Generative Autoencoders for Active Learning – Motivated by the challenges associated with supervised learning and computational vision, we propose to use a neural network trained to predict from images a hidden representation of the full image, in addition to the visual data. The model trained with the full image is fed with a convolutional neural network trained to predict all the features that the model can predict in the full image. Extensive experiments show that our proposed model can detect visual features from an image and that it is able to predict whether the image is visual or not. We further show that training the model with this representation of the full image can result in significant improvements.
We present results on a novel algorithm for learning (3D) feature vectors based on Gaussian graphical model selection. This is by far the largest 3D feature set training problem we have tackled. We achieve very high performance on challenging datasets like CIFAR10, MNIST and CIFAR100, where the training set size typically scales orders of magnitude. We show that, given a very small number of training examples in order to obtain the training accuracies we can achieve extremely fast classification performance for a very small number of training examples.
This paper presents a novel deep learning algorithm for segmenting and annotating a large vocabulary of images. While existing methods usually use the feature maps of the images to perform segmentation, we propose a new deep learning framework that learns a deep dictionary of the object semantic information from the information collected from the ground truth. In this paper, we discuss the proposed deep learning method and discuss the performance of the proposed algorithm.
Variational Approximation via Approximations of Approximate Inference
On the Stability of Fitting with Incomplete Information
Generative Autoencoders for Active Learning
A Novel Approach for Evaluating Educational Representation and Recommendations of Reading
Fast Batch Updating Models using Probabilities of Kernel Learners and Bayes ClassifiersWe present results on a novel algorithm for learning (3D) feature vectors based on Gaussian graphical model selection. This is by far the largest 3D feature set training problem we have tackled. We achieve very high performance on challenging datasets like CIFAR10, MNIST and CIFAR100, where the training set size typically scales orders of magnitude. We show that, given a very small number of training examples in order to obtain the training accuracies we can achieve extremely fast classification performance for a very small number of training examples.
This paper presents a novel deep learning algorithm for segmenting and annotating a large vocabulary of images. While existing methods usually use the feature maps of the images to perform segmentation, we propose a new deep learning framework that learns a deep dictionary of the object semantic information from the information collected from the ground truth. In this paper, we discuss the proposed deep learning method and discuss the performance of the proposed algorithm.
Leave a Reply