Multi-modal retrieval is emerging as a new search paradigm that enables seamless information retrieval from various types of media. For example, users can simply snap a movie poster to search relevant reviews and trailers. To solve the problem, a set of mapping functions are learned to project high-dimensional features extracted from data of different media types into a common lowdimensional space so that metric distance measures can be applied. In this paper, we propose an effective mapping mechanism based on deep learning (i.e., stacked auto-encoders) for multi-modal retrieval. Mapping functions are learned by optimizing a new objective function, which captures both intra-modal and inter-modal semantic relationships of data from heterogeneous sources effectively. Compared with previous works which require a substantial amount of prior knowledge such as similarity matrices of intramodal data and ranking examples, our method requires little prior knowledge. Given a large training dataset, we split it into minibatches and continually adjust the mapping functions for each batch of input. Hence, our method is memory efficient with respect to the data volume. Experiments on three real datasets illustrate that our proposed method achieves significant improvement in search accuracy over the state-of-the-art methods.

We study the design of arbitrarily shaped electromagnetic (EM) concentrators and their potential applications. To obtain closed-form formulas of EM parameters for an arbitrarily shaped concentrator, we employ nonuniform rational B-spline (NURBS) to represent the geometrical boundary. Using the conformally optical transformation of NURBS surfaces, we propose the analytical design of arbitrarily shaped concentrators, which are composed of anisotropic and inhomogeneous metamaterials with closed-form constitutive tensors. The designed concentrators are numerically validated by full-wave simulations, which show perfectly directed EM behaviors. As one of the potential applications, we demonstrate a way to amplify plane waves using a rectangular concentrator, which is much more efficient and easier than the existing techniques. Using NURBS expands the generality of the transformation optics and could lead toward making a very general tool that would interface with commercial softwares such as 3D STUDIOMAX and MAYA.

An elliptical invisible cloak is proposed using a coordinate transformation in the elliptical-cylindrical coordinate system, which crushes the cloaked object to a line segment instead of a point. The elliptical cloak is reduced to a nearly-circular cloak if the elliptical focus becomes very small. The advantage of the proposed invisibility cloak is that none of the parameters is singular and the changing range of all parameters is relatively small.Comment: to appear in Appl. Phys. Let

Complex databases are challenging to explore and query by users unfamiliar with their schemas. Enterprise databases often have hundreds of inter-linked tables, so even when extensive documentation is available, new users must spend a considerable amount of time understanding the schema before they can retrieve any information from the database. The problem is aggravated if the documentation is missing or outdated, which may happen with legacy databases.In this paper we identify limitations of previous approaches to address this vexing problem, and propose a principled approach to summarizing the contents of a relational database, so that a user can determine at a glance the type of information it contains, and the main tables in which that information resides. Our approach has three components: First, we define the importance of each table in the database as its stable state value in a random walk over the schema graph, where the transition probabilities depend on the entropies of table attributes. This ensures that the importance of a table depends both on its information content, and on how that content relates to the content of other tables in the database. Second, we define a metric space over the tables in a database, such that the distance function is consistent with an intuitive notion of table similarity. Finally, we use a Weighted -Center algorithm under this distance function to cluster all tables in the database around the most relevant tables, and return the result as our summary. We conduct an extensive experimental study on a benchmark database, comparing our approach with previous methods, as well as with several hybrid models. We show that our approach not only achieves significantly higher accuracy than the previous state of the art, but is also faster and scales linearly with the size of the schema graph.

Based on transformation optics, we present a shrinking device, which can transform an arbitrary object virtually into a small-size object with different material parameters as one desires. Such an illusion device will confuse the detectors or the viewers, and hence the real size and material parameters of the enclosed object cannot be perceived. We fabricated and measured a shrinking device by using metamaterials, which works at the nonresonant frequency and has low loss. The device has been validated by both numerical simulations and experiments on circular and square objects. Good shrinking performance has been demonstrated.