A. Social Media and Digital Preservation theme

1. Social Sensor: Through this project, the students will become familiar with a distributed architecture for monitoring, analyzing and indexing social media content from multiple social networks, as well as with some fundamental theoretical principles underpinning modern online social networks. The base architecture has been developed by the SocialSensor project and leverages widely used open technologies, such as MongoDB, Solr, Hadoop, and Nutch. Furthermore, the students will become familiar with two approaches for social network analysis, i.e.: community discovery and influence calculation in heterogeneous social networks. The point is the discovery of natural groupings of nodes, and within each groups the most prominent nodes.

Supervisors: M. Schinas, K. Iliakopoulou, S. Papadopoulos, D. Vogiatzis

2. ARCOMEM: Affect recognition and Sentiment Analysis (SA) refer to the automatic identification and the assessment of human behavior. Concerning the affect modeling, there are three main approaches: categorical, dimensional and appraisal. The main postulate of the categorical approach is the existence of a small set of universally recognised emotions and the problem is then formulated as classification of the relevant segments/units. More recent work considers the multimodal, continuous and complex nature of human interactions focusing on continuous prediction of emotions. The applied methodologies exploit several feature sets corresponding to the different modalities and communicative scenarios. In this project, students will get acquainted with the latest trends and some of the major challenges in the field. Based on available datasets and a rich set of extracted features, students will deal with the feature selection task and investigate the most suitable machine learning techniques for the recognition.

Supervisors: H. Papageorgiou, D. Spiliotopoulos

3. Argument Extraction and Ontology Modelling A mind map is a diagram used to visually outline information. A mind map is often created around a single word or text, placed in the centre, to which associated ideas, words and concepts are added. Major categories radiate from a central node, and lesser categories are sub-branches of larger branches. Categories can represent words, ideas, tasks, or other items related to a central keyword or idea. From a philosophical perspective, ontology is the study of the nature of being, becoming, existence, or reality, as well as the basic categories of being and their relations. In Computer Science, the term ontology stands for a formal explicit specification of a shared conceptualization. The depiction of an ontology through mind-maps is, thus, a way to explore the relations among the different concepts in a formal hierarchical specification. The first part of this project will explore ways for effective visualisation and exploration of ontology-based representations for non ICT-skilled stakeholders, through simple Web User Interfaces. Argumentation is a branch of philosophy that studies the act or process of forming reasons and of drawing conclusions in the context of a discussion, dialogue, or conversation. Being an important element of human communication, its use is very frequent in texts, as a means to convey meaning to the reader. As a result, argumentation has attracted significant research focus from many disciplines, ranging from philosophy to artificial intelligence. The second part of this project will explore ways for identifying arguments on textual data (texts from social media), and associate these arguments to elements of a predefined semantic model or ontology, which would be created by the first part of the project.

Supervisors: G. Petasis, E. Kanoulas, S. Dalianis, V. Tountopoulos, G. Kiomourtzis

B. Cognitive Systems and Ineractive Robotics theme

4. Robust Distributed Human-Machine Cognitive Systems Traditional cognitive systems are built by custom-designing a system that fulfills requirements, subsequently purchasing hardware and software components to account for the processing, sensing, and actuation needs of the system, and then physically building a system which is usually dedicated to performing a specific function. Now, there are multiple shortcomings to this approach. First, traditionally cognitive systems are made up of machine-only components; however, there are tasks, for example certain pattern recognition or sensing tasks, where human performance is far superior to machines; or locations where only human sensing might be available. Second, very seldom are the components of traditional cognitive systems reusable or interchangeable; and it is usually impossible to replace underperforming or malfunctioning components on the fly, without interrupting the operation of the cognitive system. To remove these limitations, in this project we will explore the idea of creating a human-machine cloud, where the sensing, processing and actuation services needed for constructing cognitive systems are offered by both machine and human providers. We take our inspiration from the way that cloud computing has changed the computing needs of companies: instead of buying and maintaining their own computer systems, they outsource their computing needs by buying computing services from the cloud. This project will serve as proof-of-concept, building the basic tools for our proposed framework, demonstrating that cognitive systems can be built by fusing dynamically both human and machine components and implemented through specially selected services procured and operated through the human-machine cloud.

