Arnau Marín-Llobet

Arnau Marín-Llobet

Round Pushpin United StatesGraduation Cap PhD StudentClassical Building Harvard UniversityBooks Machine Learning/Technical AI, Neuroscience/Cognitive Science


Project

ALIGN: Assessing Learning and Internal Geometry of Neural Vision Models through Human Data

Project Overview

AI vision models learn rich representations that can encode complex concepts, including basic-level object and social categories. Understanding how these representations align or diverge from human conceptual representation and social perception is a question of both scientific and ethical importance. This project investigates the internal geometry of conceptual representations, including social concepts, in large-scale vision models, contrasting self-supervised models (e.g., DINOv2) with label-supervised (e.g., ResNet50) and language-supervised counterparts (e.g., CLIP, SigLIP). Using Representational Similarity Analysis (RSA) and other measures of concept geometry, we ask how well each model's representational space reflects the structure of human conceptual structure – including both within and between-category structure, demographic category perception and social trait attributions – as captured by established human behavioral norms and neural activity. 
Measured human reference grounds the comparisons, moving beyond prior work that relies on language-based probing. We maintain principled distinctions: for example, demographic category recovery is a legitimate measure of model competence, whereas trait over-generalization is evidence of shared bias and should never be framed as a target to maximize. By systematically comparing models with different training objectives and levels of linguistic supervision, we can characterize which aspects of conceptual geometry emerge from perception alone versus grounding in language and semantics.

Mentor

Name
Title
Afiliation
 Michael J Tarr 

https://slite.com/api/files/V6IWN12Glogemn/Michael%20Tarr.png?apiToken=eyJhbGciOiJIUzI1NiIsImtpZCI6IjIwMjMtMDUtMDQifQ.eyJzY29wZSI6Im5vdGUtZXhwb3J0IiwibmlkIjoiRmpKZDZ4cjNGcmx3OTYiLCJpYXQiOjE3ODk0NDA1OTEsImlzcyI6Imh0dHBzOi8vc2xpdGUuY29tIiwianRpIjoiT1ZyNThwT1p0VmJXODIiLCJleHAiOjE3OTIwMzI1OTF9.UbNvGeiVjHlkXJ922f8iWNdKdPYU9GmBYm2YyKe8JdY
Carnegie Mellon University, USA
Collaborators:
  • Jorge Almeida – Coimbra
  • Mahzarin Banaji – Harvard


About the Scholar

I am a PhD candidate at Harvard working mostly in Computational Neuroscience. Before that, I received my BSc in Electrical Engineering from the UPC-BarcelonaTech in 2022. My research specifically centers on understanding minds and reasoning in biological and artificial systems, with an emphasis on mechanistic, autonomous and causal understanding. I combine electrophysiology with computational methods to probe brain circuit mechanisms. Recently, I’ve been thinking on how to extend or adapt these autonomous interpretability-driven approaches to large language and vision models to support safer and more reliable AI systems.


Links

Link LinkedIn: arnau-marin-llobet
Link GitHub:  https://linktr.ee/alirezakr