Calls 2021

Dialogue agents usually work in limited domains with clear and well defined policies, with little adaptation capabilities
to the contextual and social situations. In this scenario it has been shown that endowing neural models with a consistent user/machine
profile is important to produce more coherent and natural conversations. Still, the representation of such profiles is very limited, usually based on simple facts. The goal of this PhD Thesis is investigating end-to-end approaches and generation strategies such as guided decoding, in combination with large pre-trained language models and external knowledge, in order to improve the naturalness of
dialogues. During the PhD we will always be mindful of the societal impact of the technologies we develop.

In recent years there has been a growing interest for psychoeducational interventions delivered by means of mobile applications and personal assistants to support self-care of patients, including those coping with chronic conditions. Although the validity of psychoeducation has been proved repeatedly by previous research, the design of effective behavioral intervention technologies for virtual coaching in the area of healthy coping remains a challenge. The aim of the PhD project is to investigate key features of smart coaching solutions for healthy coping interventions that are engaging to use by patients and produce effective outcomes from a clinical perspective. The ideal candidate will be strongly motivated in developing design skills in the field of behavioral intervention technologies and conversational agents for applications in healthcare. The PhD position is offered in co-tutoring between the Digital Health Lab of FBK and UNIPD.

With this phd, we want to investigate how recommendation systems can take advantage of integrating background knowledge expressed in some symbolic form as e.g. logical formulas with standard numeric optimization method, based on machine learning approaches.
In particular we want to investigate how the state-of-the art methods for recommended systems based on embeddings can be complemented with logical reasoning on formulas that expresses structural properties and constraints about properties and relations between users and items.

The Limadou project gathers some Italian institutions participating in the China Seismo Electromagnetic Satellite (CSES) mission. CSES consists of a constellation of satellites, designed to pursue the deepest campaign of observation of the ionosphere. One of the most important scientific goals of the mission is to look for correlations between transient phenomena in the ionosphere and seismic events. Among payloads, a set of particle detectors is devoted to the detection of charged particles trapped in the Van Allen Belts, to monitor the solar activity and to measure galactic cosmic rays of very low energy. The APP group of the Physics Department in Trento looks for candidates to a PhD programme on the analysis of the scientific data from the payloads on board the CSES-01 and those to be launched on board the satellite CSES-02 in 2022. The student will focus on time-series analyses and participate in the development of the event reconstruction software. These studies will be carried out using the most modern machine learning techniques for clustering and anomaly detection, using full information from CSES payloads. The activity will be carried out in collaboration with INFN-TIFPA, Fondazione Bruno Kessler and the Institute of the High Energy Physics of Beijing. Candidates familiar with the experimental techniques for the detection of charged particles in space are welcome, as well as basic knowledge of Machine Learning/Deep Learning is recommended.

Quantum Chromodynamics (QCD) is the theory of the strong nuclear forces. At ultrarelativistic energies the degrees of freedom are quarks and gluons and their interactions can be calculated with weak coupling methods. For sufficiently high energies, the gluon density becomes large leading to strong non-linear effects whose description is the goal of the Color Glass Condensate (CGC) effective theory. It is important to apply the latter for studying observables in the forthcoming Electron Ion Collider (EIC).

We propose to investigate the optimization of quantum simulations on trapped-ion quantum processors. The Ph.D. candidate will explore the use of quantum optimal control techniques to tailor ‘analog’ gates at the laser pulse level, as well as the optimization of ‘digital’ quantum circuits built on predetermined primitive gates. The study will identify the most effective methodology to translate near-term trapped-ion quantum computing into meaningful quantum simulations of microscopic systems.

Neuroimaging methods, like structural, functional, and diffusion MRI as well as MEG/EEG can be used to investigate the anatomical and functional connectivity of the brain. In this project, the candidate will pursue research on machine learning methods for neuroimaging data to study and characterize brain connectivity, with applications to longitudinal studies and clinical practice. The ideal candidate should have a mixed background in neuroimaging techniques and numerate disciplines, like computer science, engineering, physics, or mathematics. This project is in collaboration with the Division of Neurosurgery, S. Chiara Hospital, Trento (IT).

Deep neural networks (DNNs) have reached a prominent position amongst machine learning systems [4, 1] due to an increasing experimental evidence of their flexibility, expressivity, and effectiveness in addressing functional approximation problems. At the same time, a complete and satisfactory mathematical theory, explaining for instance how to optimally design and train a DNN on some specific task, is largely missing. It is worth noticing that typical DNNs need millions or billions of trainable parameters for achieving such outstanding performances, and must execute a proportional amount of arithmetic operations during the inference step. These characteristics make DNNs very demanding in terms of storage, memory, and energy consumption. This creates an obstacle towards their deployment on low-memory and low-power architectures such as embedded devices and microcontroller units (MCUs). For this reason, the use of quantized neural networks (QNNs), that is, of specific DNNs whose parameters take values in small, finite sets and whose activation functions have a finite range, allows to represent the operands with fewer bits with respect to standard DNNs, thus providing significant benefits with respect to digital hardware constraints [10, 5]. In this sense, QNNs have an enormous potential of applications, for instance in biomedical and environmental research, where in-situ analyses of real-time data could be effectively performed by low-powered & portable devices. At the same time, the lack of a rigorous mathematical theory in support of QNNs is even more striking, also due to the non-differentiability properties of such networks, so that a better comprehension of the mathematics involved seems crucial for their effective application. The objectives of the PhD project will be of both theoretical and applicative nature. The theoretical part of the project will consist in studying some mathematical aspects of the theory of DNNs, including: • techniques for layer-wise regularization & approximation of QNNs; • expressivity properties of suitable classes of QNNs;

