Reserved topic scholarships - 2019 - 2nd call
Department of Information Engineering and Computer Science
The project aims at studying, modeling and developing new techniques and systems for Human-Machine conversation.
In particular, Question Answering systems will be developed and adapted to automatically carry out conversation with users. The research will mostly based on the design and application of machine learning models based on neural networks.
Contact: alessandro.moschitti [at] unitn.it
A2 - Quantum Estimation of Distribution Algorithms for Network Optimization Problems (Q@TN Project) (1 grant)
Quantum Annealing, the process through which specialized hardware can, over repeated samplings, probabilistically approximate the optimal solution to a given optimization problem, is a powerful tool for solving a number of problems which can be too difficult to solve using classical approaches. However, while the number of qubits available in modern quantum annealing machines is constantly increasing, the size of the problems that quantum annealers can tackle is still somehow limited.
Evolutionary Algorithms (EAs), a kind of optimization algorithms that iteratively "mutate" and "recombine" a certain number of solutions, have on the other hand proven successful on several hard-to-solve problems, even at large scales. Typically, EAs use a large number of solutions. However, a class of EAs called Estimation of Distribution Algorithms (EDAs) is characterized by a limited amount of memory as they model and evolve a "compact" probability distribution from which solutions are iteratively sampled, rather than using an actual "population". Therefore, a parallel between Quantum Annealers and EDAs can be drawn.
The primary goal of this PhD project is then to design and implement EDAs on quantum annealers. Firstly, the performance of quantum EDAs will be compared with that of traditional ones. Secondly, the developed techniques will be applied to challenging network optimization problems, such as the optimal routing of traffic demands under the constraints of packet- or circuit-based technologies, or the design of Quantum Key Distribution sub-networks at minimal cost.
Contact: giovanni.iacca [at] unitn.it
A3 - Hybrid Quantum-Annealer Algorithms for Tabu search and Data-Driven Computation (Q@TN Project) (1 grant)
The current availability of limited quantum hardware requires to devise hybrid quantum-classical algorithms that take advantage of the existing hardware and overcome their limitations by combining classical and quantum computation. One of the existing hardware, the D-Wave quantum annealer, solves optimization problems however its architecture suffers from limitations in the actual encoding of the target functions. A novel quantum-classical technique is based on a local search where already-visited solutions are penalized to avoid a redundant exploration of the solution space (so-called tabu search paradigm). Tabu search is implemented in the quantum annealer by a sequence of re-initializations of the Ising Hamiltonian of the qubit network iterated within the hybrid quantum-classical structure. An algorithm of Quantum Annealer Tabu Search (QATS) has been proposed in a recent paper (https://arxiv.org/abs/1810.09342).
The PhD candidate will implement and test hybrid quantum classical algorithms that can run on a quantum annealer. Initially, the PhD program will focus on the existing QATS algorithm. The expected result is a complexity-characterized, implemented and empirically evaluated tabu search quantum- classical algorithm for quantum annealers. In a second phase the PhD activity will extend to optimization of more general target functions and also towards data representation into the quantum annealer for data-driven computation.
The candidate should be familiar with:
- Programming (Python experience preferred);
- Linear algebra;
- Algorithm Theory.
Experience in the following areas would also be useful, but can be acquired as part of the PhD:
- Quantum Computing (especially pertaining quantum annealing);
- Optimization algorithms (particularly tabu-search);
- Quantum Mechanics;
- Ising model.
Contact: enrico.blanzieri [at] unitn.it
The project plans to investigate modern machine learning approaches for behavior analysis in video and event prediction. Specifically, the project will cover two significant computer vision topics, i.e., video-surveillance in urban scenarios and behavior understanding of people during a museum visit. We believe that the effort in computer vision should go behind scene segmentation and target detection and tracking towards a predictive capability. As such the project will explore the capabilities of algorithms and novel solutions in visual artificial intelligence to predict different types of events. Moreover, the project will predict the effect of collaborative interactions between people and between people and the context in which the analysis is taking place. State-of-the-art deep networks, such as Generative Adversarial Networks, Recurrent Neural Networks and Autoregressive Autoencoders will be used together with standard Convolutional Networks in order to define an effective predictive visual intelligence for new images and temporal reasoning. The research will be supported by the Artificial Intelligence for Cultural Heritage (AI4CH) project as part of the joint Italo-Israeli laboratory financed by MAECI and by the PRIN 2017 PREVUE (PRediction of activities and Events by Vision in an Urban Environment) project (CUPn: E64I19001120001).
