Reflective Artificial Intelligence

TL;DR

This paper asks what reflective AI might look like, drawing on notions of reflection in complex systems, cognitive science, and agents, and sketches an architecture for reflective AI agents, and highlights ways forward.

Abstract

Artificial Intelligence (AI) is about making computers that do the sorts of things that minds can do, and as we progress towards this goal, we tend to increasingly delegate human tasks to machines. However, AI systems usually do these tasks with an unusual imbalance of insight and understanding: new, deeper insights are present, yet many important qualities that a human mind would have previously brought to the activity are utterly absent. Therefore, it is crucial to ask which features of minds have we replicated, which are missing, and if that matters. One core feature that humans bring to tasks, when dealing with the ambiguity, emergent knowledge, and social context presented by the world, is reflection. Yet this capability is utterly missing from current mainstream AI. In this paper we ask what reflective AI might look like. Then, drawing on notions of reflection in complex systems, cognitive science, and agents, we sketch an architecture for reflective AI agents, and highlight ways forward.

Figures (8)

• Figure 1: Kolb's 'Experiential Learning Model'. Source: Kolb:1984. The model shows captures the cognitive cycle in humans that is responsible for learning from experience.
• Figure 2: Critic Agent Architecture russell2021artificial. We introduce this architecture as a baseline AI architecture that manages to capture various aspects of perceiving, learning, planning, reasoning and acting as different qualitative processes. One can visually contrast this architecture with with the other mainstream architectures, old and new, in AI.
• Figure 3: ANN architecture (left, by Colin Burnett from https://en.wikipedia.org/wiki/Connectionism) and GAN goodfellow2014generative. Considering these common machine learning architectures, it is clear that there is a lack of any reflective 'loop'. Although these achieve different outcomes, they are qualitatively equivalent in the sense that they both operate at a single level of abstraction when it comes to information processing. There is no self-reference: the loops in both cases are for feedback, in much the same way that the Critic Agent operates. Additionally, even though Kolb's model of experiential learning is a model of learning in humans, it also presents (albeit at a high level) qualitative processes that ANNs and GANs do not.
• Figure 4: PRS architecture (Source: georgeff1987reactive). The PRS architecture is similar to the Critic agent architecture in the sense that it allows us to break down different qualitative processes. The difference between the Critic architecture and the PRS is that the PRS does not include a learning component, but it has a richer representation of the processes and elements responsible with driving the reasoning behind the actions that are executed in the environment. PRS also allows for an eventual learning component to be plugged into the system interface which feeds data into the belief base.
• Figure 5: ACT-R architecture anderson1997act. It is crucial to note that ACT-R is not an AI agent architecture, rather a cognitive architecture that was used as an expert system. The original purpose of ACT-R is to map and understand human cognition as a set of modular components that execute procedures to produce behaviour in a specific domain. ACT-R assumes that all cognitive components are represented and driven by declarative and procedural memory.