CaReLearner: Causal Reasoning with Logical Interpretable Learning

Researchers are accumulating increasing evidence that machine learning models based on correlation are brittle. These models do not distinguish between correlation and causation. As such, they provide limited insight and cannot reason about the effects of actions. The traditional fields of logical engineering, causal inference, and deep learning have struggled with learning causal models at scale for several decades, each approach facing substantial obstacles. The CaReLearner project seeks to adopt a unified approach to address the core challenges of each paradigm. Building on the Tsetlin machine, the project deals with the unresolved learning challenge in logical engineering, the scaling challenge of probabilistic causal models, and the correlation-reliance of deep learning. We will thus bring about causal and counterfactual reasoning at scale. This entails interconnecting images, time series, tabular data, and natural language text across time and space on challenging multi-modal datasets from knowledge-intensive domains. A medical case will validate the degree to which the learned models capture causal mechanisms through multi-modal observations assessed by human domain experts. In the first phase of the project, we have initiated research in three directions: 1) Multi-modal Tsetlin machines that combine text and image; 2) Scalable learning of Bayesian networks from data; and 3) Modeling and analysis of how the Tsetlin machine captures uncertainty.

Researchers are accumulating increasing evidence that machine learning models based on correlation are brittle. These models do not distinguish between correlation and causation, and, thus, they provide limited insight and cannot reason about the effects of actions for decision-making. The traditional fields of logical engineering, causal inference, and deep learning have struggled with learning causal models at scale for several decades, each approach facing substantial obstacles. The CaReLearner project seeks to adopt a unified approach to address the core challenges from each paradigm, dealing with the unresolved learning challenge in logical engineering, the scaling challenge of probabilistic causal models, and the correlation-reliance of deep learning. The unified approach merges: (1) recursive logical Horne clauses for modeling all possible functions that can be calculated; (2) causal Tsetlin machine learning for distilling causal mechanisms from data into causal Horn clauses; (3) large-scale probabilistic causal reasoning over sparse truth tables in Horn clause form. We will evaluate the ability to do causal and counterfactual reasoning at scale, which entails interconnecting images, time series, tabular data, and natural language text across time and space, on challenging multi-modal datasets from knowledge-intensive domains involving safety-critical decision-making. Compared to deep learning architectures, the performance target measures are: (1) surpassing or maintaining accuracy, (2) reducing energy by 100x, (3) reducing memory by 10x, and (4) reducing the overall computation cost by 10x. A medical case will validate the degree to which the learned models capture causal mechanisms through multi-modal observations, assessed by human domain experts. Overall, the most far-reaching impact of CaReLearner will be to synergize the above three fields, unifying the artificial intelligence community under a new research paradigm.