The Ph.D. student will evaluate control measures against different infectious diseases, both retrospectively and prospectively, by developing mathematical models calibrated against observed epidemiological data and informed by other data relevant to the infection under study. The modelling approach will be tailored to the addressed problems and may include compartmental models, generative models, individual-based simulations as well as Bayesian approaches and will include scenario analysis to compare alternative intervention strategies.

Research activity conducted during the Ph.D. will focus on the development of epidemiological models informed by real-world data aimed at investigating the main determinants of the disease spread in humans. Envisioned approaches range from the study of mechanistic models mimicking the spatio-temporal transmission dynamics of infectious diseases to the use of bayesian approaches applied to detailed epidemiological records.

Interpretability has become a crucial requirement to support translation of reliable AI models into clinical practice. Throughout the graduate course, the candidate will explore how to make a DL model interpretable and reproducible, and she/he will test such methodologies comparing predictive models trained on integrated imaging (CT/PET/MRI or Digital Pathology), omics and clinic data, both publicly available and original, with the final goal of defining a framework or a pipeline leading from the available data to the explainable and repeatable models.

Cyber-Physical Systems (CPS) are ubiquitous systems that integrate computation, networking and physical processes. The correctness and dependability of their control software is critical in many high-assurance domains such as space, mobility, and energy. However, their design requires complex component-based hybrid models describing the continuous dynamics of the physical components and the discrete interaction with the control and monitoring components. The objective of this PhD research is that of advancing the state of the art in the formal verification of CPS control design models, integrating various techniques such as SMT-based model checking, contract-based design, abstract interpretation, simulation, test-case generation, and fault injection. The candidate will work on theoretical aspects of the problem as well as its practical applications in relevant case studies, drawn from the domains of aerospace, automotive, railways, and energy.

3D scene understanding using learning-based approaches is becoming largely employed in several application sectors including industry, automotive, surveillance and cultural heritage. Computational algorithms that are traditionally used to process 2D images, such as object detection, tracking and segmentation, are nowadays being successfully extended to process 3D data (e.g. point clouds). However, traditional 3D approaches rarely use the image content directly for their task, but often rely on mid-level representations (e.g. voxels, sparse point clouds) that disregard the rich context provided by images. Moreover, the availability of 3D data is limited because the effort for annotationing it is greater than its 2D counterpart. The goal of this PhD position is to advance the state of the art about 3D scene understanding by focusing on the aspects of learning-based approaches that can potentially leverage different sensor modalities and the lack of human-annotated data.

Cyber-physical systems are ubiquitous systems that integrate computation, networking and physical processes. The correctness and dependability of their control software is critical in many high-assurance domains such as space, mobility, and energy. However, their design requires hybrid models that combine continuous dynamics with discrete states and is typically verified with simulation-based testing. The objective of this PhD research is that of advancing the state of the art in the formal verification of hybrid system models, integrating various techniques such as SMT-based model checking, simulation, test-case generation, and fault injection. The candidate will work on both theoretical aspects of the problem, as well as its practical applications in relevant case studies, drawn from the domains of aerospace, automotive, railways, and energy.

Automated Planning is the problem of synthesizing courses of actions guaranteed to achieve the desired objective, given a formal model of the system being controlled. A class of problems particularly interesting for applications is temporal planning (also called planning and scheduling) where the discrete decisions of "what to do" are coupled with the problem of scheduling (deciding "when to do"). Unfortunately, planning and scheduling techniques suffer from scalability issues and are often unable to cope with the complexity of real-word scenarios, despite the plethora of approaches available in the literature. Recently, efforts such as Deepmind AlphaGO and OpenAI Five hit the headlines, with groundbreaking advancements in the field of reinforcement learning. These techniques are able to automatically learn policies to decide what to do in order to achieve the desired goal. However, they offer no formal guarantee and are not model-based. The research objective of this PhD scholarship is to investigate techniques that combine the formal guarantees offered by automated planning and scheduling with the performance and self-improving capabilities offered by recent advances in deep reinforcement learning to construct self-adaptive planners that can improve over time their performance on specific application scenarios.

The need to translate audio input from one language into text in a target language has dramatically increased in the last few years with the growth of audiovisual content freely available on the Web. Current speech translation (ST) systems are now required to be flexible and robust enough to operate in different application scenarios and diverse working conditions. On one side, the industry calls for key features like real-time processing, domain adaptability, extended language coverage and the capability to adhere to application-specific constraints (e.g. length or lip-synch constraints in the subtitling and dubbing scenarios). On the other side, the society calls for new efforts towards inclusiveness with respect to specific categories and groups (e.g. gender-sensitivity, customization to the needs of impaired users). Both dimensions (industry and society) face the orthogonal challenges posed by the variability of audio conditions (e.g. background noise, strong speakers’ accent, overlapping speakers). The objective of this PhD is to advance the state of the art in speech translation to make ST flexible and robust to these and other factors.

