University Projects | Matteo De Sanctis

matteo.desanctis2002@gmail.com

Write to the mail above to get code access.

carried out during my Bachelor's Degree in Applied Computer Science and Artificial Intelligence and Master's Degree in Computer Science (curricula in Artificial Intelligence and Data Science) at Sapienza University of Rome.

On the Spectral Dynamics of Diffusion Models Sampling

Master's Degree in Computer Science - curricula in AI and Data Science

Thesis Advisor: Iacopo Masi; Thesis co-advisor Maria Rosaria Briglia

My Master's Degree Thesis involved analyzing behavior of diffusion models during reverse diffusion, specifically during generation of images. These models, at a specific point in time, "speciate" and hence commit to generating a specific class among the many ones seen at training. My work focused on perturbing (and "attacking") this reverse trajectory to study semantic structure that emerges during generation and its preservation, along with the timing of the speciation phenomenon.

Beyond Gradient Descent for Deep Learning: Assessing the Volume Bias Hypothesis through Particle Swarm Optimization

Bachelor's Degree Thesis in Applied Computer Science and Artificial Intelligence

Thesis Advisor: Iacopo Masi

Research on the volume bias hypothesis: Gradient free optimizers compared against the Implicit regularization of Stochastic Grdient Descent, which is at the core of the generalization capabilities of neural networks (surprisingly avoiding many bad minima---where train accuracy is high but test evaluation is low---and reaching a good minima).

Beyond Edge Deletion: A Comprehensive Approach to Counterfactual Explanation in Graph Neural Networks

Machine Learning | Paper under submission

Master's Degree in Computer Science - curricula in AI and Data Science

Graph counterfactual explainer (XAI) that allows for edge additions, deletions and node features perturbations. Extending CF-GNNExplainer (Lucic et al.) and improving over state-of-the-art. Currently submitted to ICLR 2026.

Wavelet Transform for Frequency-Domain Learning Non-Stationary Time Series Analysis

Deep Learning and Applied Artificial Intelligence

Master's Degree in Computer Science - curricula in AI and Data Science

Time series forecasting: enhance frequency-domain MLPs (Yi et al., 2023b) by integrating the Dual-Tree Complex Wavelet Transform on non-stationary datasets.

CUDA kernel implementation of CNN training (Complementary Educational Activity)

Architectures for Artificial Intelligence

Master's Degree in Computer Science - curricula in AI and Data Science

Roofline-analysis workflow for CUDA kernels that simulate a convolutional neural network (CNN) training performed on the Leonardo HPC node proding NVIDIA A100 accelerators.

Quantum Machine Learning

Topics in Physics

Master's Degree in Computer Science - curricula in AI and Data Science

Seminar on Quantum Machine Learning: theory, architecture and training pipeline of a Quantum Neural Network. Papers: Training deep quantum neural networks, Efficient Learning for Deep Quantum Neural Networks.

Aligning Minerva LLMs

Natural Language Processing

Master's Degree in Computer Science - curricula in AI and Data Science

Aligning Minerva to human values with Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO).

Study of Kolmogorov-Arnold Networks Architectural Transfer

Advanced Machine Learning

Master's Degree in Computer Science - curricula in AI and Data Science

Implementation and analysis of performances of Kolmogorov-Arnold Networks in image classification tasks: full pipeline.

Imitational Learning via Diffusion-based Behavioral Cloning

Reinforcement Learning

Master's Degree in Computer Science - curricula in AI and Data Science

Expert imitation learning for Reinforcement Learning agents using denoising diffusion in Gymnasium environments.

Deploying an Image-Generation Application on AWS Lambda: Performance and Scalability Evaluation

Cloud Computing

Master's Degree in Computer Science - curricula in AI and Data Science

Large-scale deployment of an image-generation diffusion model using AWS Lambda, S3, ECR, EC2, CloudWatch, API Gateway, Docker and Locust.

Robotic Navigation

Computer Vision

Master's Degree in Computer Science - curricula in AI and Data Science

LLMs and visual robotic navigation in a Webots simulated environment. NAO robot navigation integrating image question answering and captioning.

Poisoning of Neural Networks

Data and Network Security

Master's Degree in Computer Science - curricula in AI and Data Science

Survey on Neural Networks poisoning, backdoor attacks and possible mitigations.

AutoMIND: Autonomous Multimodal Intelligent Navigation Dashboard

Multimodal Interaction

Master's Degree in Computer Science - curricula in AI and Data Science

Multimodal vehicle driving dashboard developed in a Webots simulated environment integrated with a SUMO interface.

Verification Guided RL

Formal Methods for AI-based System Engineering

Master's Degree in Computer Science - curricula in AI and Data Science

Formal verification of a Reinforcement Learning agent with SPIN to guarantee safe actions execution in Gymnasium environments.

Policy Gradient Descent from scratch

Deep Learning

Bachelor's Degree in Applied Computer Science and Artificial Intelligence

Implementing REINFORCE Algorithm with no use of Deep Learning Libraries.

Stock Prediction

AI Lab: Computer Vision and NLP

Bachelor's Degree in Applied Computer Science and Artificial Intelligence

Prediction of stocks through LSTMs.

NBA

Statistics

Bachelor's Degree in Applied Computer Science and Artificial Intelligence

Thorough analysis of NBA data and statistics and prediction of future winners.