Category: Uncategorized (page 3 of 4)

Multimedia content is ubiquitous and immense; the retrieval of multimedia content represents an important and hard challenge. This talk will focus on the analysis and retrieval of multimedia content, most notably images and videos. It will start by considering which data granularities are useful for retrieval, and will discuss methods for the temporal segmentation of video to shots and scenes. It will continue with discussing the problem of image and video indexing with both low-level features and high-level concepts (semantic indexing). For the latter case, the use of Web-based retrieval services towards automatically generating training corporal will also be briefly examined. The talk will then proceed to discussing the indexing of video with complex event labels, using either training video samples or simply a short textual description of each sought event. Throughout the talk, machine learning methods that are in the core of multimedia indexing with concept and event labels will also be sketched. Finally, ideas for future research will be discussed.

Lecturer: DR. VASILEIOS MEZARIS

In this tutorial we elaborate on recent advances in semantic processing of language and more specifically on entity and event recognition. We go deeper into the joint processing of language and visual data, where language data forms weak annotations to recognize information in images and video (e.g., recognitions of persons and events). We also consider examples where visual data provide context for disambiguating language. The methods regard unsupervised alignment techniques as well as weakly supervised graphical models, structured support vector machines and neural networks. A substantial focus is on representation learning and the use of topic models and neural embeddings. We will study how the recognized semantics will be integrated in retrieval models (such as vector and language models for retrieval). We discuss tasks, algorithms and evaluation.

The tutorial is composed of 5 parts:
1. Introduction and problem setting
2. Semantic processing of language
3. Semantic processing of multimedia
4. Integration of semantic recognitions in retrieval models
5. Ideas for future research.

Lecturer: Professor Marie-Francine Moens

Vasileios Mezaris is a Senior Researcher (Researcher B) with the Information Technologies Institute / Centre for Research and Technology Hellas, Thessaloniki, Greece. He received his bachelor’s and Ph.D. in Electrical and Computer Engineering from the Aristotle University of Thessaloniki, Thessaloniki, Greece, in 2001 and 2005, respectively. His research interests include image and video analysis, content- based and semantic image and video retrieval, event detection in multimedia, machine learning for multimedia analysis, application of image and video analysis technologies in specific domains (medical images, ecological data). He is the co- author of 29 papers in refereed international journals, 13 book chapters, two patents and more than 100 papers in international conferences. He has participated in many European and National Projects, having leading roles in several of them. Currently, he is the Research Manager of the FP7 IP project “ForgetIT: Concise Preservation by combining Managed Forgetting and Contextualized Remembering”. He is an Associate Editor for the IEEE Transactions on Multimedia and a Senior Member of the IEEE.

Lecture: Multimedia Analysis and Retrieval

Stefano Mizzaro (University of Udine, Italy) is Associate Professor at Udine University since 2006. He has been university researcher (assistant professor) from 2000 to 2006. His current research interests include Information Retrieval (IR), digital libraries and scholarly publishing, and mobile contextual information access. In the last 20 years he has specifically focused on several aspects of IR evaluation: user evaluation, novel effectiveness metrics, and mining of test collection data. He published more than 100 refereed papers, several as a single author, received two international awards, for two best papers, and authored two books on Java programming. He had an active role in several research projects at regional, national, and European level. In December 2013 he obtained the full professorship habilitation in Italy. In 2014 he has spent a sabbatical year at RMIT University in Melbourne.

Lecture: Axiometrics – An axiomatic approach to evaluation metrics

Effectiveness evaluation is of paramount importance in Information Retrieval (IR). The effectiveness metric being used is a fundamental parameter in any evaluation, and metric choice is neither a simple task, nor it is without consequences: an inadequate metric might mean to waste research efforts improving systems toward a wrong target. The problem is exacerbated by the large number of metrics existing, more than one hundred when counting the system-oriented metrics only. In this talk I will introduce some of the most common IR metrics and propose an axiomatic approach to metrics. I will then present a general framework based on Measurement theory that can be used to express axioms, thus making explicit the desirable metric properties, and will hopefully allow to better understand them.

