COVID-19 Cough Audio Classification with Deep Learning

An automatic COVID-19 cough detection algorithm would allow for a screening process that would be both immediate and low-cost. We propose an audio classification model which is trained and tested on publicly-available datasets of user-submitted cough recordings. Our model utilises the ResNet-50 architecture coupled with a set of pre-processing and data augmentation techniques such as Mel-Spectrograms, SpecAugment or cough segmentation, which we believe provide useful insights for further work on this problem and on audio classification tasks in general. Additionally, our model only makes use of the audio data - avoiding the need for any additional clinical information which is often unavailable or unreliable. Report

Winning team @ AI Coliseum 2020

p

Martín and me teamed up for AIColiseum 2020, an artificial intelligence competition in Java. The contest consisted on implementing the strategy for a game (similar to Battlecode) in which you move different units. One of the key points of our winning strategy was the path finding algorithm. While most contestants decided to go for parallel BFS, we obtained a map informing how near each position is to our base. Then we exploited the fact that maps were symmetric (there were three possible symmetries in the game), thus obtaining a map that tells you how to go to the enemy base.

p

p

Wasserstein-Procrustes for Unsupervised Cross-Lingual Learning (Bachelor Thesis @ MIT)

Word embeddings, which are non-sparse vector representations of words, are at the core of the ongoing neural revolution in Natural Language Processing (NLP). There have been a number of attempts to align the embedding spaces across different languages, which could enable a number of cross-language NLP applications. Performing the alignment using unsupervised cross-lingual learning (UCL) is especially attractive as it requires little data and often rivals supervised and semi-supervised approaches. Here, we analyze popular methods for UCL and we find that often their objectives are, intrinsically, versions of the Wasserstein-Procrustes problem. Hence, we devise an approach to solve Wasserstein-Procrustes in a direct way, which can be used to refine and to improve popular UCL methods such as iterative closest point (ICP), multilingual unsupervised and supervised embeddings (MUSE) and supervised Procrustes methods. Our evaluation experiments on standard datasets show sizable improvements over these approaches. Paper and Github repository.

p

Winners of the Epic Games challenge (Junction 2018)

p

The challenge consisted of creating a game coach for players to improve their game performance in the Darwin project, a survival game similar to Fortnite. We analyzed real data from different game players, and created a series of indicators, such as kill efficiency, distance covered or exposition to other players. Voronoi diagrams were useful for indicating the players' dominance towards electronics, a scarce resource in the game that randomly appears in a map and gives players a huge advantage. Shout out to Pau for writing this awesome blog entry. I really recommend reading it! Github repository.

p

Migrating a database to the cloud @ BaseTIS

p

Implementation of a Database "on the cloud" using AWS and Microsoft Azure. This project involved using Amazon Redshift, Amazon Lambda and Amazon Glue. We then created different indicators (SQL queries) of the company that are automatically updated on the Data Studio.

p

Winners of the 1st CFIS Alumni Datathon

p