I'm Matteo Suez

Ray Tracer

Made with:

Advanced Ray Tracing software based on Nori.

On top of it, I implemented an Octree as acceleration data structure, Monte Carlo sampling and Ambient Occlusion, Distribution & Whited-style Ray Tracing, Path Tracing, Microfacet models and Multiple Importance Sampling.

Recommendation System

Made with:

The project creates a movie-ratings pipeline out of the MovieLens dataset by using Apache Spark.

It aims to create a recommendation system using Locality-Sensitive Hashing & Collaborative Filtering.

Sustainable placements of city lamps

Made with:

Interactive tool for Analysis II students aiming to implement concepts of sustainability.

The objective is to find the optimal placement of city lamps across the EPFL campus in order to minimize the energy consumption and the quantity of CO2 emitted.

Lattice Consensus

Made with:

Implementation of a fully asynchronous weak version of consensus, named Lattice Consensus, on top of some distributed system abstractions, such as Fair-Loss Links, Stubborn Links, Perfect Links and Best-Effort Broadcast.

Predicting poverty in Africa through time using publicly available data

Made with:

Machine Learning project developed by the iGH lab to understand with greater accuracy the distribution of poverty in Africa, in order to aid the planning of humanitarian interventions and managing the allocation of limited resources.

Worked in a team of 3 people to optimize and extend the geographic scope of a previous work to 3 new African countries, predicting poverty through time from multimodal, remotely sensed data such as satellite images. Thereafter, we optimized the previous CNN architecture, by adopting a ResNet152, which resulted in significant performance gains.

Software Transactional Memory

Made with:

Implementation of a STM by adapting a custom version of Transactional Locking II, achieving a speedup of 3x with respect to the coarse lock-based reference implementation.

Higgs Boson Classification

Made with:

Machine Learning project to estimate the likelihood that a given decay signature was the result of a Higgs Boson (signal) or some other process or particle (background).

Amazon Telegram Bot

Made with:

Telegram bot to help tracking and monitoring prices of products on Amazon across some european countries.

GLAD: Gossip Learning Averaging Distance

Made with:

This paper analyzes Federated Learning and Decentralized Learning approaches by presenting use cases, applications, metrics involved, and the variety of architectures and topologies. Finally, it proposes a novel system called GLAD, comparing it with the state-of-the-art research.

GLAD (Gossip Learning Averaging Distance) is a decentralized system based on the gossip protocol that introduces some metrics, such as resource capability and data quality, in the aggregation of several models. By computing a score for each node based on such metrics, a weighted averaging process is effectively executed when the Euclidean distance among the weights of same neurons within a NN model between different nodes is under a specified threshold. This enables better performance in terms of global accuracy, especially for heterogeneous networks, compared to other traditional algorithms that take advantage of nodes with better resources and data quality over poorer nodes during the merging step.

Software for management of the air quality

Made with:

The goal of the software is to provide functionalties for AQI index evaluation and comparison between locations, basing on the provided dataset of registered data. The program works using the concentration of pollutants associated to bad air quality: SO2, O3, PM10 and NO2 and compares them with the European Enviroment Agency standards.