Search Papers | Poster Sessions | All Posters
Poster B62 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Do diffusion models generalize on abstract rules for reasoning?
Binxu Wang1 (), Jiaqi Shang1, Haim Sompolinsky1,2; 1Harvard University, 2Hebrew University
Can diffusion models trained on a sample dataset acquire abstract relational reasoning ability? To explore this question, we train diffusion models on the RPM visual reasoning dataset. We find that diffusion models are capable of generating novel samples conforming to relational rules without directly memorizing training data. Moreover, the models successfully generate samples that conform to rules of similar structure unseen in training, suggesting generalization in the abstract relation space. Notably, the models exhibit ordered learning dynamics in rule acquisition, with local data structure learned earlier than global structure.
Keywords: visual reasoning generative model diffusion model abstract rule