Lecturer: Professor Stefano Mizzaro

Evangelos Kanoulas is an Assistant Professor at the University of Amsterdam. His expertise lies in the fields of information retrieval and text mining, with specializations in experimental design, evaluation methodology, and statistical analysis. Prior to joining the University of Amsterdam he was a postdoctoral research scientist in Google. In 2010 Dr. Kanoulas was awarded the Marie Curie Fellowship to explore the efficient and effective training and evaluation of information retrieval systems, as a postdoctoral research scientist at the University of Sheffield, UK. Evangelos has extensively published his work in top-tier conferences in the field, including SIGIR, CIKM, ECIR, and VLDB. He has written over 40 peer-reviewed journal articles and conference papers, which have received over a 1000 citations to this date. He has served as a program chair for the Conference and Labs of the Evaluation Forum (CLEF) in 2014, and Information Retrieval Facility Conference (IRFC) in 2013. Since 2007 together with others he has proposed and organized numerous search benchmark exercises under the umbrella of the Text Retrieval Conference (TREC), funded by the US National Institute of Standards and Technology, all of which led to large-scale testing collections to foster research and development in information retrieval. Since 2014 he is a member of the steering committee of the Conference and Labs of the Evaluation Forum (CLEF), the European counterpart of TREC

Lecture: Experimental design for collection-based comparative evaluation of search engines

Information retrieval effectiveness evaluation typically takes one of two forms: batch experiments based on static test collections, or online experiments tracking user’s interactions with a live system. Test collection experiments are sometimes viewed as introducing too many simplifying assumptions to accurately predict the usefulness of a system to its users. As a result, there is great interest in creating test collections that better model the variability encountered in real-life search scenarios. This includes experimenting over a variety of queries, corpora or even users and their interactions with the search results. In this talk I will discuss different ways of incorporating user behaviour in batch experimentation, how to model the variance introduced to measurements of effectiveness, and how to extend our statistical significance test arsenal to allow comparing search algorithms.

Lecturer: Dr. Evangelos Kanoulas

Lecturer: Dr. Barla Cambazoglu

Large-scale retrieval systems are indispensable tools for accessing the information available in the Web. In practice, there are two fundamental challenges faced by these systems: i) achieving high effectiveness when serving user queries and ii) doing this efficiently. In this context, effectiveness refers to understanding users’ information needs, often expressed by a few query terms, and providing high-quality search results that satisfy these needs. Efficiency refers to the speed at which a retrieval system is able to respond to search requests and its capability to operate under heavy query workloads. This tutorial aims to give an overview of the techniques employed by the state-of-the-art web retrieval systems to tackle the above-mentioned effectiveness and efficiency challenges. The main body of the tutorial is accordingly composed of two parts (they are roughly balanced in terms of length). In the first part of the tutorial, we first take a system-centric view and discuss document indexing and query processing in web retrieval systems (covering many technical issues that may affect the quality of generated search results, such as spam filtering, deduplication, ranking, result diversification, snippet generation). We then take a user-centric view and discuss various evaluation metrics employed for estimating user satisfaction with search results. In the second part of the tutorial, we shift our focus to efficiency and scalability issues in web retrieval. In particular, we present alternative indexing and search architectures as well as some specific optimizations, such as index compression, skipping, early termination, and result caching. Moreover, we briefly talk about the implications of efficiency improvements on users’ engagement with the retrieval system.

Berkant Barla Cambazoglu received his BS, MS, and PhD degrees, all in
computer engineering, from the Computer Engineering Department of
Bilkent University in 1997, 2000, and 2006, respectively. He has then
worked as a postdoctoral researcher in the Biomedical Informatics
Department of the Ohio State University. He is currently employed as a
senior researcher in Yahoo Labs, where he is heading the web retrieval
group. He has many papers published in prestigious journals including
IEEE TPDS, JPDC, Inf. Syst., ACM TWEB, and IP&M, as well as papers and
tutorials in top-tier conferences, such as SIGIR, CIKM, WSDM, WWW, and
KDD.

Lecture: Effectiveness and efficiency issues in web retrieval systems

Lecturer: Dr. CARLOS CASTILLO

Social software enhances or intermediates communications through the Internet. These communications, particularly the one-to-many ones, can be analyzed for a number of applications in domains such as economics, public policy and health. For these applications to be useful, at least two conditions have to be met. First, the data should be processed in real time or with low latency: this creates important algorithmic challenges in areas such as event detection, online classification/clustering, credibility/veracity assessment, among others. Second, the application designer must be well acquainted with the application domain, which is usually achieved by working in interdisciplinary teams. The lecture will introduce methods for processing social media data exemplifying through concrete applications, particularly in the areas of emergency response and disaster relief.