87213955efbe48b46586e37bf2f1fe5b-Supplemental-Conference.pdf

However, crucially, we do not make two assumptions used to derive the AVAE objective. Firstly, we do not assume that the decoder is a one-to-one mapping between latent samples and a corresponding generated sample. Interestingly,theDCIresponsibility matrices dooftenresemble theconditioned response matrices, suggesting that relying on correlations instead of a full causal analysis, can yield similar results. Infact, then the learned causal structure estimated using the latent response matrix may be used in tandem to develop a structure-aware disentanglementmetric. Given the complexity of the underlying data manifold, aviable alternativeisbased on riemanian geometry [78]which has previously been investigated for alternativeprobabilistic models likeGaussian Process regression [79].