Currently thinking about identifiability of latent representations, and other necessary conditions for causal representation learning.
