Project description
Researchers are gathering increasing evidence that machine learning models based on correlation are fragile. 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 logic, causal reasoning, and deep learning have struggled for decades to learn large-scale causal models. Each approach faces significant challenges. The CaReLearner project aims to develop a unified approach to address the core challenges of each paradigm.
The project builds on the Tsetlin Machine and tackles the unresolved learning challenge in logic, the scalability challenge in causal reasoning, and the correlation dependence in deep learning. This will enable large-scale causal and counterfactual reasoning. It involves linking images, time series, tabular data, and text across time and space to handle complex multimodal datasets from knowledge-intensive application domains. A medical dataset will be used to validate the extent to which the learned models capture causal mechanisms through multimodal observations assessed by domain experts.
In the first phase of the project, we have initiated research in three directions:
- Multimodal Tsetlin Machines that combine text and images;
- Scalable learning of Bayesian networks from data; and
- Modeling and analysis of how the Tsetlin Machine captures uncertainty.
