When causal inference meets deep learning

Bayesian networks can capture causal relations, but learning such a network from data is NP-hard. Recent work has made it possible to approximate this problem as a continuous optimization task that can be solved efficiently with well-established numerical techniques.