MAIN PhD Seminars 2026

 

Date	Speaker	Title
March, 4th	Filippo Concas	Learning to Translate with Scarce Data Foundations and Recent Advances in Low-Resource NMT
March, 11th	Luca Garau	How to build a PhD: from 3D printing to advanced industrial fabrication
March, 18th	Marco Ferrara	Implementing the Simply Typed Lambda Calculus with De Bruijn Levels
March, 25th	Davide Antonio Mura	Bridging Modalities: Fusion Strategies and Vision-Language Alignment
April, 1st	Roberta Angioni	Healthcare Timetabling and Routing: State of the Art and Emerging AI-Driven Approaches
April, 8th	Emily Priyadarshini	DeFi Protocols and Attacks: A glimpse into the Wild West
April, 15th	Simone Pinna	Can machines deeply understand us? The role of AI in modern psychology
April, 22nd	Giulio Casti	Machine Learning Applications in Economics: Case Study on Startup Survival Prediction
April, 29th	Federico Meloni	Third Year in the third dimension: a voyage into Volumetric Tetrahedral Maps
May, 6th	Fabio Littera	Beyond Static Profiling: Navigating User Intent with Conversational Agents and Reinforcement Learning
May, 13th	Martina Salis	Hybrid AI methods for data management and behavioral analysis
May, 20th	Marco Murgia	Challenges and strategies for robust AI-Generated Text Detection
May, 27th	Marco Damele	Problems on skew left braces
June, 3rd	Andrea Cabriolu	A Bayesian and heuristic approach to an optimization algorithm for the dynamic scheduling of astronomical observations
June, 10th	Giuseppe Scarpi	From uncertainty to formalisation: mitigating limits of Language Models in practical scenarios

All the seminars at 13:00 in Aula Magna di Fisica

 

Filippo Concas: Learning to Translate with Scarce Data Foundations and Recent Advances in Low-Resource NMT
Research Topic: Natural Language Processing

Neural Machine Translation (NMT) has shown strong performance in high-resource settings, where large and clean parallel corpora are available. These results, however, are harder to reproduce for low-resource languages, where data is scarce, heterogeneous, and often noisy or affected by domain mismatch. In this seminar, we introduce machine translation from its probabilistic formulation to modern neural approaches, highlighting the central role of training data and the assumptions underlying current NMT models.

We then focus on low-resource scenarios, discussing how data limitations influence training and evaluation, and reviewing common strategies such as data augmentation, transfer learning, and multilingual modeling. The seminar concludes by presenting recent research trends that adopt a data-centric perspective on neural machine translation, emphasizing data selection, filtering, and quality assessment as key components, and outlining open questions and directions for future research.

Luca Garau: How to build a PhD: from 3D printing to advanced industrial fabrication
Research Topic: Computer Graphics – Digital Fabrication

This seminar analyzes the evolution of digital manufacturing, comparing additive processes (based on material deposition) with subtractive ones (characterized by material removal). Positioned between these two techniques are casting processes, widely used in various industrial fields, which involve pouring molten material into molds. Starting with a general overview, the seminar will focus on the technical challenges related to the design of casting molds. We will delve into the unresolved problems affecting the efficiency and precision of reusable rigid molds, with particular emphasis on the fabrication of complex and unconventional geometries..

Marco Ferrara: Implementing the Simply Typed Lambda Calculus with De Bruijn Levels
Research Topic: Implementation of formal systems based on typed lambda calculus

The lambda calculus is a foundational formalism for describing computation, and it underlies many logical and programming systems. Its standard presentation relies on named variables, which are human-friendly but introduce complications related to alpha equivalence and substitution in formal and computational settings. Nameless representations of binding, such as de Bruijn levels, offer an alternative approach that avoids these issues while retaining a precise account of variable binding and scope. Using these representations, formalizations of logics based on the lambda calculus in a proof assistant such as Rocq provide an environment for establishing properties of programming and formal languages whose definitions or semantics are based on such logics.

Davide Antonio Mura: Bridging Modalities: Fusion Strategies and Vision-Language Alignment.
Research Topic: Computer Vision

In recent years, Deep Learning has revolutionized a multitude of tasks, yielding outstanding results in the analysis of single-modality data, such as images in Computer Vision or text in Natural Language Processing. However, many real-world applications require combining different types of information to capture complementary features and richer semantic context, making the analysis of multimodal data essential. This seminar introduces the fundamental concepts of Multimodal Deep Learning, presenting the main data types involved and demonstrating how neural models can effectively combine diverse information sources. The primary multimodal fusion techniques will be discussed, highlighting their characteristics, advantages, and limitations. Finally, the seminal architecture Contrastive Language–Image Pretraining (CLIP) will be examined as a concrete example of alignment between natural language and images in a shared representation space.

