Thesis & projects proposals

GOOGLE CLASSROOM FOR THESIS' STUDENTS

All my thesis' students are cordially invited to join the google classroom (classroom code mwx76vm), also joinable with this [LINK]. There you will find all the information and the registration form in order to keep in contact. The classroom will also be used in order to send information regarding the weekly meeting and it's meet link, which you can however find in my calendar under the code [OHT].

Tutti i miei studenti sono inviati ad iscriversi alla google classroom (classroom code mwx76vm), anche attraverso questo [LINK]. Lì troverete tutte le informazioni ed il form di registrazione anagrafica che sarà utile per mantenerci in contatto. La classroom sarà anche usata per mandare informazioni riguardo ai meet settimanali e per condividere il meet link, che comunque potete anche trovare nel mio calendario con il codice [OHT]

Here you will find some thesis proposals for students

The topics are available both for undergraduate projects (Laurea) and graduate master project (Laurea Magistrale)

SIGNAL PROCESSING & ANALYSIS

[LM01] Motion repeatability modeling

Investigating using new data processing strategies for motion features extraction and repeatability modeling (i.e Scalogram and Convolutional Neural Networks, Recurrent Neural Networks for time series analysis).

[LM02] Thermal drift modelling

investigating using Transformer Neural Networks and Recurrent Neutral Networks for modelling the nanometric thermal expansion of bodies. Investigate the contribution of positional and thermal information in the modeling problem and assess if complex models allow to simplify the existing concepts for data driven compensation systems.

[LM03] Modeling of the geometrical distortion

Modeling of the geometrical distortion of a beam cross section pattern to retrieve misalignment properties of the beam (rotations and transportations): Investigating new feature extraction from the imaged pattern or using convolutional neural network directly on images of the cross section

[AP01] Spatiotemporal Transformerss for anomaly prediction in multidimensional signals

A project with Tales Alenia Space. This project investigates how transformer models with spatiotemporal attentioncan be used to detect anomalies in complex satellite data. By using a unique multi-level attention mechanism, the model captures both time-based and spatial relationships across multiple variables (like temperature, power, or velocity) and geographic locations. This approach is tailored for the intricate patterns in satellite data, aiming to improve accuracy in anomaly detection while offering insights into where and when anomalies occur.

[AP02] Multi-head Self-attention Transformer for Interpretable Anomaly Detection in Satellite Time Series Data

A project with Tales Alenia Space. This project focuses on creating an interpretable anomaly detection model using a Transformer architecture with multi-head self-attention to analyze satellite time series data from sources like ESA or the Mars Express Orbiter. Unlike tra ditional models that merely flag anomalies, this approach identifies the specific sensors or variables contributing to each anomaly, providing detailed insights into the nature and origin of these irregularities. Through the multi-head self-attention mechanism, the Transformer can cap ture complex temporal relationships across multiple sensors, allowing for a more accurate, sensor-level interpretation of anomaly sources.

[AP03] Enhancing Graph Deviation Networks (GDN) for Anomaly Detection on Satellite Time Series

A project with Tales Alenia Space. This project seeks to improve the Graph Deviation Network (GDN) model for detecting anomalies in satellite time series data by building on GDN’s capacity to model relationships between sensors as a graph. The enhancement involves integrating recurrent or transformer-based layers to better capture both long-term temporal and spatial dependencies, which are crucial for dynamic satellite data. Additionally, an online training mechanism will enable the model to adaptin real-time as new data arrives, making it more effective for tracking sudden atmospheric changes or long-term geophysical trends. This adaptation aims toprovide a robust, interpretable anomaly detection solution that is specifically tailored to the evolving nature of satellite data.

RENDERING

[GD01] Inverse and differentiable rendering

For this project, you are asked to design and implement a novel application that utilizes one of primary rendering techniques: Neural Radiance Fields, Soft Rasterization, and Differentiable Physically Based Rendering, with a corresponding framework that implements the technique, along with examples of applications.. You may draw inspiration from the examples presented in class or other existing applications; however, the application you propose should be unique and original.

MANIFOLD TRANSFORM & MODELING

[IT01] Geodesic CNNs for Spatial-Temporal EEG Analysis

Design convolutional neural networks that perform convolutions along geodesics on the EEG manifold. This allows the network to respect the non-linear geometry of the data whenearning spatial and temporal features. By incorporating manifold-valued kernels, the model can better capture the complex relationships in EEG signals, leading to improved performance in tasks such as seizure detection or mental state classification.