Contact: niculae.sebe [at] unitn.it (body: niculae.sebe%40unitn.it)
Research will focus on protocols and techniques to improve the data rate, energy efficiency, reliability and scalability of solutions based on ultra-wideband (UWB) radios. These radios can be used for communication but also for accurate distance estimation (ranging) and therefore localization. The proposed research will focus on both aspects, communication and localization, as well as their integration.
Contact: gianpietro.picco [at] unitn.it
The research activity will aim at investigating and developing novel techniques, methodologies and support tools advancing the state of the
art in Optimization Modulo Theories (OMT). In particular:
- extending OMT with optimization with non-linear constraints and objectives;
- extending OMT for dealing with Constraint Programming in MiniZinc.
All such tools will be implemented inside OptiMathSAT OMT solver developed in Trento (http://optimathsat.disi.it/) o top of the MathSAT.5 SMT platform (http://mathsat.fbk.eu/), and provided as an API.
See also http://disi.unitn.it/rseba/PHD-OMT19-recruit.pdf for more details.
Contact: roberto.sebastiani [at] unitn.it
The goal of this thesis is to develop a platform for learning from personal data streams, and to use it in large scale real world experiments. The platform will have to be able to deploy state of the art machine learning algorithms in a robust manner. The data collected will be used to inform the machine learning algorithms which, in turn, will have to create user friendly services. This thesis will be part of the PRIN project Delphi (CUP: E64I19001100001).
Contact: Fausto.Giunchiglia [at] unitn.it (body: Fausto.Giunchiglia%40unitn.it)
The topic of this PhD thesis will cover inter-disciplinary topics at the boundary between computer science, social sciences and privacy. The goal is to develop a general methodology for the design of person-centric data collection pilots and for the follow-up exploitation of these data in experiments. the methodology will have to be GDPR compliant and make data available according to the state-of-the-art best practices in data management This thesis will be part of the PRIN project Delphi (CUP: E64I19001100001).
Contact: Fausto.Giunchiglia [at] unitn.it (body: Fausto.Giunchiglia%40unitn.it)
The research activities related to this PhD position are related to planetary radar sounders that are instruments for the study of the subsurface of celestial bodies. These radars operate from satellite platforms and acquire data related to the subsurface. The research will be focused on the development of a new generation of data analysis techniques for planetary data that exploit the most recent developments in the framework of artificial intelligence and deep learning. The activity will related to the definition, design, implementation and validation of:
1) Radar signal processing algorithms based on deep learning techniques;
2) Data analysis techniques based on artificial intelligence for the automatic extraction of the semantic from the data;
Part of the research will be related to the activities in progress on the Radar for Icy Moon Exploration (RIME) developed in the framework of the JUpiter ICy moon Explorer (JUICE) mission of the European Space Agency. RIME is developed under the leadership of the University of Trento in cooperation with Italian Industry and the Jet Propulsion Laboratory (JPL) in USA with the funding of Italian Space Agency and NASA, respectively. RIME is a radar sounder defined to study the geology, the geophysics and the possible presence of water in the subsurface (up to 9 km) of the Jupiter icy moons, i.e., Ganymede, Europa and Callisto. For more information on RIME refer to https://rslab.disi.unitn.it/rime/
The PhD student will work at the Remote Sensing Laboratory of the University of Trento.
For more information on the activity of RSLab refer to: http://rslab.disi.unitn.it/
Contact: lorenzo.bruzzone [at] unitn.it
Fondazione Bruno Kessler (FBK)
Representation of context has been one of the main approaches in the knowledge representation and reasoning area for the management of large knowledge bases: here, modelling and reasoning on knowledge is relativized with respect to the contexts (situations, set of hypothesis) in which it is supposed to hold. Introducing forms of non-monotonic and default reasoning in logic-based contextual frameworks is a challenging research issue which, in particular, plays a role in the propagation and inheritance (with overriding) of knowledge across hierarchies of contexts. The objective of this PhD will be to investigate the possibilities for extending the theory and implementation of the current methods for representing default information in contextual frameworks and explore their applications.
