Machine Unlearning in Hyperbolic vs. Euclidean Multimodal Contrastive Learning: Adapting Alignment Calibration to MERU

Àlex Pujol Vidal, Kamal Nasrollahi, Thomas B. Moeslund, Sergio Escalera; Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 1644-1653

Machine unlearning methods have become increasingly important for selective removal of harmful concepts in large multimodal models. While recent work has explored unlearning in Euclidean contrastive vision-language models, the effectiveness of concept removal in hyperbolic spaces remains unexplored. This paper explores this gap by adapting Alignment Calibration to MERU, a model that embeds images and text in hyperbolic space to better capture semantic hierarchies. Through systematic experiments and ablation studies, we demonstrate that hyperbolic geometry offers distinct advantages for concept removal, achieving lower accuracy at zero classification in forget set with reasonable performance on retain set, particularly when scaling to multiple concept removal. Our approach introduces hyperbolic-specific components including entailment calibration and norm regularization that leverage the unique properties of hyperbolic space. Comparative analysis with Euclidean models reveals fundamental differences in unlearning dynamics, with hyperbolic unlearning reorganizing the semantic hierarchy while Euclidean approaches merely disconnect cross-modal associations. These findings provide insights into the geometric properties that influence concept removal, setting a direction towards better control and safer deployment for large multimodal models. Code available at https://github.com/alex-pv01/HAC.