Αναλυτικές πληροφορίες για Υποτροφίες 2019 – 2020
Institutions: Rice University & NCSR Demokritos
Field: Artificial Intelligence / Robotics
Considering human behaviour ambiguities in robot motion planning
The goal of this PhD project is to transform the way robots collaborate with humans by enabling robots to take appropriate action even when human behaviour is ambiguous. Possible research directions include strategies for interpreting ambiguities given different modalities of human behaviour, e.g. speech and movement, methods for accumulating past experiences and resolving ambiguities based on the environmental context and previous experience and methods grounding such ambiguities into feasible robot actions.
Why it is important: Human behaviour can often have multiple interpretations, especially if each behavioural modality is treated in isolation. However, if multiple modalities are taken into account along with environmental context and history, such behavioural ambiguities could be eliminated to a degree that allows robots to plan feasible alternative actions and choose the most appropriate among them.
Institutions: Houston University & NCSR Demokritos
Field: Artificial Intelligence/ Machine Learning
Deep Learning for the analysis of multimedia content
The aim of this PhD thesis is to adopt state-of-the-art machine learning methodologies focusing on deep supervised and semi-supervised learning and transfer learning, for content-based analysis of multimedia. This requires leveraging knowledge from all types of information, such as faces and facial expressions, objects, voices, music, audio events, speech emotions, user comments, web pages, video text, low level visual cues, accelerometer measurements, EEGs and user/expert-generated annotations.
Why it is important: Modern multimedia databases can contain millions of files such as videos, digital music collections, image archives and measurements from wearables. High-level semantic descriptions of such multimedia content are crucial for several application domains such as hybrid recommender and personalization systems, multimodal security applications, environmental monitoring and health monitoring systems.
Field: Artificial Intelligence / Deep Learning / Big Data / Linked Data
Towards using linked data for interpreting deep learning models
The proposed PhD thesis will extend current state-of-the-art research at the intersection of big data and deep learning, with a focus on the wider security domain. It will make use of heterogeneous big data, for example earth-observation and weather data, to define predictive models based on deep learning. In addition, it will define novel semantic structures for linking inputs and human concepts to the latent features of these models, eventually creating knowledge. This work will build both on advances in deep learning as well as on metadata schemas and ontologies.
Why it is important: Despite the widespread availability of big data and analytics, artificial intelligence systems to predict adverse events remain underutilised. This is in part due to the difficulty to associate learned features with human factors, which can be communicated and argued about. One such example would be to interpret the role of specific weather patterns in a classification outcome.
Field: Artificial Intelligence/ Text Mining / Bioinformatics
Extracting knowledge from multiple modalities in biomedical publications
In this PhD thesis we will study methods and algorithms for extracting, analysing and fusing knowledge from multiple modalities in biomedical publications.
Why it is important: Biomedical text mining deals with the analysis of biomedical text, such as publications. Most of the work done so far in this area examines text in isolation, ignoring the non-textual parts of the publications (like tables and images), which are often very knowledge-rich. Potential applications of such a text mining process include, among others, biological semantic indexing and/or semantic Question Answering.
Field: Artificial Intelligence / Natural Language Processing / Argument Mining
Content Analysis in natural language texts
This PhD thesis aims to adopt state-of-the-art machine learning approaches on tasks related to the identification of communication and rhetoric devises on natural language texts. The field of application can range from a relatively new and exciting challenge in corpus-based discourse analysis, argument mining, which tries to automatically identify human reasoning in unconstrained natural language texts, to the identification of elements of populistic rhetoric, hate speech or types of misinformation (such as fake news), according to the interests of the applicant. However, the focus will remain on how to automatically detect these linguistic devices in texts through the use of machine learning. The PhD candidate will be able to explore theories developed in the humanities for describing relevant rhetoric devices and approaches, based on machine learning and statistical techniques from computational linguistics.
