Adaptive Sparse Convolutional Features For Deep Neural Network-based Audio Classification – In this paper, we propose an end-to-end, fully convolution network which allows for efficient extraction of the low-level information in speech and visual data. The proposed model is a multi-stage, fully convolutional network and utilizes the convolutional layers together to learn a hierarchical representation. After learning, the extracted high-level information is used as a discriminator for inferring the audio patterns to be extracted, and then a sequence of the high-level information is then extracted from the discriminator. Based on the proposed model, the neural network is trained without any additional preprocessing step. To the best of our knowledge, this is the first fully-convolutional neural network that can be used for speech retrieval tasks.
We present a method for joint learning of segmentation and recognition using deep learning. The segmentation method is the basis for several deep learning architectures to address the problem of object detection in video. As a technique, segmentation is trained using deep learning. By using CNNs for embedding and training, one achieves an object detection performance comparable to that of CNNs trained on object detectors. In contrast, the object detection performance can be measured using linear or nonlinear discriminant analysis. The segmentation method can use a combination of both linear and nonlinear discriminant analysis in order to improve the performance of the final target. We discuss our approach in the paper and propose a technique for joint learning segmentation.
An iterative k-means method for minimizing the number of bound estimates
Learning to Distill Similarity between Humans and Robots
Adaptive Sparse Convolutional Features For Deep Neural Network-based Audio Classification
Stochastic Learning of Graphical Models
Multi-Modal Deep Convolutional Neural Networks for Semantic SegmentationWe present a method for joint learning of segmentation and recognition using deep learning. The segmentation method is the basis for several deep learning architectures to address the problem of object detection in video. As a technique, segmentation is trained using deep learning. By using CNNs for embedding and training, one achieves an object detection performance comparable to that of CNNs trained on object detectors. In contrast, the object detection performance can be measured using linear or nonlinear discriminant analysis. The segmentation method can use a combination of both linear and nonlinear discriminant analysis in order to improve the performance of the final target. We discuss our approach in the paper and propose a technique for joint learning segmentation.