- ✨ Open-Source Contributor of AlphaZero.jl:
13xspeed improvement - 🔬 Research on Distributed Reinforcement Learning in Robotics & Self-Supervised Natural Language Processing @ LRDE & CEA (France)
- 💻 Double CS Master and Engineering degree @ Epita & Sorbonne in Computer Vision
- 🔥 Interested in Reinforcement Learning, Computer Vision & High Performance Computing
- 🚴 Love Cycling, Climbing & Piano
- 🇬🇧 Speaking English & French
I was introduced to the world of AI and research during my first internship in my school's AI laboratory, where I used Machine Learning (sklearn) methods to improve the detection of malicious traffic in network cores. My contribution led to a 50% reduction in errors compared to the previous model.
Recently, I contributed to the Open-Source community through the Google Summer of Code program. AlphaZero, an algorithm developed by Google Deepmind, has made significant advancements in the fields of chess, Go, and Shogi through its Reinforcement Learning approach. To make this algorithm more accessible, AlphaZero.jl has been created as a powerful Open-Source implementation of this algorithm in Julia. I parallelized the environment exploration using CUDA, which resulted in a 13x speedup compared to the previous implementation, as described in this accessible Medium article.
During my studies in France at Epita and my double master's degree at the Sorbonne, I specialized in Machine Learning and Deep Learning methods for Computer Vision, extensively experimenting with Tensorflow and PyTorch Framework. Meanwhile, I joined Epita's Research and Development Laboratory, where I am researching and developing tools to automatically detect and correct texts extracted from old documents. My latest project involves a supervised approach using a sequence-to-sequence model (BERT), which we plan to publish a paper on in May. This experience has allowed me to deepen my research background and gain experience with RNNs and Transformers in the Natural Language Processing context. To explore further details about this project, please feel free to consult my final report, available here.
Further on the path of computational resources optimization, I am currently undertaking a research internship at the French Alternative Energies and Atomic Energy Commission. In this position, I am responsible for distributing the computing load for training a state-based Reinforcement Learning algorithm, which is being used to control a multi-task robotic arm, by utilizing Jax (\cite{jax}). The goal of this project is to reduce the current four-day training time down to just one hour, resulting in a significant 100x speedup. This acceleration will greatly enhance our ability to iterate through experiments and generate meaningful results, in the hope to reproduce results of SayCan with fewer resources.
In addition to my academic pursuits, I am also involved in several side-projects, including the President of my School's Chess club, Co-Organizer of an AI Safety discussion group, Head of my school's Student Office Events, and participation in Reinforcement Learning side-courses.
I value teamwork and effective communication within a team. I was able to practice the latter when I co-organized a discussion group on AI safety for nine months and was president of Epita's chess club for three years. I am autonomous, hardworking, and passionate. By actively contributing to developing these technologies, I hope to become one of the critical actors in the following decades.
- Mail - guillaume1.thomas@protonmail.com
- LinkedIn - @guillaume-thomas
- Phone - (+33) 6 17 28 50 87