Why it is important: Content analysis of texts and especially discourse analysis has recently emerged as a research area that can help in the analysis, understanding and comprehension of a wide range of societal phenomena, as these are expressed in the social web and especially in user generated content that is ubiquitous, mainly driven by the constant increase in the use of social media. This largely uncontrolled means of communication has brought the opinion of citizens to the forefront, but at the same time it has facilitated the spread of misinformation, “fake news”, conspiracy theories, populistic rhetoric, hate speech, and has amplified public opinions that are ill-founded, misinformed, misguided and in error; elements that can damage, rather than benefit, the society. As a result, the analysis and automated identification of such communication devices in texts has become a necessity with increasing areas of applicability.
Field: Artificial Intelligence / Machine Learning
Knowledge transfer for rapid adaptation of machine perception to new environments
This PhD project aims at unsupervised or weakly supervised machine perception that is able to generalize sensory input into a library of recognizable objects classes. Such classes are expected to accurately reflect the natural segmentation of a scene into objects and their inherent attributes, as opposed to circumstantial co-occurrence of parts and transient or accidental attributes. The PhD candidate is expected to experiment with transfer learning, learning with privileged information and, in general, machine learning methods that exploit the differences observed between different environments in order to accurately separate universal generalizations form circumstantial correlations.
Why it is important: There are several applications where a system’s ability to collect new data so that it can adapt, depends on the system’s ability to perform adequately almost immediately after deployment. Examples range from commercial apps where early adopters need to remain engaged long enough to generate adaptation data to planetary exploration where early mistakes can incapacitate the system without the possibility to repair.
Institutions: Dalhousie University & NCSR Demokritos
Field: Big data/ Machine Learning / Knowledge representation
Big data analysis for precision medicine
This PhD thesis aims to introduce methods and algorithms to bring together disparate data sources (e.g. scientific literature, demographic, cognitive, images, genomics, EHRs of patients) and fuse the knowledge that will be extracted, to help realize the vision of precision medicine.
Why it is important: Large amounts of heterogeneous medical data have become available nowadays. This introduces an opportunity and a challenge, namely to bring together disparate data sources and fuse the knowledge that will be extracted. Such integrated knowledge could prove particularly valuable for improving the diagnosis and treatment of various diseases. Innovative data analysis and knowledge extraction methods could even allow the design of effective personalised treatments, in the wider context of precision medicine. While data are becoming ever easier to obtain, they are difficult to process using common database management tools or traditional data processing applications. The challenges include capturing, storing, integrating, searching, sharing, and analysing.
Field: Artificial Intelligence / Data Mining / Big Data
Data mining for forecasting anomalies in maritime data
The main objective of this PhD proposal is to develop data mining models that are able to anticipate anomalous deviations in the maritime domain, combining vessel trajectories with other useful dynamic and static information, such as weather forecasts, bathymetry data and geographical areas of interest (such as protected areas). The work to be carried out will involve 3 main questions/topics: (1) which properties of maritime data should be used/derived in order to facilitate the task of detecting and forecasting anomalies; (ii) which methods should we use to face the fact that anomalies will be rare in these data and thus the forecasting models will have difficulties in facing this imbalanced distribution; (iii) how to properly obtain forecasting models for maritime data streams, and which methods should be used to estimate the performance of these models, given the existing dependencies among the data observations.
Why it is important: Trajectory data is becoming ubiquitous due to the widespread use of devices with geo-reference capabilities and also the profusion of cheap sensors. All these technological advances have led to a considerable growth in the number of data sets with information on the trajectories of different moving objects, like humans, animals or ships. In many application domains, being able to forecast anomalous deviations is of very high importance. In particular, being able to anticipate these deviations with sufficient lead time to allow for preventive actions to be put in place can be highly rewarding. Examples include monitoring elderly people with critical health conditions living at home, monitoring trajectories of endangered marine species, or monitoring the behavior of ships in the Ocean.