A0489
Title: Harnessing geometric signatures in causal representation learning
Authors: Yixin Wang - University of Michigan (United States) [presenting]
Abstract: Causal representation learning aims to extract high-level latent factors from low-level sensory data. Existing methods often identify these latent factors by assuming they are statistically independent. However, correlations between latent factors are prevalent across applications. It is explored how geometric signatures of latent causal factors can facilitate causal representation learning without any assumptions about their distributions or dependency structure. The key observation is that the absence of causal connections between latent causal factors often carries geometric signatures of the latent factors' support (i.e. what values each latent can possibly take). Leveraging this fact, latent causal factors are identified for permutation and scaling with data from perfect do interventions. Moreover, block affine identification with data can be achieved from imperfect interventions. These results highlight the unique power of geometric signatures in causal representation learning.
