Identifying Coarse-grained Independent Causal Mechanisms with Self-supervision

Current approaches for learning disentangled representations assume that independent latent variables generate the data through a single data generation process. In contrast, this manuscript considers independent causal mechanisms (ICM), which, unlike disentangled representations, directly model multiple data generation processes in a coarse granularity. In this work, we aim to learn a model that isolates each mechanism and approximates the ground-truth ICM from observational data. We outline sufficient conditions under which the ICM can be learned and isolated using a single self-supervised generative model with a mixture prior, simplifying previous methods. Moreover, we implement a generative model with an identifiable structural latent space by combining the ICM with a shared latent space. We compare this ICM approach to disentangled representations on various downstream tasks, showing that the ICM is more robust to intervention, co-variant shift, and noise due to the isolation between the data generation processes.