Causal Discovery and Injection for Feed-Forward Neural Networks

Neural networks have proven to be effective at solving a wide range of problems but it is often unclear whether they learn any meaningful causal relationship: this poses a problem for the robustness of neural network models and their use for high-stakes decisions. We propose a novel method overcoming this issue by injecting knowledge in the form of (possibly partial) causal graphs into feed-forward neural networks, so that the learnt model is guaranteed to conform to the graph, hence adhering to expert knowledge. This knowledge may be given up-front or during the learning process, to improve the model through human-AI collaboration. We apply our method to synthetic and real (tabular) data showing that it is robust against noise and can improve causal discovery and prediction performance in low data regimes.