Àlex R. Atrio

Hi, I am a Senior Deep Learning Scientist at Pi School, in Rome, Italy. I am leading EVE, an open-source LLM for Earth Observation and Earth Sciences: a project in collaboration with Φ-lab, the Earth Observation innovation lab at the European Space Agency - ESA. Work on the project involves data gathering, continued pretraining, instruction fine tuning on relevant downstream tasks, alignment, and retrieval augmented generation. I also work in Videoloc, a 1-year project in collaboration with Fondazione Bruno Kessler on automatic subtitling.

Before that, I completed my PhD (“Regularization Techniques for Low-Resource Machine Translation”) in EPFL and HEIG-VD (Switzerland) under the supervision of Andrei Popescu-Belis, on Natural Language Processing, and specifically Machine Translation and Language Generation.

I have mainly worked on low-resourced languages or domains, both in machine translation and language generation. I have worked on improving sequence-to-sequence models by studying the relation between regularization and the flatness of their loss landscape. I have also worked with multilingual and multitask models, and studied how to train more efficient machine translation systems. Regarding text generation with language models, I have studied both rules-based, hybrid, and neural (with synthetic data creation) generation on low-resource domains.

Before starting my PhD, I did my masters on Cognitive Science, with a master’s thesis on cross-lingual sentiment analysis under the supervision of Toni Badia and Jeremy Barnes, in Universitat Pompeu Fabra, in Barcelona.

Before that, I earned a bachelor’s degree in Philosophy at Universitat de Barcelona, specializing in Logic, and Philosophy of Language and Science.