Paul, L., Ullman, T. D., De Freitas, J., & Tenenbaum, J.
(2023, July 8). PsyArXiv
https://doi.org/10.31234/osf.io/vzwrn
Abstract
To think for yourself, you need to be able to solve new and unexpected problems. This requires you to identify the space of possible environments you could be in, locate yourself in the relevant one, and frame the new problem as it exists relative to your location in this new environment. Combining thought experiments with a series of self-orientation games, we explore the way that intelligent human agents perform this computational feat by “centering” themselves: orienting themselves perceptually and cognitively in an environment, while simultaneously holding a representation of themselves as an agent in that environment. When faced with an unexpected problem, human agents can shift their perceptual and cognitive center from one location in a space to another, or “re-center”, in order to reframe a problem, giving them a distinctive type of cognitive flexibility. We define the computational ability to center (and re-center) as “having a self,” and propose that implementing this type of computational ability in machines could be an important step towards building a truly intelligent artificial agent that could “think for itself”. We then develop a conceptually robust, empirically viable, engineering-friendly implementation of our proposal, drawing on well established frameworks in cognition, philosophy, and computer science for thinking, planning, and agency.
The computational structure of the self is a key component of human intelligence. They propose a framework for reverse-engineering the self, drawing on work in cognition, philosophy, and computer science.
The authors argue that the self is a computational agent that is able to learn and think for itself. This agent has a number of key abilities, including:
- The ability to represent the world and its own actions.
- The ability to plan and make decisions.
- The ability to learn from experience.
- The ability to have a sense of self.
The authors argue that these abilities can be modeled as a POMDP, a type of mathematical model that is used to represent sequential decision-making processes in which the decision-maker does not have complete information about the environment. They propose a number of methods for reverse-engineering the self, including:
- Using data from brain imaging studies to identify the neural correlates of self-related processes.
- Using computational models of human decision-making to test hypotheses about the computational structure of the self.
- Using philosophical analysis to clarify the nature of self-related concepts.
The authors argue that reverse-engineering the self is a promising approach to understanding human intelligence and developing artificial intelligence systems that are capable of thinking for themselves.
