Dreyfus

Dreyfus claims that following rules, as is done by expert systems is not enough. It can't capture the richness, the ``thickness'' of human experience. See Mind over Machine by Dreyfus and Dreyfus (PSY KH:D 081). For a summary, see Crevier pages 127-132.

Part of the Dreyfus' argument is based on their five-stage model of progress from novice to expert (page 22 of Mind over Machine)

1. Novice. Operates by consciously-learnt context-free rules. Lacks any sense of the overall task.

2. Advanced beginner. Uses more sphisticated rules, which refer to situational elements as well as context-free ones. These situational elements are features such as the pattern of behviour which distinguishes a drunken from a sober driver. They're learnt by experience, and the advanced beginner can't formalise them.

3. Competent. Has now learnt to recognise many many context-free and situational elements. Still lacks any sense of their overall importance to the task, and rapidly becomes overwhelmed. Tries to overcome this by hierarchical goal-based planning. This hierarchical decomposition of the task means that, at any time, the competent pays attention only to that small number of features relevant to a particular subgoal, thus avoiding being overwhelmed.

4. Proficient. Most of the time, now performs his task intuitively, without analytical thought. But this deep involvement in the task will be broken when certain elements present themselves as particularly important. The proficient then stops and thinks analytically about what to do next.

5. Expert. Performs his task intuitively, almost all the time. Occasionally has to stop and deliberate, but this involves critical reflection on his intuitions, rather than goal-based planning.

So this is a progression from rule-based problem-solving to a different approach based on matching against past experiences. The first is similar to classical symbolic AI, which divides problems down into bits, divides the job between different components, and puts the results together. The second involves some kind of holistic pattern recognition.

Classical AI assumes the mind also divides problems down into bits, divides the job between different components, and puts the results together. Hence it's not a good model for expert performance. Digital computers require one to organise tasks in this way: hence they can't achieve expert performance. In Dreyfus' terms, such systems he calls ``machines''.

What is required for holistic pattern recognition is something like a holographic recogniser (page 60), example in Transforming the prospects for robot vision, AI photocopy B165. Like computers, such systems are physical systems. Unlike computers, they don't divide the job up into separate pieces, giving one to each component. Dreyfus does not call systems of this kind machines.

Dreyfus does not claim machines (in our sense) can't behave intelligently; just that digital computers and other machines (in his sense) can't. He does not deny the brain is a machine (in our sense).