Computationally-Embedded Perspective on Continual Learning

This paper introduces a novel perspective on continual learning by framing the agent as a computationally-embedded automaton within a universal computer. This approach provides a new way to understand and address the challenges of continual learning, particularly in the context of the 'big world hypothesis'. The paper's strength lies in its theoretical foundation, establishing a connection between embedded agents and partially observable Markov decision processes. The proposed 'interactivity' objective and the model-based reinforcement learning algorithm offer a concrete framework for evaluating and improving continual learning capabilities. The comparison between deep linear and nonlinear networks provides valuable insights into the impact of model capacity on sustained interactivity.

• •Proposes a novel perspective on continual learning by embedding the agent within a universal computer.
• •Introduces the 'interactivity' objective to measure an agent's ability to adapt.
• •Develops a model-based reinforcement learning algorithm for interactivity-seeking.
• •Finds that deep linear networks sustain higher interactivity than deep nonlinear networks as capacity increases.