Beyond Curve Fitting: Causation, Counterfactuals, and Imagination-based AI

In recent years, Artificial Intelligence and Machine Learning have received enormous attention from the general public, primarily because of the successful application of deep neural networks in computer vision, natural language processing, and game playing (more notably through reinforcement learning). We see AI recognizing faces with high accuracy, Alexa answering English spoken questions efficiently, and Alpha-Zero beating Go grandmasters. These are impressive achievements, almost unimaginable a few years ago. Despite the progress, there is a growing segment of the scientific community that questions whether these successes can be extrapolated to create general AI without a major retooling. Prominent scholars voice concerns that some critical pieces of the AI-puzzle are still pretty much missing. For example, Judea Pearl, who championed probabilistic reasoning in AI and causal inference, recently said in an interview: “To build truly intelligent machines, teach them cause and effect” (link). In a recent OpEd in the New York Times, Cognitive Scientist Gary Marcus noted: “Causal relationships are where contemporary machine learning techniques start to stumble” (link).

These and other critical views regarding different aspects of the machine learning toolbox, however, are not a matter of speculation or personal taste, but a product of mathematical analyses concerning the intrinsic limitations of data-centric systems that are not guided by explicit models of reality. Such systems may excel in learning highly complex functions connecting input X to an output Y, but are unable to reason about cause and effect relations or environment changes, be they due to external actions or acts of imagination. Nor can they provide explanations for novel eventualities, or guarantee safety and fairness. This symposium will focus on integrating aspects of causal inference with those of machine learning, recognizing that the capacity to reason about cause and effect is critical in achieving human-friendly AI. Despite its centrality in scientific inferences and commonsense thinking, this capacity has been largely overlooked in ML, most likely because it requires a language of its own, beyond classical statistics and standard logics. Such languages are available today and promise to yield more explainable, robust, and generalizable intelligent systems.

Our aim is to bring together researchers to discuss the integration of causal, counterfactual, and imagination-based reasoning into data science, building a richer framework for research and a new horizon of applications in the coming decades. Our discussion will be inspired by the Ladder of Causation and Structural Causal Models (SCM) architecture, which unifies existing approaches to causation and formalizes the capabilities and limitations of different types of causal expressions (link). This architecture provides a general framework of integrating the current correlation-based data mining methods (level 1) with causal or interventional analysis (level 2), and counterfactual or imagination-based reasoning (level 3). We welcome researchers from all relevant disciplines, including, but not limited to, computer science, cognitive science, economics, social sciences, medicine, health sciences, engineering, mathematics, statistics, and philosophy.