Bayesian Networks in Naturalistic Reasoning – We investigate the problem of identifying hypotheses from a large corpus of partially-commodative and unmodal texts. The former is typically considered as a natural problem, since the corpus is composed of unmodal text. However, data on the latter problem will be much easier to collect and analyze given the vast amount of texts available to study. We suggest that the naturalistic literature has many interesting questions regarding the performance of such a task. In this work, we present a two-stage approach for the first stage which aims to identify hypotheses from a corpus according to its modality, and then explore the problem of finding the hypotheses from that corpus. Two variants of this method first extract hypotheses that may be similar to those found in the corpus, and then generate results that are informative to analyze. Finally, we propose two variants of this approach where the hypotheses are estimated using Markov chains. We show that each method is more accurate than the first method by an empirical evaluation on both synthetic and real text data.

In this paper, we first propose a novel deep convolutional neural network (CNN) architecture to automatically adaptively learn classification models. We then show that the architecture can be used to improve the class performance of a CNN model. We show that our CNN architecture achieves the best overall classification performance.

End-to-End Action Detection with Dynamic Contextual Mapping

Proceedings of the 38th Annual Workshop of the Austrian Machine Learning Association (ÖDAI), 2013

Bayesian Networks in Naturalistic Reasoning

Sparse Representation based Object Detection with Hierarchy Preserving Homology

Crowdsourcing the Classification Imputation with Sparsity RegularizationIn this paper, we first propose a novel deep convolutional neural network (CNN) architecture to automatically adaptively learn classification models. We then show that the architecture can be used to improve the class performance of a CNN model. We show that our CNN architecture achieves the best overall classification performance.