Intro. Video: https://youtu.be/yqEuCho6V8Y
- Project Goal -
The goal of the Causal Curiosity Project is continuous and open-ended causal symbolic learning.
Causal symbolic learning lies at the intersection of two branches of AI - reinforcement learning and causal inference. It represents an approach to AI which prioritizes discovering the causal connections between actions and objects in an agent’s environment.
This is achieved probabilistically through interventions taken by the agent and then storing the results in a causal graph. It is a combination of both intrinsic curiosity in the reinforcement learning sense, that is exploration for its own sake, and the systematic recognition of causal connections through trial and error learning. Causally connected relationships in the agent’s environment are discovered through reinforcement learning and then expressed in terms of DAGs or causal equations that are understandable to humans.
- Relevance -
Most real world problems to which an embodied AI such as a robot might be tasked will take place in non-stationary Markov environments, environments which are dynamic and evolving over time. Lacking a causal symbolic understanding of its own actions and how these relate to objects in the environment, an agent would neither be able to develop good priors, what we might call “common sense” or the ability express its decisions to humans in a way that can be interrogated or reverse engineered.
In this sense causal curiosity has the potential to enable explainable AI, where we can lift the hood on the results of deep learning and view the causal connections that the agent is forming. Causal symbolic AI is therefore critical to enabling both the kind of general artificial intelligence that can operate in real world environments and one that is capable of communicating those decisions to humans.
- Supporting Literature and References -
Have questions, reach out at contact below.
Email: aaron.krumins@gmail.com
Or check out our other offerings: