Keynote Lectures

Abstract
The popularity of web-based mapping services such as Google Earth/Maps and Microsoft Virtual Earth (Bing), has led to an increasing awareness of the importance of location data and its incorporation into both web-based search applications and the databases that support them in the past, attention to location data had been primarily limited to geographic information systems (GIS), where locations correspond to spatial objects and are usually specified geometrically.
However, in the web-based applications, the location data often corresponds to place names and is usually specified textually.
An advantage of such a specification is that the same specification can be used regardless of whether the place name is to be interpreted as a point or a region. Thus the place name acts as a polymorphic data type in the parlance of programming languages. However, its drawback is that it is ambiguous. In particular, a given specification may have several interpretations, not all of which are names of places. For example, ''Jordan'' may refer to both a person as well as a place.
Moreover, there is additional ambiguity when the specification has a place name interpretation. For example, ''Jordan'' can refer to a river or a country while there are a number of cities named ''London''.
In this talk we examine the extension of GIS concepts to textually specified location data and review search engines that we have developed to retrieve documents where the similarity criterion is not based solely on exact match of elements of the query string but instead also based on spatial proximity. Thus we want to take advantage of spatial synonyms so that, for example, a query seeking a rock concert in Tel Aviv would be satisfied by a result finding a rock concert in Herzliyah of Petach Tikva. This idea has been applied by us to develop the STEWARD (Spatio-Textual Extraction on the Web Aiding Retrieval of Documents) system for finding documents on website of the Department of Housing and Urban Development. This system relies on the presence of a document tagger that automatically identifies spatial references in text, pdf, word, and other unstructured documents. The thesaurus for the document tagger is a collection of publicly available data sets forming a gazetteer containing the names of places in the world. Search results are ranked according to the extent to which they satisfy the query, which is determined in part by the prevalent spatial entities that are present in the document. The same ideas have also been adapted by us to collections of news articles as well as Twitter tweets resulting in the NewsStand and TwitterStand systems, respectively, which will be demonstrated along with the STEWARD system in conjunction with a discussion of some of the underlying issues that arose and the techniques used in their implementation. Future work involves applying these ideas to spreadsheet data.

Abstract
Computers can now do things that their programmers cannot explain or understand in detail: today's Artificial Intelligence has found a way to bypass the need for understanding a phenomenon before we can replicate it in a computer. The technology that made this possible is machine learning: a method to program computers by showing them examples of the desired behaviour. And the fuel that powers it all is DATA. For this reason, data has been called the new oil: a new natural resource, that businesses and scientists alike can leverage, by feeding it to massive learning computers to do things that we do not understand well enough to implement them with a traditional program. This new way of working - often called Big Data - is all about predicting, not explaining. It is about knowing what a new drug will do to a patient, not why.
While the opportunities created by this approach are becoming clear, and the technological challenges are gradually being met, we still struggle to come to terms with the implications of this transition. Was not science meant to help us make sense of the world? Or is it just meant to deliver good predictions? Are patterns in data capable of replacing theoretical understading? It is also important to remember that the fuel that powers this revolution is very often our own personal data, and that we still do not have a clear cultural framework to think about all this.

Abstract
Traditional pattern recognition and machine learning techniques take typically a reductionist view, in the sense that they tend to view each object in isolation, and are inherently essentialist, in the sense that they are deeply grounded in the assumption that the objects populating the world are endowed with a number of “essential” (intrinsic) features which determine the category to which they belong. Starting from the common-sense observation that in the real world objects do not live in a vacuum, and that contextual, or relational, constraints do provide a rich and often unexplored source of information, I’ll argue for a holistic perspective to pattern recognition, with a view to overcome the limitations of today’s approaches. I’ll maintain that game theory offers an elegant and general conceptual framework that serves well this purpose and I’ll provide game-theoretic formulations of some specific pattern recognition and machine learning problems.

Abstract

The emergence of cheap structured light sensors, like the Kinect, opened the door to an increased interest in all matters related to the processing of 3D visual data. Applications for these technologies are abundant, from robot vision to 3D scanning.
In this presentation we will go through the main steps used on a typical 3D vision system, from sensors and point clouds up to understanding the scene contents, including key point detectors, descriptors, set distances, object segmentation, recognition and tracking and the biological motivation for some of these methods. We will present several approaches developed at our lab and discuss some current challenges.