Ayrton SAN JOAQUIN personal page

Ayrton SAN JOAQUIN

Ayrton SAN JOAQUIN

  • Position and research topics

Ayrton is a research assistant at CNRS@CREATE under its DesCartes Program, a collaboration between the French and Singaporean governments to build Hybrid AI for the Smart City. He works mainly under the WP5 (Natural Language Processing) and affiliated with WP1 (Human Computer Interaction).

His mission is to develop Trustworthy AI: AI that is harmless when misused, transparent & well-understood, fair to its users, and privacy-preserving. He is especially interested in the interactions of AI Systems and humans, the role of the technical designer, and the design of socio-political institutions to govern AI systems. He has previously worked with researchers from NUS, AI Singapore, Meta, and Google DeepMind.

Personal webpage: https://ajsanjoaquin.github.io/about/

  • Current or intended projects in relation with IPAL

Efficient Fine-tuning of Large Language Models using Influence Functions: In this project, we investigate using influence functions to provide a data-centric method to improve the capabilities of open-source Large Language Models on diverse instruction-following tasks. We do this by developing an algorithm to find a significantly smaller subset of the original fine-tuning data, called the coreset, and fine-tune the base model on that coreset. We demonstrate that we can fine-tune better models with less data.

Analyzing explanations in a Human-In-The-Loop Reasoning / Planning System: In this project, we analyze various ways a human user can use explanations within a Human-In-The-Loop System, especially when the user is a critic of the AI model outputs and the system iterates to produce an ideal solution. We are especially interested in how this framework informs a user’s level of trust in the AI model, their ability to collaborate to solve complex reasoning tasks, and the overall user experience.

  • IPAL Research Theme

– Theme 1: Explainable and Trustable AI

– Theme 2: AI & HCI (Augmented Human, Augmented Cognition, etc.)

– Theme 3: Natural Language Processing

– Theme 5: Efficient AI

  • French and/or Singaporean collaborators

Philippe Muller (IRIT)

Brian Lim (NUS)

Nicholas Asher (CNRS)

Nancy Chen (A*STAR)