How vision-language and generative models represent, name and reproduce color, and how this compares with human color perception. Supervised by Alexandra Gomez-Villa at the Computer Vision Center.
What changes inside a diffusion model when it is adapted, and how to measure it. In collaboration with Jesus Malo.
Fine-tuning or unlearning a text-to-image diffusion model also damages concepts it was never meant to touch, and standard metrics such as FID and KID do not show it. DriftScope takes two model checkpoints and returns a ranked list of the tokens whose visual concepts shifted the most, without real data or access to model internals.
Camera traps and acoustic sensors produce large amounts of data for biodiversity monitoring, but models trained in one place often fail in another. My master's work studied how foundation models and multimodal learning can close that gap. It included research stays at Télécom Paris (domain adaptation) and Universidad de los Andes (wildlife image segmentation).
Models trained on camera-trap images from one region lose accuracy when deployed in another. WildIng learns representations that stay stable across geographical domains.
A contrastive model that uses text descriptions from foundation models to guide camera-trap image recognition, so it transfers better to new camera locations.
A comparison of audio and image models for monitoring the critically endangered Antioquia brushfinch with very little labeled data. This was my undergraduate thesis.