Key Features: Cover three types of neural networks, formulate deep learning as an optimal control problem and use Pontryagin's Maximum Principle for network training; Deep learning for nonlinear mediation and instrumental variable causal analysis.Construction of causal networks is formulated as a continuous optimization problem; Transformer and attention are used to encode-decode graphics. RL is used to infer large causal networks; Use VAE, GAN, neural differential equations, recurrent neural network (RNN) and RL to estimate counterfactual outcomes; AI-based methods for estimation of individualized treatment effect in the presence of network interference.