Supervisors: I. Vetsikas, N. Mavridis

5. Multimodal Speaker Recognition and Localization Speech detection and diarization, synchrony detection, source separation, as well as multi-party speech recognition can be greatly improved by prior knowledge of the number of speakers in the scene; and, even more so, by prior knowledge of speakers’ location in order to, for example, predict the location of the speaker's mouth region in realistic situations where the speaker or speakers are not situated in a steady frontal head pose. In the field of robotics, the similar tasks of object segmentation and recognition are often treated outside of, or in fusion with, the audio-visual domain. Planar range data from laser scanners are mainly used for obstacle avoidance but have also been used to detect human and other motion patterns; 2D depth maps from triangulation sensors are typically used to recognize human figures and gestures in gaming and other human-computer interfaces. Given the above, the project will explore ways in which research on human pattern recognition using laser range data and depth maps can be used to improve the robustness of audio-visual speech processing technologies. Furthermore, the project will investigate ways to achieve autonomous adaptation to dynamic environments by transferring knowledge gained in one modality in order to improve recognition in another.

Supervisors: T. Giannakopoulos, S. Konstantopoulos

6. Human-Robot Interaction and Machine Perception Traditional Robotics was mainly concentrated on industrial and manufacturing application, where robots were usually performing automated repetitive tasks with no human intervention, except for an expert programmer while the robots was being programmed, and often also a human supervisor. But as robots move on to numerous other application domains, and as they start playing the roles of receptionists, museum robots, or household helpers, among others, the need for fluid interaction with humans becomes vital. However, achieving such a state of affairs requires overcoming several technical obstacles; and also often necessitates an interdisciplinary approach where artificial intelligence, psychology, linguistics, sociology and even philosophy become highly relevant. Also, several technical areas become deeply intertwined with the goal of fluid human-robot interaction: from machine learning, to perception, to natural language processing, affective computing, social networks and beyond. Most importantly, the state-of-the-art of the field is advancing rapidly, opening up not only a wide array of futuristic and highly useful applications, but also creating highly exciting problems as well as tools for researchers to solve, in a quest that will not only create intelligent machines and hopefully make our lives more enjoyable, but which will also let us get a deeper view of ourselves humans, by understanding and examining cognition and interaction in an unprecedented level of detail and clarity, In this summer school, students will embark upon an exciting project on Human-Robot Interaction and Machine Perception. Through this project, the students will become familiar with multiple aspects of this exciting interdisciplinary field, and gain highly valuable hands-on real-word experience.

Supervisors: N. Mavridis

C. Others

7. BioTransfer Bioinformatics concerns a wide spectrum of biological research, including genomics and proteomics, evolutionary and systems biology, etc. Data mining in bioinformatics targets the investigation of learning statistical models to infer biological properties from new data. Various techniques such as supervised learning and unsupervised learning have been developed. A typical assumption in biological data mining is that a sufficient amount of annotated training data is required, so that a robust classifier model can be trained. Sometimes this data can only be obtained by paying a huge cost. In response to the above problem, various novel machine-learning methods have been developed. Among them is the transfer learning framework, which refers to a new machine learning framework which reuses knowledge from other domains. Transfer learning aims to extract knowledge from some related domains where the data is annotated, but it cannot be directly used as the training data for a new domain. In this summer school project, the students, based on different biological datasets, will investigate a new transfer learning approach. Through this project, the students will become familiar with transfer learning techniques, as well as biological data.

Supervisors: G. Paliouras, A. Krithara

8. COLLONA: Collaborative Ontology Alignment The numerous automatic approaches for producing mappings between ontologies and schemas have produced significant and valuable results. However, and especially in cases where there is a demand for highly precise mappings, these methods have certain drawbacks and limitations. It is, therefore, essential for human agents to get involved in the matching process. This semi-automatic approach to the problem presents its own issues. It has been shown that humans struggle with handling large ontologies, remembering their decisions, providing evidence/arguments for their choices etc. A semi-automatic system for ontology mapping has to take into consideration all of the above parameters and be able to provide the means for presenting ontologies in an understandable way to users, make clear the results of automatic alignment methods, give the users the ability to understand the rationale behind automatic suggestions and provide a straight-forward way for humans to correct the proposed mappings and provide their own suggestions. The core objective of the project is to produce an extensible framework for visualizing mappings of ontological entities (both at the conceptual and the instance level) and enabling human users to edit/ correct the proposed mappings or submit mapping suggestions of their own.

Supervisors: P. Karampiperis, A. Koukourikos, G. Vouros