What is intelligence and why is it needed? Current research in the cognitive and life sciences presents only fragmented views on this question. We examine the recent proposal that intelligence can be understood as accurate prediction, and develop it behaviourally to include adaptive control. We argue that this framework, albeit already applied to diverse domains of intelligence research, needs conceptual clarification. We specifically refine the notions of “accuracy” in terms of practically exploitable representation and “prediction” in terms of action-oriented probabilistic inferences at the subpersonal level of information processing. We then argue that this interpretation does not fully explain the mechanisms of intelligence in artificial systems, but that it nevertheless provides a useful analysis. Firstly, this view allows us to demarcate cognition from intelligence, the latter of which comes out as a more sophisticated form of efficient and robust information processing that may be realised in systems that are non-cognitive (e.g., bacteria and large-language models). Secondly, this view provides a unified platform for researchers from different disciplines to integrate their diverse theoretical perspectives, share fragmented data, and effectively discover intelligence in non-human systems. We also explore some obstacles and avenues to scaling the view up to adequately capture intelligence in social and collective systems.
Our framework ties the different notions of intelligence together in a natural way by integrating them into a control theory framework of adaptive control and model-based reinforcement learning. We posit that memory, knowledge, experience, and understanding can be interpreted as having to do with an adaptive controller’s predictive model, where memory pertains to the ability to store sensory traces, knowledge indicates that memory traces are structured and consolidated to afford effective prediction, and experience indicates the sensory traces in the context of an agent being situated and continuously sampling the world, learning both its environment’s structure and statistics, as well as the statistics of its own body. So long as it is the case that there is uncertainty both about the state of the environment and which behaviour best achieves the agent’s goals, the agent needs to make decisions and choices based on a value system where some outcomes are better or worse than others. Together, these constitute the ability of the agent to make judgements, weighting and prioritising behaviours and outcomes in relation to its internal needs.