[IT02] Adversarial Training with Manifold Constraints for EEG Augmentation:

Implement a generative adversarial network (GAN) where the generator produces synthetic EEG data constrained to the Riemannian manifold. By enforcing manifold constraints, the synthetic data maintains the geometric fidelity of real EEG signals. This method enhances data augmentation strategies, providing more realistic samples to improve modelgeneralization and robustness against overfitting.

[IT03] Manifold-Based Diffusion Models for EEG Data Augmentation

Implement diffusion models that generate synthetic EEG data within the Riemannian manifold of SPD matrices. By modeling the diffusion process on the manifold, the generatedsamples maintain the intrinsic properties and variability of real EEG signals. This method provides high-quality data augmentation that enriches training datasets without violating the geometric constraints of EEG data. It enhances model robustness and generalization, especially in scenarios with limited data.

[IT04] Transfer Learning via Manifold Alignment:

Develop a transfer learning framework that aligns EEG data from different subjects onto a common Riemannian manifold. By using manifold alignment techniques, the model reduces inter-subject variability, allowing knowledge transfer without extensive retraining. This personalization enhances the applicability of EEG-based models in brain-computerinterfaces, making them more adaptable to individual differences.

[IT05] Denoising Diffusion Probabilistic Models on Riemannian Manifolds:

Design denoising diffusion probabilistic models (DDPMs) that operate on the Riemannian manifold to reconstruct clean EEG signals from noisy observations. The diffusion andreverse processes are defined using manifold-valued stochastic differential equations, ensuring that the denoising respects the data's geometry. This approach effectively removesnoise and artifacts while preserving essential neural information, improving the quality of EEG data for analysis.

[IT06] Sparse Coding with Riemannian Constraints for EEG Decomposition:

Implement sparse coding algorithms that operate on the manifold of EEG covariance matrices. By enforcing sparsity in the manifold domain, the method decomposes EEGsignals into a set of geometrically meaningful basis functions. This facilitates the extraction of salient features associated with specific neural processes, aiding in tasks like sourcelocalization or artifact removal.

[IT07] Self-Supervised Contrastive Learning with Manifold Embeddings for EEG:

Develop a self-supervised learning framework where EEG data is projected onto a manifold, and contrastive loss is used to learn meaningful embeddings without labels. Dataaugmentations respecting the manifold structure generate positive pairs, encouraging the model to learn invariant features. This method is particularly useful when labeled EEG data is limited, enabling the model to leverage large amounts of unlabeled data.

GRAPH NEURAL NETWORKS

[VP01] Enhancing Traffic Prediction Robustness with a Bayesian GRU-GAT Architecture: Managing Uncertainty in Dynamic Traffic Networks

This research introduces a novel Bayesian GRU-GAT (Graph Attention Network) model specifically designed to handle the inherent uncertainties in dynamic traffic data. The model combines the capabilities of GRU (Gated Re current Units) for temporal sequence modeling with GAT (Graph Attention Networks) for spatial attention across traffic nodes. By employing Bayesian in ference, the GRU and GAT layers treat their weights as probabilistic variables, enabling the architecture to quantify and manage uncertainties in spatial and temporal patterns. This approach is crucial for real-world traffic applications where data can be sporadic or noisy due to sensor failures, incomplete information, or environmental disruptions.

[VP02] A GNN-based traffic prediction framework that incorporates hierarchical masks to enable explainable predictions.

While GNNs achieve high predictive accuracy, their opaque decision-making process remains a challenge. This project proposes a flexible framework designed to enhance both prediction accuracy and interpretability in traffic forecasting. Using a perturbation-based interpretation generator with hierarchical spatial and temporal masks, the framework selectively highlights critical traffic features without altering the original model structure.

[VP03] Enhancing Next POI Recommendations with Cluster-Aware Graph Attention Networks

Predicting users’ next Points of Interest (POI) is essential for developing personalized location-based services and optimized route planning. This project proposes a novel approach that will leverage Cluster-Aware Graph Attention Networks (GATs) to enhance POI recommendation accuracy by dynamically focusing on relevant spatial-temporal patterns in user movement data. The intended architecture aims to improve GATs by incorporating user clustering, embedding both cluster centroids and prominent user preferences within eachgroup.