The increasing complexity of software systems calls for the development of new methods and tools to design and test software systems characterized by high variability from the point of view of the space of the possible functional configurations, the space of the release architectures, and of the aspects related to dynamic reconfiguration. The goal of this PhD thesis is that of exploring new approaches to the testing, verification and validation of these systems that involve the joint use of model-based and AI based techniques such as planning, machine learning and optimization.

Multi-Access Edge Computing leverages the network's edge to store and process data and applications locally, and provide fast reactions and efficient use of network and computing resources. Future communication systems and networks, both terrestrial and non-terrestrial, will increasingly require solutions based on AI/ML, virtualization and softwarisation techniques besides traditional communication technologies. The goal of this PhD Thesis is to design and explore novel approaches that leverage on the close cooperation with these domains from an overall system perspective.

In the context of the green deal transition 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, target detection and change detection in long and dense timeseries of remote sensing images.
The candidate will be requested to deal with both multi-/hyper-spectral images acquired by passive satellite sensors and Synthetic Aperture Radar (SAR) images acquired from active systems for Earth Observation. Copernicus data acquired by the new ESA Sentinels will be considered. The goal is to design novel methods able to use temporal correlation to model landcover behaviors, changes and trends for a better understanding of phenomena over the past and the future for a better modeling and understanding of climate change.
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, deep learning, image/signal processing, statistic/remote sensing, passive/active sensors.

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, target detection and change detection in radar and radar sounder images.
The PhD activity will be developed in the context of European Space Agency (ESA) space mission JUpiter ICy moons Explorer (JUICE) in the Jovian system. The candidate will be requested to deal with images acquired from active radar systems including Synthetic Aperture Radar (SAR) images and sub-surface radar sounding data from airborne Earth Observation missions and satellite planetary exploration missions. The activity amins in improving the understanding of subsurface structure and their impact on planetary body climate.
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, deep learning, image/signal processing, statistic/remote sensing/radar.

Machine learning can impact people when it is used to automate decisions in areas such as insurance, lending, hiring, and predictive policing. In many of these scenarios, several works have shown that training data can be unfairly biased against certain populations and groups, for example those of a particular race, gender, or sexual orientation. Since training data may be biased, machine learning predictors must account for this to avoid perpetuating or creating discriminatory practices. This Phd student will work on designing and implementing innovative approaches for fair and explainable machine learning and deep learning algorithms. The selected student will have the possibility of collaborating with the activities of the Human-centric Machine Learning program of the ELLIS society (https://ellis.eu/).

Gamification principles have proven to be very effective in motivating target users in keeping their engagement within everyday challenges, including dedication to education, use of public transportation, adoption of healthy habits, and so forth. School closures due to the COVID-19 pandemic and thus the sudden change in the management of the students' educational pathways has opened up the need for methods and digital systems able to support teachers in defining educational content and objectives for their classrooms and to keep students engaged in their training path. The goal of this PhD Thesis is to investigate approaches, techniques and tools to design and release educational digital systems for personalized and cooperative learning plans. This will be done exploiting AI techniques for adaptive gamification and will support teachers in the process of defining and monitoring dedicated learning plans for their students. At the same time, it will facilitate learning, will encourage motivation and engagement, will improve student’s participation and cooperation, and will stimulate students to expand their knowledge through dedicated learning plans and personalized feedback.

End-2-End models for speech recognition have been steadily improved recently, achieving performance comparable to state-of-the-art systems. This paves the way to the adoption of such solutions also for the extraction of semantically higher information, directly from the raw speech. This allows avoiding approaches based on the combination of speech recognition followed by text processing, with consequent propagation of errors from the intermediate stages. The goal of this thesis is to develop innovative end-to-end systems, eventually based on the transformer model and sequence-to-sequence learning, to address tasks like spoken language understanding, named entity recognition, intent classification and so forth. Ideally, at the end of the doctoral thesis, the candidate would have developed a neural audio processing front end that can be applied to a variety of semantic down-stream tasks.

In the context of dialogues with chatbots it has been shown that endowing neural models with a persona profile is important to produce more coherent and meaningful conversations. Still, the representation of such personas is still very limited, usually based on simple facts. The goal of this PhD Thesis is to make a step forward, trying to grasp more profound characteristics of human personality (such as opinions, values, and beliefs) to drive language generation of conversational agents in multiple domains and languages.

Conversational agents are designed to interact with users in multiple domains on several topics using natural language. Recently end-to-end systems have started to be tested to fight fake news and hate speech in single turn settings. Still, scaling to full dialogue interactions is a challenging topic, requiring world and domain knowledge together with a deep understanding of argumentative tactics. The goal of this PhD Thesis is to overcome the shortcomings of traditional end-to-end applications in which all components are trained from the dialogs themselves, by incorporating several dialogue, argumentation and domain features.