Deep Learning of Spatio-temporal Event Knowledge with Recurrent Neural Networks – We propose a novel algorithm for the automatic retrieval of spatio-temporal temporal dependencies in a real-time manner. We present efficient and interpretable algorithms for different domain-specific spatio-temporal dynamics. We test our algorithms on both synthetic and real world data sets. Finally, we show how to use our algorithms to build a neural network that models and predicts future spatio-temporal temporally dependent behaviors.
While the majority of the methods used for video classification make use of linear features derived from the target sequence, many existing models use a series of feature vectors instead of image features. We propose a novel class of features which is a mixture of linear and nonconvex representations of image labels that is significantly richer in information and is more appropriate for classifying a class of images. The new feature representation can be generalized to any nonlinear or non-convex matrix or is trained as a linear model using the class of image labels as training data. We illustrate how the new representation is used for learning and learning-based classification using both synthetic and real neural networks.
Sketch-Based Approach to Classification of Unstructured Data for Mobile Sensing
Multi-Resolution Video Super-resolution with Multilayer Biomedical Volumesets
Deep Learning of Spatio-temporal Event Knowledge with Recurrent Neural Networks
Interpretable Deep Learning with Dynamic Label Regularization
Training of Deep Convolutional Neural Networks for Large-Scale Video ClassificationWhile the majority of the methods used for video classification make use of linear features derived from the target sequence, many existing models use a series of feature vectors instead of image features. We propose a novel class of features which is a mixture of linear and nonconvex representations of image labels that is significantly richer in information and is more appropriate for classifying a class of images. The new feature representation can be generalized to any nonlinear or non-convex matrix or is trained as a linear model using the class of image labels as training data. We illustrate how the new representation is used for learning and learning-based classification using both synthetic and real neural networks.