Optimizing parameter selection in Datalog transformations – We propose a new method for minimizing the loss of the parameters, via maximizing their regret in terms of the expected regret squared. This strategy is especially well suited for situations where the loss is not sensitive to the transformation’s behavior, such as when the transformation is a morphologically rich structure, or when the transformation is an optimization problem. Specifically, this strategy makes use of the notion of the least-squares minimizer when learning the parameters from data, and uses it to guarantee the optimality of the minimizer, which is the result of a priori assumptions. We apply this strategy to transform prediction by using both the maximum and minimum-margin assumptions, and apply this strategy to Datalog predictions from the same data. Our results suggest that it is possible to obtain a more natural optimization-inducing minimizer: a minimizer which maximizes the risk of the model over the space of the minimizers. Based on this optimization-inducing minimizer, our algorithm minimizes a risk of $1-$f$.

The goal is to use deep learning approach to automatically produce high quality images from a large database. But there are many reasons why it is not easy to do so. In this paper, we propose a novel approach to create a large-scale, low-cost image retrieval using multi-view semantic segmentation. Our architecture consists of two main components: (1) a robustly model-driven deep neural network (DRNN) module, (2) an image captioning approach that simultaneously learns a semantic model of the underlying image. Extensive experiments were conducted on the COCO dataset to show that our approach achieves state-of-the-art retrieval performance on a huge set of images.

Adversarially Learned Online Learning

Comparing the Learning-Model Classroom Approach, Constraint-Based Approach, and Conceptual Space

Optimizing parameter selection in Datalog transformations

Learning from Incomplete Observations

Towards Knowledge Based Image RetrievalThe goal is to use deep learning approach to automatically produce high quality images from a large database. But there are many reasons why it is not easy to do so. In this paper, we propose a novel approach to create a large-scale, low-cost image retrieval using multi-view semantic segmentation. Our architecture consists of two main components: (1) a robustly model-driven deep neural network (DRNN) module, (2) an image captioning approach that simultaneously learns a semantic model of the underlying image. Extensive experiments were conducted on the COCO dataset to show that our approach achieves state-of-the-art retrieval performance on a huge set of images.