Adaptive Stochastic Learning – We present a novel approach, based on an extended version of the recently proposed deep convolutional neural networks (CNNs) learning from input images. At a higher level of abstraction, we use two iterative steps for learning a global feature for each image. When the feature is a high-dimensional feature, the CNNs will learn a sparse representation of the feature with respect to the input image. When the feature is a low-dimensional feature, the CNNs will learn a low-dimensional representation from the input image. This approach allows for both direct and indirect feedback loops where the input is the source domain and the outputs of a CNN are the output domain. The proposed approach is demonstrated on MNIST and ImageNet datasets. The method achieved comparable performance to state-of-the-art CNNs by only training on three datasets and outperforming the state-of-the-art CNNs on two of them by a large margin.
This paper proposes a framework for learning dense Markov networks (MHN) over a large data set. MHN is a family of deep learning methods focusing on image synthesis over structured representations. Recent studies have evaluated three MHN architectures on a range of tasks: 1) text recognition, 2) text classification, and 3) face recognition. MHN models provide a set of outputs, that can be useful for learning a novel representation over images. However, it may take many tasks without good input data. Therefore, MHN model is a multi-task learning system. First, we learn MHN from data. We then use a mixture of both learned inputs and output outputs for learning MHN. Second, we use the same inputs in two different tasks, namely object detection and visual pose estimation.
A Novel Concept Space: Towards Understanding the Emergence of Fusion of Visual Concepts in Video
Recurrent Neural Network based Simulation of Cortical Task to Detect Cervical Pre-Canger Pathways
Adaptive Stochastic Learning
Predicting Precision Levels in Genetic Algorithms
Tangled Watermarks for Deep Neural NetworksThis paper proposes a framework for learning dense Markov networks (MHN) over a large data set. MHN is a family of deep learning methods focusing on image synthesis over structured representations. Recent studies have evaluated three MHN architectures on a range of tasks: 1) text recognition, 2) text classification, and 3) face recognition. MHN models provide a set of outputs, that can be useful for learning a novel representation over images. However, it may take many tasks without good input data. Therefore, MHN model is a multi-task learning system. First, we learn MHN from data. We then use a mixture of both learned inputs and output outputs for learning MHN. Second, we use the same inputs in two different tasks, namely object detection and visual pose estimation.