About Research Teaching Education Proficiencies Cool Stuff Depression Fuel


Guilhem Xavier Piat, PhD

My doctorate is in Artificial Intelligence, on the topic of integrating structured knowledge into Transformer-based Language Models. My research interests revolve around Artificial General Intelligence (AGI). I got into language models based on the following idea:

As text (often in the form of natural language, but also program source code, mathematical equations, etc.) is the most versatile and universal tool we use to communicate ideas, inputs and outputs for cognitive tasks and descriptions of non-cognitive processes, it appears sensible that proficiency in the communication medium would entail proficiency in reasoning about the concepts being handled.

It has been my opinion since the beginning of my graduate studies in 2016 that mastering this modality is critical for an AGI, and if no new fundamental breakthroughs are necessary, the first AGI will be largely based on language modeling. This belief has driven and shaped my work thus far. My work on language models has been focused on attempting to leverage knowledge without requiring a model to internalize it. This approach makes language models more parameter- and data-efficient, and allows for greater explainability.

In the medium to long term, I intend to pivot my research into AI safety, particularly AI alignment, as current generative language models are exhibiting near-AGI behaviors and characteristics, and it seems likely to me that AGI will be achieved before alignment.

Outside of work, my interests include tabletop RPGs (D&D in particular, but others as well); outdoor sports such as hiking, climbing and skiing; fitness; and free and open-source software. When I can, I advocate for consumer rights, right to repair, independance of governments and people from big tech, and digital privacy.

guilhem.piat[at]proton.me



Research

PIAT G., 2024. Incorporating expert knowledge in deep neural networks for domain adaptation in natural language processing. Doctoral dissertation, Université Paris-Saclay.

PIAT G., Semmar N., Tourille J., Allauzen A., Essafi H., 2023. What does KnowBert-UMLS forget? 20th ACS/IEEE International Conference on Computer Systems and Applications (AICCSA 2023).

PIAT G., Kirby E., Tourille J., Semmar N., Allauzen A., Essafi H., 2023. Intégration de connaissances structurées par synthèse de texte spécialisé. 30e Conférence sur le Traitement Automatique des Langues Naturelles.

PIAT G., Semmar N., Allauzen A., Essafi H., Gaël B., Tourille J., 2022. Adapting Without Forgetting: KnowBert-UMLS. IEEE International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob 2022).

PIAT G., Semmar N., Allauzen A., Essafi H., Tourille J., 2022. Enriching Contextualized Representations with Biomedical Ontologies: Extending KnowBert to UMLS. SAI Computing Conference 2022.

Teaching

2022-2023 : Université Paris Dauphine

1er Semestre

2nd Semestre

Other work experience

Higher Education

Year span Degree Subject Alma Mater
2019-2023 Doctorate Transformers, NLP, Structured Knowledge Université Paris-Saclay
2016-2018 Master’s Machine Learning for Data Science Université Paris Descartes
2013-2016 Bachelor’s Computer Science Université Paris Descartes

Proficiencies

Category Specific proficiencies
Neural Nets Transformers, PyTorch, Huggingface, AllenNLP
NLP Leveraging Knowledge Graphs, Classification, Entity Linking, Word/Document/Graph Embeddings
ML Active Learning, Dimensionality reduction & Clustering, Retrieval
Programming Python, R, Bash, Java, Prolog, Haskell, Ocaml
Other Teaching, scientific writing in English and French, working in a team

Cool stuff

Here’s some cool stuff I found.

Depression Fuel

Abandon all hope, ye who enter here.