Matteo Massetti

Leveraging Physical Cues for Learned Representations in Visual Question Answering

Inferring knowledge from various sources and data, such as natural language and visual data, is challenging. Several tasks were presented to reach this aim, however, it is not just a matter of solving the task, but it is the assessment of the models' ability to ground natural language information in the visual world. GuessWhat?! is an evaluation framework aiming at assessing the performances of multi-modal conversational models. It is structured as a game in which two players are collaborating for reaching a common objective, by the means of generating and answering questions related to a visual scene. This work presents a new version of the Imagination Module, which is part of both player architectures and helps them to improve their understanding of textual and visual information. The presented version integrates the information about object attributes in the learned representation, to further improve the generalization and grounding capabilities of the models.

Brain Tumor Detection

In this project, I implemented a convolutional neural network to perform semantic segmentation of MRIs depicting brains. For each input image, a binary mask is produced in which the area of the brain affected by the tumour is highlighted. Given the nature of the problem, I implemented two different architectures from scratch: Fully Convolutional Network and U-net.

Irony and Sarcasm detection in tweets

In this project, developed in collaboration with a colleague from the University of Pisa, I implemented several machine learning models to solve the tasks of the EVALITA challenge. We tested different architectures including an implementation of BERT, convolutional layers, GRU and LSTM so as to detect the presence of irony and sarcasm in tweets.

Smart Traffic Lights

In this project, developed in collaboration with two colleagues from the University of Pisa, we conceived and designed a smart solution to the problem of traffic generated at traffic light-controlled intersections. The solution uses IoT sensors to detect the busiest lanes and to detect weather conditions (rain), together with a controller that uses Fuzzy Logic to modify traffic light timers to avoid traffic and accidents

Low Rank Approximation with Normal Equation and Cholesky Factorization

In this project, developed in collaboration with a colleague from the University of Pisa, we used an altered optimisation algorithm to calculate a low rank approximation of a matrix A. This optimisation makes use of the Normal Equation and the Cholesky factorisation. We have both provided a mathematical analysis of the convergence of the algorithm and implemented the Cholesky algorithm from scratch.

Index for Predecessor problems

In this project, I have designed and implemented a compressed data structure, which is able to veil the finding of a certain element. Specifically, an integer is provided and one must search within the data structure for the largest number preceding the one passed as a parameter.

Analysis of Purchasing Data

In this project, I have designed and implemented a compressed data structure, which is able to veil the finding of a certain element. Specifically, an integer is provided and one must search within the data structure for the largest number preceding the one passed as a parameter.

Genetic Algorithm and Parallelism for Traveler Salesman Problem

In this project, I implemented a genetic algorithm to solve the travelling salesman problem on a parameterisable number of nodes. To speed up the evolution process, I used and compared two parallelisation techniques: C++ threads and the FastFlow library.

MoneyTrack

MoneyTrack allows the user to track and review all the earning and expenses. The main goal of MoneyTrack is to be simple to use, to provide a quick way to enter a new expense or income (in app called Bids) and an intuitive category management.

Machine Learning Engineer

Machine Learning Engineer employed in implementing products and solutions for different types of customers, mainly using the Google Cloud Platform.

Software Developer

Full Stack Developer of products, services and platforms in banking and industry.
Within the Research and Development group, I've contributed to the study of new technologies, including Machine Learning and Artificial Intelligence, aimed at improving company's products and identifying new possibile solutions for the customers.

Stage

Programming of a plug-in for the Moodle platform, which could facilitate and automate the administration and correction of multiplechoice questionnaires in order to test the knowledge of a large group of university students efficiently.

Università di Pisa

Final mark: 110/110 e lode

Thesis: Leveraging Physical Cues for Learned Representations in Visual Question Answering

Google Cloud

Trinity college London