[VP04] Kolmogorov-Arnold Graph Attention Networks: Enhancing Complex Relationship Modeling in Graph Structures

This project introduces Kolmogorov-Arnold Graph Attention Networks, an innovative architecture that leverages the Kolmogorov-Arnold representation theorem to enhance the modeling of complex, multi-variable relationships within graph structures. This architecture utilizes the theorem’s principles to decompose high-dimensional functions, allowing the model to capture intricate dependencies among nodes in a structured and computationally efficient manner. By embedding this approach within Graph Attention Networks (GAT), the model dynamically focuses on essential nodes and connections, while also simplifying and breaking down relationships into manageable components.

HUMAN-MACHINE & HUMAN-ROBOT INTERACTION

[EF01] Improving Eye-tracking Performances with YOLO/SSD and ViTs

Combining ViT-YOLO/SSD algorithm aims to enhance eye-tracking performance, particularly in the y-axis, by integrating YOLO (You Only Look Once) or Single Shot MultiBox Detector (SSD) with Vision Transformers (ViTs) for robust gaze estimation. YOLO or SSD are employed to localize eye regions in real-time video streams, ensuring accurate detection even under varying lighting and head orientations. The ViT further processes these areas for iris and pupil pixel segmentation, allowing the structure to predict gaze direction with improved precision, specially during challenging vertical eye movements, a common limitation in traditional eye-tracking systems. The combined model is trained on a diverse dataset of eye images, enriched with annotations for gaze direction.

[EF02] Handless Interaction with Technology

The proposed gaze-controlled interaction system enables users to operate a device hands-free by registering prolonged fixations as clicks in specific locations on the screen throughhigh-precision eye tracking technology. The system monitors gaze direction and detects whenthe user fixates on a particular area for a predetermined duration, such as one second. Once this fixation threshold is met, the system interprets it as a click, triggering the corresponding action. With a calibration phase to adapt to individual gaze patterns, this system is particularly beneficial for individuals with mobility impairments or in scenarios where hands-free operation is essential.

[EF03] Modeling of Human-Environment Interactions Graphs through GNNs and ViTs

This project models human-environment interactions using Graph Neural Networks (GNNs) and Vision Transformers (ViTs) to extract and represent dynamic interactions through interaction graphs. Given a hand-crafted dataset consisting of labeled sequences of images or videos capturing individuals as they move and interact with their surroundings, the system uses ViTs for feature extraction, followed by GNNs to model spatial and temporal relationships. The output is a structured graph that captures the interactions, including object manipulation, proximity, and sequential movements, allowing for a detailed analysis of behavioral patterns.

[PS01] A Machine Learning Approach to assess the interlocutor's attention and understanding during robotic interactions

Related to the Hermes(WIRED) project.. A The student will address the problems of interlocutor's attention and understanding when interacting with a robotic device. The domain of interest is the social robots industry, while both machine learning and image processing algorithms will be applied.

[PS02] A Machine Learning Approach to assess cognitive conditions in fragile individuals

A inter-department and inter-university project .The student will address the problems of cognitive testing and mental impairment diangosis on fragile individuals (eg. children, elderly people, impaired individuals, etc...). The domain of interest is machine learning applied to mental healthcare.

[PS03] A Machine Learning Approach to adaptive interfaces for mentally impaired individuals

A project within the European collaboration HASFRIENDS . The student will address the problems of self-adaptive user interfaces to be used by mentally impaired people. The domain of interest is machine learning applied to mental healthcare.

[SA01] Autonomous Robots as a Resource for Mass Epidemic Sustainability assisting in critical care wards.

A medical robotics and AI project with several hospitals and healthcare facilities. The goal is to implement robots as acting remote interface for physicians at home in smart-working mode, to perform diagnostic tasks that do not require high precision manual interactions, and therefore decreasing the workload of physically available operators in critical care wards. The secondary goal is to implement an autonomous tool in order to spare a large amount of time that is generally wasted by the caregivers in sanitization operations, tampering with number and duration of visits to patients who are normally left alone for a long time, also lowering the quality of service perceived by the patients, as well as harming them also on the psychological side, with significant fallbacks on the recovery speed.

[SA02] A machine learning based approach for Healthcare Emergency Resources Management during Mass Epidemics.

A computational medicine project with several hospitals and healthcare facilities. The goal is to implement Machine Learning Algorithms to provide a preemptive planning strategy for intensive care units' accesses in order to enforce middle and long term sustainability in pandemic scenarios. The student will implement machine learning algorithms to predict the bed availability in critical care units, as well as predicting the workload availability due to potential contagions of caregivers and also to improve the emergency management of hospital facilities.

Page updated

Report abuse