Roberta Angioni: Healthcare Timetabling and Routing: State of the Art and Emerging AI-Driven Approaches
Research Topic: Hybrid Optimization Methods and Generative AI in Healthcare

The effective optimization of human and logistic resources constitutes a critical socio-economic imperative in the healthcare sector. This necessity drives the search for advanced solutions to complex combinatorial challenges, such as the Home Healthcare Routing and Scheduling Problem (HHCRSP) and the Integrated Healthcare Timetabling Problem (IHTP). The foundational work of Mankowska et al. (2014) and the benchmarks defined by the winning approaches of the Integrated Healthcare Timetabling Competition 2024 establish robust baselines for these problems; however, the application of exact solvers to these domains often faces limitations, and the NP-hard nature of these problems calls for continuous research to mitigate the intrinsic limitation of an intractable global optimality. 

Addressing these challenges, the seminar first provides a critical overview of state-of-the-art resolution methods, then the focus shifts to emerging AI-driven architectures: we investigate how Large Language Models (LLMs) can function as active heuristic components within frameworks based on classic metaheuristics and genetic algorithms, or within specific hybrid systems incorporating other AI techniques based on Monte Carlo methods. Finally, drawing from ongoing research efforts, an overview of the obtained experimental results is also provided, critically assessing both the competitive advantages of these AI-integrated solutions and the specific limitations that have emerged.

Emily Priyadarshini: DeFi Protocols and Attacks: A glimpse into the Wild West
Research Topic: Blockchain Security

Decentralized finance (DeFi) has enabled a suite of financial services to be offered without relying on trusted intermediaries. Ethereum, an account-based blockchain model, implements DeFi using smart contracts. In this seminar, we give an overview of the DeFi landscape through a few widespread DeFi protocols. Specifically, we use these protocols to witness how blockchains have given rise to new market dynamics and subsequently novel attack vectors. To conclude, we present a few general approaches and common tools used to mitigate such attacks and discuss some potential areas for research.

Simone Pinna: Can machines deeply understand us? The role of AI in modern psychology
Research Topic: AI in Psychology

In recent years, artificial intelligence has become increasingly present in the field of health, covering various care roles. However, when it comes to mental health, its intervention is still seen as taboo. This mistrust is motivated by valid reasons, which are examined by the most recent research, with the aim of breaking down the barriers to understanding between humans and machines. For example, Large Language Models are commonly used for their ability to mimic reasoning patterns expressed in human natural language. However, in the field of psychology, standard patterns are not always our ally, and an overly rigid interpretation can lead to risks and clinical misjudgments. This seminar will explore the ability of AI (LLMs in particular) to approach the sensitive issue of mental health and how this technology can be used in different roles (diagnostic hypotheses, emotion recognition, patient simulators, training for therapists, etc.). The fundamental rule remains the same as in medicine: the use of AI is not intended to replace experts but to improve their ability to treat patients by capturing the nuances in their language. There are many obstacles to the introduction of these new technologies into this world, including realism, adaptability, trust, explainability, ethics and a lack of data for privacy reasons. We will look at the most promising techniques for overcoming these obstacles and how they have evolved over time to refine the ability of computers to learn the pattern of our psyche.

Giulio Casti: Machine Learning Applications in Economics: Case Study on Startup Survival Prediction
Research Topic: Machine Learning in Economics

In the highly volatile startup ecosystem, the ability to predict the survival and growth of early-stage companies is a critical challenge for investors, policymakers, and entrepreneurs alike. Traditional economic evaluation methods, such as comparative analysis,  often struggle to account for intangible assets and extreme market uncertainty. This seminar presents an approach to predicting startup survival by leveraging advanced machine Learning methodologies applied to large-scale structured data from the PitchBook platform. The discussion focuses on a comparative analysis of different algorithmic architectures, highlighting their performance in binary classification tasks. Furthermore, we explore the application of survival analysis models to handle right-censored data, allowing for the analysis of survival probabilities as they evolve over time. Key findings report that factors beyond traditional financial indicators, such as team experience, lead investor counts, and competitive positioning, play a crucial role in business longevity. Finally, the seminar outlines future research directions, including the integration of deep learning models to capture temporal market dynamics and the development of real-time monitoring platforms for investors and entrepreneurs.

Federico Meloni: Third Year in the third dimension: a voyage into Volumetric Tetrahedral Maps
Research Topic: Computer Graphics – Geometry Processing

Volumetric tetrahedral maps are a key tool in geometry processing, enabling applications such as simulation, deformation, and volumetric texturing. Compared to surface parameterization, volumetric mapping poses additional challenges in ensuring bijectivity, controlling distortion, and achieving robustness on complex meshes. This seminar introduces the problem and surveys the main state-of-the-art approaches, outlining their strengths, limitations, and open challenges. In the second part, I will present my ongoing research on volumetric tetrahedral mapping, describing my methodological perspective, how it differs from existing techniques, and the contributions I am developing toward more robust and practical volumetric maps.

Fabio Littera: Beyond Static Profiling: Navigating User Intent with Conversational Agents and Reinforcement Learning
Research Topic: AI Agents and Reinforcement Learning

Traditional Recommender Systems often struggle to capture evolving user needs, particularly in “cold-start” scenarios where prior interaction data is missing. To address this limitation, Conversational Recommender Systems (CRS) have emerged as a solution, enabling users to express complex preferences through multi-turn natural language dialogues.

