Lorenzo Pannacci

I am currently enrolled as a second year student following the Honours’ Programme and working on my thesis.

Recipient of “Exam bonus” merits from Sapienza University.

Grade: 110 / 110 cum laude

Recipient of “Outstanding student” merits from Sapienza University.

In my thesis “Generazione e utilizzo di dati sintetici da scene tridimensionali per l’object detection” (Generating and Using Synthetic Data from Three-Dimensional Scenes for Object Detection) I explored the generation of synthetic data from 3D environments to mitigate the domain shift between training and test data in the setting of precision agriculture. Using NVIDIA Omniverse, I created synthetic datasets of grape clusters and combined them with real-world data to train YOLOv5 object detection models.

The results demonstrated that synthetic data effectively reduced the domain gap, improving object detection accuracy in the target environment. Additionally, I achieved a small performance boost even within the original domain, highlighting the broader potential of synthetic data in supervised learning for computer vision tasks.

In this project I use complex systems and network analysis tools to investigate how the behavior of users on StackOverflow changed since the diffusion of LLMs at the end of 2022. In particular I study how the network of users and questions evolved over time and how the type and the complexity of questions changed.

My results show that while the network was already declining the diffusion of LLMs sped-up the phenomenon while also greatly increasing the complexity of the questions and how niche their topic is, which together with a decline in questions suggests that users may first ask questions to LLMs and if not satisfied, search for help on the website.

I developed an interactive web-based visualization tool for external sorting algorithms: 2-way merge-sort and its generalization into k-way merge-sort. The tool provides an interface where the user selects the number of input pages, frames and the amount elements per page.

The application provides an animation that makes clear each external memory access and each operation inside the buffer, while also keeping a log about what is happening and recording useful metrics about the execution. The aim of this tool is to help students understand the mechanics of external sorting algorithms with a clean and simple animation. The webpage is available here.

I implemented from scratch in JAX the distribuited training and inference algorithms for Graph Convolutional Networks (GCNs) described in the paper “Distributed Training of Graph Convolutional Networks”, a distribuied optimization scenario where each agent has data about a part of a graph and they agents that communicate over a sparse network topology.

I also compared their performances on different datasets varying the amount of agents and against the non-distribuited implementation, which code I also wrote from scratch in JAX.

I implemented from scratch in Python the components required for Yao’s Garbled Circuits, a cryptographic protocol for secure two-party computation, including the parsing of circuits form file (described in the “Bristol Format”), socket communication and cryptographic primitives as asymmetric encryption and Oblivious Transfer.

The result of many example circuits is provided, as integer and float operations, hashing, encryption and coin tossing.

My team and I conducted a comparative analysis of state-of-the-art disease gene prioritization algorithms (e.g. DIAMOnD, DiaBLE) and applied them to retinal dystrophy data, with the primary objective of drug repurposing.

We carried out every stage of the pipeline, including data preprocessing, algorithm implementation, and result interpretation. Our findings were then synthesized into a comprehensive report.

I implemented from scratch in JAX all the layers and optimization algorithms for an LSTM-based Recurrent Neural Network, including LSTM layers, embedding layers, dropout layers and the RMSprop training algorithm. The network is trained for the autoregressive task of music generation on a large collection of tokenized MIDI files, resulting in 24 million training samples.

Listenable examples are provided by processing converting the generated sequences back into MIDI files and converting them into audio using a music synthesizer.

I designed and implemented a scenario to confront two well-known data center networking topologies for distribuited computing (Fat-Tree and Jellyfish), against a baseline of a non-distribuited computation. Using simulations I obtained data about their job running cost including overhead and mean response time varying the amount of servers and summarized my findings in plots.

I researched, wrote and presented a lecture about a comprehensive overview of Datalog and the introduction of negation within the language, covering the foundations of Horn clauses up to advanced semantic extensions like perfect, well-founded, and stable models, with a focus on the computational and logical implications of negation.

Me and my team won the first place in the competition made by Leonardo Company S.p.A. for the Training Camp of September 2024 at the Master of Science in Data Science at Sapienza about “Multimodal Learning from Language and Vision Processing Neural Networks”.

The competition was about reverse-engineering a image-generating model by extracting the hidden context it has been given starting from the generated image and its original textual description.

During my 4th year of high school as part of “Percorsi per le Competenze Trasversali e per l’Orientamento” (Paths for transversal skills and orientation) I have been selected as a representative for my school for the project “Da Grande Farò il Biologo”, hosted by Roma Tre university’s biology department.

The project consisted in a full-time week of lectures and laboratory sessions made by biology professors from the university and ended with a final examination. Thanks to my score in the evaluation I have been awarded the possibility of entering the bachelor’s course of biology at Roma Tre university without taking the admission test.