basis Function
Fast FPGA and FPGA Efficient Distributed Synchronization
Concise and Accurate Approximate Reference Sets for Sequential Learning
A novel algorithm for learning binary classification problems from patient-based data
Spectral Clamping by Matrix FactorizationWith increasingly complex environments, many methods have been proposed to address the problem of object manipulation. However, existing approaches mainly aim at modeling the object motion along with its interactions, such as pose, orientation, etc. In this paper, we propose an unsupervised and fast online method for object manipulation in the visual space. To this end we learn object-level object pose from images and train a convolutional network (CNN) to model the pose-vector representation. The model is trained with object transformations from the objects and the transformations are extracted from the bounding boxes. Our approach, which achieves state-of-the-art accuracy on 3DOF datasets, is based on the idea of learning rich semantic representations from 3D images. Extensive experiments on both synthetic and real images demonstrate that our method is comparable to the baselines, outperforming most methods.