Re: two important articles on cognition and implications for KR&R for AI

> On 15 Nov 2022, at 06:00, ProjectParadigm-ICT-Program <metadataportals@yahoo.com> wrote:
> A New Brain Model Could Pave the Way for Conscious AI
> 
> https://scitechdaily.com/a-new-brain-model-could-pave-the-way-for-conscious-ai/
> 


The model in the above work is not dissimilar to what I’ve been exploring:

In the first stage, evolution ensures that animals can select appropriate actions in response to the immediate environment and their internal drives.  This involves complicated systems for perception and action, e.g. driving development of the cerebellum for real-time sensory-motor control, as well as hierarchies of perception that can detect when things aren’t behaving as expected, which entails learning how to predict behaviour at different levels of abstraction, with the help of the cerebellum.

The second stage is the evolution of System 1 cognition that can operate on memory as well as on perception, freeing the mind from fixating on the present. This allows for cognition about the past as well as potential futures, albeit only at an intuitive level that is not open to introspection.

The third stage is the evolution of System 2 cognition enabling deliberative, analytical cognition that is open to introspection, and works in tandem with System 1, as noted by Daniel Kahneman. This allows humans to create complex plans and to solve problems that other species have great difficulties with.

What I am now struggling with is how to mix inferential, i.e. logical everyday thought, as per Alan Collins, with model building, as per Philip Johnson-Laird, who showed that humans mostly reason in terms of mental models rather than logic and statistics. I have implementations for both approaches, but it isn’t yet clear how they fit together along with metacognition.

On the one hand I have a (logical) notation for plausible knowledge with properties, relationships and implications, and on the other hand I have a different notation for chunks and rules, inspired by John Anderson’s work on ACT-R.  Chunk rules are procedural with the means to manipulate knowledge expressed as chunks, including live models generated by perception, and to delegate actions to a separate system capable of real-time coordination. This poses a strong contrast between logical and procedural reasoning.

Metacognition is at a higher level, and deals with reasoning about reasoning, including reasoning how to learn to reason. This points the way to how to integrate logical and procedural reasoning. The next step is to gather some examples that can shine a light on the details. Working implementations then provide an important means to test the consistency and effectiveness of theories. You can’t find this stuff in existing work, though, which is at the same time frustrating and exciting. 

Best regards,

Dave Raggett <dsr@w3.org>