Conjecture—Amortized reasoning with on-demand inference as an essential architectural ingredient

Determine whether architectures that combine efficient amortized reasoning with introspective, on-demand deliberate inference are essential for robust, adaptable action and decision-making in embodied agents.

Background

Inspired by Dual Process Theory, the authors posit that embodied agents should employ both fast learned responses and slower deliberate reasoning.

They conjecture that the ability to amortize reasoning while retaining capacity for deeper inference when needed is an essential architectural ingredient.

References

We conjecture that an essential ingredient of such an architecture is the ability to effectively amortize reasoning while maintaining the introspection and capability to perform more elaborate inference on demand. We address this conjecture in section \ref{sec:fast-slow}.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 1 (Introduction: Guiding Questions)