Current interests (March 2017)
This is a list of ideas related to AGI that I’ve been thinking about in the recent weeks or months. I love it when people give me reading recommendations, so please let me know if you have any.
- A theory of beliefs. In artificial cognitive systems, this would help justify certain architectures and development methods. What I call a belief is anything that corresponds somewhat to the conversion from a probability to an assumption of truth.
- Actions as beliefs. A belief would be represented by a triple (concept, action, probability) and whenever the concept exists with a sufficiently high probability, the discrete action is triggered. The action in this context is not necessarily on the outside world but could be directly on the state of the system. The belief is that the action must be triggered when the concept exists and the probability is above some threshold.
- Optimization in non-euclidean spaces of unknown topology and topology discovery. The general problem is to speed up the search for optimal pairings between concepts and actions without knowing a priori which pairs of actions most of the time yield similar outcomes.
- Theories and methods related to chaos and patterns. It seems valuable for a cognitive system to learn patterns with good predictive power. On one hypothetical extreme, one could imagine someone who only remembers very abstract patterns and doesn’t remember anything about individuals. This person would very good at understanding anyone’s new situation on the spot without knowing them, but would not remember personal details about people because they’re not sufficiently reusable. The human brain seems to be doing this well and it would be good to have a solid understanding on what kind of patterns and concepts should be constructed in priority.
- Mapping between models of the world, with or without communication between systems maintaining these models. A related question is what path takes a model when it evolves and acquires greater confidence surrounding old concepts as well as new concepts replacing old ones.