Causal AI: from prediction to understanding

Machine learning (ML) has achieved remarkable success in developing predictive models for diverse applications, from recommending movies to forecasting weather patterns. However, current AI systems are almost invariably driven by data, based on probabilistic/statistical machinery, and often lack in explainability, robustness, and adaptability. There is a growing understanding that statistical associations cannot predict what happens if the environment changes or some external interventions occur, and that robust decision-making requires some knowledge of the causal mechanisms underlying the environment. This talk delves into causal AI, a field about enriching AI systems with causal reasoning capabilities, exploring its potential to move AI beyond prediction to understanding. I’ll share my research on causal inference, particularly our recent work on inferring causal effects from data, a fundamental task in various scientific fields. We study the problem of estimating a target causal effect from a combination of observational and experimental data and the underlying causal structure. We develop a novel estimator with double/debiased machine learning (DML) properties that offers robustness against model misspecification and debiasedness against slow convergence in nuisance function estimation, thus allowing for the use of modern machine learning techniques for estimating nuisances.

Post Talk Link:  Click Here 

Passcode: dk7?#2cI