In this seminar, we present an Agentic AI architecture that frames the recommendation task as a sequential decision problem. We model the system as a Markov Decision Process (MDP), formally defining its State Space, Action Space, Transitions, and Reward functions to bridge the gap between language understanding and decision-making. To ensure efficient retrieval over large catalogs, we introduce a hybrid filtering mechanism that combines deterministic constraints (Hard Filters) with semantic embedding similarity (Soft Filters).

The core of our approach is a Reinforcement Learning Director, an agent trained to optimize the trade-off between exploration (eliciting preferences via questions) and exploitation (recommending items). This decision process is driven by the semantic uncertainty of the search space, modeled through embedding-based representations. Finally, we discuss the challenges of offline evaluation in conversational settings and demonstrate our validation strategy using an LLM-based User-Simulator.

Martina Salis: Hybrid AI methods for data management and behavioral analysis
Research Topic: Hybrid data analytics

The growing adoption of data-driven systems and artificial intelligence poses new challenges in data management, governance, and behavior analysis across heterogeneous domains.

In this seminar, we will present a set of works that leverage large language models (LLMs) to address these challenges through scalable and lightweight AI-based solutions.

Specifically, we will discuss a metadata-based technique for detecting similarity among data products, LLM-based methods for the automated verification of natural-language semantic rules (policies) on data products and a system that uses LLMs and sensor data from smart homes to recognize and explain anomalous behavior with respect to examples of normal daily activities.

Marco Murgia: Challenges and strategies for robust AI-Generated Text Detection

Research Topic: Generative AI

The exponential advancements in Large Language Models (LLMs) have revolutionized text generation, making it increasingly difficult to distinguish between human-produced and machine-generated content. This indistinguishability raises significant concerns in critical domains such as disinformation, academic integrity, and online content security. This seminar will explore the inherent challenges of AI-generated text detection, focusing on current methodologies and strategies to enhance the robustness and generalization of detection systems. Beginning with an understanding of the stylistic and statistical characteristics that differentiate AI text, we will study the state of art and analyze approaches based on fine-tuned language models and discuss findings from recent research, including personal contributions to developing detectors resilient to adversarial attacks. We will conclude by outlining the ethical implications and future directions of research in this rapidly evolving field, emphasizing the continuous ‘cat-and-mouse game’ between generation and detection capabilities.

Marco Damele: Problems on skew left braces
Research Topic: Algebra

Skew left braces are algebraic structures introduced to study set-theoretical solutions of the Yang–Baxter equation, with a one-to-one correspondence between solutions and skew braces. Recently, the concept has been extended to Lie groups.

Research on skew braces focuses on understanding the relationship between the additive and multiplicative groups, including the open Byott conjecture, which asserts that a skew brace with solvable additive group has a solvable multiplicative group. In the finite case, skew braces are also studied as single algebraic objects, with interest in properties such as solvability, supersolvability, and nilpotency, as well as analogues of the Cauchy and Sylow theorems. Another line of research uses permutation braces to analyze algebraic properties of set-theoretical solutions, such as indecomposability and irretractability. In this talk, I will provide an accessible introduction to skew left braces, explaining what they are and why they are interesting. I will then present the main lines of research in the area and discuss my recent contributions to the field.

Andrea Cabriolu: A Bayesian and heuristic approach to an optimization algorithm for the dynamic scheduling of astronomical observations
Research Topic: Heuristic algorithms, dynamic scheduling

In order to maximize the scientific time used by the Sardinia Radio Telescope a Dynamic Scheduling system will be developed. An all-encompassing architecture will be designed based on a careful top-down approach. A complete system will be presented, in which the weather tools, proposal tools, scheduling system, observing system and telescope management tools will all be integrated.

Among all the components that compose the relationships between the system’s actors, the “Optimizer” component plays a key role. A central element of the entire architecture, its main role is to create long- and short-term schedules, basing its choices with Bayesian and heuristic approaches on various parameters, such as weather forecast data, antenna availability, scientific constraints, and the like.

Basically, Optimizer will involve a set of algorithms with the aim to validate if the conditions and the constraints defined by astronomers at proposal phase would be met at a specific time. Some of the tasks managed by Optimizer aim to predict observation-relevant parameters using information such as real-time weather data and time series; others assign appropriate weight to parameters; still others calculate long- and short-term schedules.

Optimizer must maximize a specific parameter called “figure of merit”, related to the semester. These are metrics that statistically establish the desired success rate for projects of various levels or the efficiency of telescope time use. Therefore, the “figure of merit” governs the quality of astronomical observation arrangements, prioritizing specific characteristics so that the activities carried out can meet the objectives established by the Observatory.

Giuseppe Scarpi: From uncertainty to formalisation: mitigating limits of Language Models in practical scenarios
Research Topic: Language Models

It was clear from the beginning, but now we are interiorizing it: language models are neither totally stochastic, nor totally deterministic. Thus, they can’t be creative, but not even completely reliable.

With the increasing integration of LM in critical applications, undestanding and mitigating their limits has become very important. In this seminar we will examine some countermeasures – like PLD formalization, usage of ontologies/nomenclatures, delegation of critical functions to deterministic software – and some practical applications of synthetic data generation and genuine data analysis.