Contact: bozzato [at] fbk.eu
Many moders software systems are highly parameterized in order to take into account various operational conditions. Testing such problem is extremely difficult, given the astronomic number of combinations corresponding to all the parameter assignments. The challenge is to maximize the coverage and the confidence provided by the generated test cases without incurring in the limitations deriving from the parameterization space. The thesis will investigate the definition of novel testing techniques starting from the integration genetic algorithms for software testing and test case generation approaches based on formal methods.
Contact: cimatti [at] fbk.eu
Space and avionic systems are reaching an unprecedented degree of complexity. The process of safety analysis attempts to characterize the likelyhood of faults and failures, and to assess the effectiveness of the adopted mitigation measures. Unfortunately, traditional techniques are becoming ineffective, unable to deal with large scale systems.
This thesis will investigate novel methods for safety analysis, based on the adotpion of formal models of system and software (nominal and faulty) behaviours. Particularly interesting are the analysis of iming aspects in the propagation of multiple faults to failures and errors, the ability to explain the causality of propagation, and the definition of techniques for on-the-fly fault detection, isolation and recovery policies.
Contact: cimatti [at] fbk.eu
The current trend of automation and data exchange in manufacturing technologies allows to gather huge amounts of information. One of the challenges is to assess the condition of an equipment and foresee a failure before it actually occurs, so that fine-tuned, predictive maintenance can be applied. Although machine learning techniques can be used to tackle this problem, a pure data-driven approach is bound to re-discovering background information that is already available, such as operation modes and procedures, failure modes, and maintenance records. The thesis will investigate the integration of symbolic background knowledge within modern machine learning approaches, in order reduce the training time and improve the accuracy of predictions.
Contact: cimatti [at] fbk.eu
We are looking for candidates willing to develop novel methodologies based on machine learning, deep learning pattern recognition and artificial intelligence for information extraction, classification and target detection and change detection in radar images. The candidate will be requested to deal with both Synthetic Aperture Radar (SAR) images acquired from Earth Observation satellite missions, and sub-surface radar data from airborne Earth Observation missions and satellite planetary exploration missions. The latter activity is developed in the framework of the Radar for Icy Moons Exploration (RIME) payload on board of European Space Agency (ESA) JUpiter ICy moons Explorer (JUICE). Besides the requirements established by the rules of the ICT school, preferential characteristics for candidates for this scholarship are: • master degree in Electrical Engineering, Communication Engineering, Computer Science, Mathematics or equivalents; • knowledge in pattern recognition, image/signal processing, statistic/remote sensing/radar
Contact: bovolo [at] fbk.eu
University of Trento - Department of Information Engineering and Computer Science - Consiglio Nazionale delle Ricerche (CNR)
Deep Learning approaches, powerful Big Data Analytics processes based on multi-layer neural networks, are gaining popularity even in the clinical imaging field, thanks to their excellent performances in processing and classification of medical imaging for diagnostic purposes.
So far, Deep learning approaches have been only marginally applied on ultrasound, despite its non-invasiveness, low cost, wide availability and extensive use in several clinical settings; all these advantages make ultrasound imaging particularly suitable for the definition and development of innovative diagnostic systems employable on large populations. Furthermore, Artificial Intelligence approaches may help to overcome the limitations of ultrasound imaging, such as the operator-dependency and the low reproducibility.
Lung ultrasound imaging shines among the others as a prime example of potential application of Deep Learning methods, given the artefactual nature of the diagnostic features and the lack of a model that is sufficiently accurate for the quantitative interpretation of the data.
In view of these considerations, the present project will be aimed at creating a multi-application platform which uses Deep Learning approaches for the processing and classification of ultrasound images characterizing different pathological conditions. Results will also be compared with state-of-art model based approaches and possibly explored for the definition of new models and image processing algorithms.
This platform will take as input ultrasound images acquired with different ultrasound modalities following acquisition protocols specifically defined for each application. The employment of Deep Learning techniques together with the definition of specific acquisition protocols is aimed at overcoming the ultrasound-related limitations (e.g. its strong operator-dependent nature) and the implementation of a precision medicine approach.
In conclusion, the proposed framework will represent a powerful tool to be used in clinical practice for addressing important clinical needs, facilitating prevention programs and diagnostic strategies which will turn out to be appropriated, safe and aimed at resources optimization.
Contact: libertario.demi [at] unitn.it