Calls 2022

Automated Planning is the problem of synthesizing courses of actions guaranteed to achieve a 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"). Planning and scheduling techniques are important in several application domains such as flexible manufacturing and robotics. Unfortunately, these techniques suffer from scalability issues and are often unable to cope with the complexity of real-word scenarios, despite the significant advances in the field.

Recently, efforts such as Deepmind AlphaZero 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 a 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 adaptive planners that can learn strategies capable of solving problems in a specific application scenario and improve their performance (in terms of both speed and quality) over time.

AI components are more and more used in safety-critical systems in different application domains such as automotive or space. In particular, the increased availability of sensor data gives the opportunity to increase the autonomy these systems with advanced perception, optimized control, and efficient fault detection and recovery. The validation, verification, and safety assurance of AI components in these systems are therefore of paramount importance. However, the uncertainty of Machine Learning (ML) algorithms poses hard challenges for traditional approaches. In this PhD project, we aim at investigating new model-based design techniques to ensure the safe usage of AI/ML components. We will explore the definition of new formal models to represent the uncertainty of the ML models and the related errors, as well as formal verification techniques for the evaluation of the reliability of the system with AI components, and will design and evaluate architectural schemas in specific application scenarios.

While the task of automatically recognising hateful content online has been extensively explored in the last years within the NLP community, what is the best strategy to respond to such messages has only recently entered the research agenda. One of the main issues related to this task is indeed how to best measure the effects of computer generated counter-narratives (i.e. textual responses to hate messages), in order to identify the most promising approaches. This thesis will explore this topic across NLP, NLG and complex networks in order to combine content-based, emotion-based and network-based metrics and apply them effectively to fight online hate via analysis of Social Media content spreading.

Machine and deep learning methods are entering also the industrial sector to automatise 3D monitoring and analysis tasks. The research should investigate the use of AI-based methods to boost photogrammetric 3D inspections for industrial quality control operations. Innovative and advanced AI-based solutions should be developed in order to inspect non-collaborative surfaces (reflective, transparent, etc.) and derive precise 3D results useful for quality control.

Machine learning and deep neural networks are extensively and successfully used to process multimodal data (e.g., audio, video, environmental data) on powerful computers. At the same time, several challenges still need to be solved to bring AI on low consumption devices (e.g., end nodes in an IoT) with limited resources. Recently TinyML approaches are emerging to distribute the intelligence at the far edge in the edge-to-cloud continuum. Exciting research scenarios emerge, spanning from novel, innovative hardware for always-on and event-based sensing to tiny deep learning solutions for inference on resource-constrained platforms based on distillation, quantization, or neural architecture search. The complexity grows if we want to move learning to the edge. Motivated by these scenarios, the research aims to (i) define novel hardware/software approaches to optimize AI at the very edge on energy-efficient embedded devices, in particular for audio processing and/or computer vision; (ii) to explore the potential of distributing and fuse the intelligence in heterogeneous nodes of an IoT (iii) to demonstrate the advantages of the investigated approaches in real-world application scenarios, such as those of smart cities.

Data is often collected at the edges of the network but processed centrally fueled by the availability of computing power provided by the cloud. However, the edge of wireless networks can play a role as a distributed platform for ML mitigating the latency and privacy concerns as well as alleviating backhaul network from the transmission of data to the cloud.

The main goal of this PhD is to investigate the impact of bringing learning at the edges of wireless networks, considering an edge-cloud network which is AI aware and where machine learning algorithms interact with the physical limitations of the wireless medium.

The ability of modeling, understanding and predicting human behaviors, mobility routines and social interactions is fundamental for computational social science and has a range of relevant applications for individuals, companies, and societies at large. In this project, the goal is merging approaches from machine learning and network science (e.g., graph neural networks, multi-agent deep reinforcement learning, etc.) and using data on mobility routines (e.g., GPS and other mobile phone data), face-to-face interactions and communication data in order to develop methods for quantify daily habits, individual dispositions and traits, and behavioral changes. A special attention will be given to the changes on daily human behaviors due to the emergence and spread of the Covid-19 pandemic and other shocks. The Ph.D. project will be conducted within the FBK MobS research unit but with collaborations with several international groups (i.e., MIT Connection Science) and with the ELLIS program of the Human-Centric Machine Learning.

The European open data policies have led to the definition of the concept of data space: ecosystem of data within a specific application domain and based on shared policies and rules where users are enabled to access data in a safe, transparent, reliable way, easy and unified.
In this project, the goal is to provide Human-Centered AI tools capable of enabling a data space for mobility , in the context of the European green deal, keeping a balance between users' freedom and companies' constraints.

In the context of the green deal transition and climate change 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 multi-/hyper-spectral images acquired by passive satellite sensors and/or Synthetic Aperture Radar (SAR) images acquired from active systems for Earth Observation. Among the others, data from ESA Copernicus (Sentinels), ASI PRISMA and COSMO-SkyMed 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 detecting trends and changes for modeling and understanding their impacts on climate and environment.
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.

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. 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. On the other side, 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 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 Ph.D. is to make ST flexible and robust to these and other factors.

Conversational agents are designed to interact with users through various communication channels, such as social media platforms, using natural language. Recently neural end-to-end systems have started to be tested to fight misinformation using argument generation to debunk fake news. Still, Neural Language models suffer from limitations such as hallucination and knowledge lack. Scaling to credible, up-to-date and grounded arguments requires 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 neural language models, by incorporating several knowledge sources, argumentation and domain features into a constrained generation pipeline.

The research project will focus on the development of tiny machine learning models for learning continuously over time and under domain shift.  The research will focus on developing compact deep learning models (i.e. with reduced memory footprint and computational cost) for domain adaptation and continual learning. Techniques for network pruning and Neural Architecture Search methods will be investigated.

The goal of the thesis is to develop AI models for speech processing which are aware of the computational resources and of the application requirements and are capable of dynamically adapting in order to meet such limitations. This entails not only the search for a trade-off between resources and inference performance but also the possibility to dynamically exploit additional computational resources, eventually expanding the model. The project will address both training and inference phases, starting from state of the art supervised techniques as model compression, neural architecture search, distillation and continual learning and pushing them towards continuous and unsupervised solutions.