In this talk, I will present a generic methodology to encode hierarchical causal structure among observed variables into a neural network to improve its prediction performance. The proposed causality-informed neural network (CINN) leverages three coherent steps to systematically map the structural causal knowledge into the layer-to-layer design of neural network while strictly preserving the orientation of every causal relationship. In the first step, CINN discovers causal relationships from observational data via directed acyclic graph (DAG) learning, where causal discovery is recast as a continuous optimization problem to avoid the combinatorial nature. In the second step, the discovered hierarchical causal structure among observed variables is encoded into neural network through a dedicated architecture and customized loss function. By categorizing variables as root, intermediate, and leaf nodes, the hierarchical causal DAG is translated into CINN with a one-to-one correspondence between nodes in the DAG and units in the CINN while maintaining the relative order among these nodes. Regarding the loss function, both intermediate and leaf nodes in the DAG are treated as target outputs during CINN training to drive co-learning of causal relationships among different types of nodes. In the final step, as multiple loss components emerge in CINN, we leverage the projection of conflicting gradients to mitigate gradient interference among the multiple learning tasks. Computational experiments across a broad spectrum of UCI datasets demonstrate substantial advantages of CINN in prediction performance over other state-of-the-art methods. In addition, we conduct an ablation study by incrementally injecting structural and quantitative causal knowledge into neural network to demonstrate their role in enhancing neural network’s prediction performance.

This talk provides a holistic lens on emerging uncertainty quantification (UQ) methods for ML models with a particular focus on neural networks and gives a tutorial-style description of several state-of-the-art UQ methods: Gaussian process regression, Bayesian neural network, neural network ensemble, and deterministic UQ methods focusing on spectral-normalized neural Gaussian process (SNGP). Established upon the mathematical formulations, we subsequently examine the soundness of these UQ methods quantitatively and qualitatively (by a toy regression example) to examine their strengths and shortcomings from different dimensions. Based on the findings of the comparison, we exploit the advantages of SNGP in UQ and develop an uncertainty-aware deep neural network to detect the defects of steel wire rope. Computational experiments and comparisons with state-of-the-art models suggest that the principled uncertainty quantified by SNGP not only substantially enhances the prediction performance, but also provides an essential layer of protection for neural network against out-of-distribution data.