Convolutional Kernels for Graph Signals – The state of the art of graph signal processing is hampered by the large-margin, monotonicity bound in the dimensionality of the signal. In this work we focus on the problem of learning a network for the real-world domain by combining several techniques from natural language processing. We propose a novel approach by incorporating the concept of hidden Markov models under a unified framework. We show that this framework can be extended to the problem of graph signals, where this framework also benefits from the novel structure and high degree of independence of the data. Specifically, we consider a network in which each node contains the most important bits of the input data, and the other nodes contain the small bits. We provide an efficient inference scheme capable of solving the problem, which allows us to make the network learnable for graph signals. We show that our network can be applied to a variety of graphs, and provide experimental validation on synthetic graphs in the context of supervised classification of graphs.
This paper proposes a method to solve the continuous temporal reasoning question of DPT (discovery and re-iscovery of temporal information). The core assumption underlying the proposed method is that each object is a temporal entity, and its event-related events cannot be represented by any semantic or linguistic properties. We propose the concept of re-orging (orging) temporal entities to model the entity’s event-related events. As long as objects are moving in temporal space, this concept should be sufficient to represent them as temporal entities. The key innovation is the concept of re-orging-ness (the ability to re-org as many objects as it can). We show that, according to the proposed method, all temporal entities in the temporal space can belong to the same entity. To the best of our knowledge, this is the first step toward temporal reasoning in this setting, and we demonstrate that our method performs well in practice and can be applied to any temporal knowledge processing system that is given an input of time series data.
Scalable Sparse Subspace Clustering with Generative Adversarial Networks
Towards a unified paradigm of visual questions and information domain modeling
Convolutional Kernels for Graph Signals
Generative model of 2D-array homography based on autoencoder in fMRI
Learning Discrete Event-based Features for Temporal ReasoningThis paper proposes a method to solve the continuous temporal reasoning question of DPT (discovery and re-iscovery of temporal information). The core assumption underlying the proposed method is that each object is a temporal entity, and its event-related events cannot be represented by any semantic or linguistic properties. We propose the concept of re-orging (orging) temporal entities to model the entity’s event-related events. As long as objects are moving in temporal space, this concept should be sufficient to represent them as temporal entities. The key innovation is the concept of re-orging-ness (the ability to re-org as many objects as it can). We show that, according to the proposed method, all temporal entities in the temporal space can belong to the same entity. To the best of our knowledge, this is the first step toward temporal reasoning in this setting, and we demonstrate that our method performs well in practice and can be applied to any temporal knowledge processing system that is given an input of time series data.
Leave a Reply