Heiddeggerian Artificial Intelligence Part 2

After reading Hubert L Dreyfus paper on Heiddeggerian Artificial Intelligence I was left with many impressions and thoughts, but amidst it all I had the feeling that he is onto something. Dreyfus is a philosopher and is thus inclined to speak in very general, abstract and impressionistic terms. Take this key comment by Dreyfus, for example:

Rather, acting is experienced as a steady flow of skillful activity in response to one's sense of the situation Part of that experience is a sense that when one's situation deviates from some optimal body-environment gestalt, one's activity takes one closer to that optimum and thereby relieves the "tension" of the deviation. One does not need to know what that optimum is in order to move towards it. One's body is simply solicited by the situation [the gradient of the situation’s reward] to lower the tension. Minimum tension is correlated with achieving an optimal grip.

I think I understand that: The goals one aims for are not literally envisaged by our minds but are implicit in that one is less aware of goals than how one is supposed to move toward them. The situations confronting us cause a response but not necessarily by way of an internalized model that allows us to envisage the goal of that response. However, if Dreyfus’ ideas are to be realized in hardware and software how does one reduce terms like “optimal body-environment gestalt”, “tension”, “soliciting” “optimal grip” to bits, bytes and “if then elses”?

From the outset I was immediately attracted to Dreyfus phenomenology: phenomenology is a philosophy founded in the realization that our experience of the world and our thoughts about it are to all intents and purposes the extent of our cosmos. This existential philosophy side steps the horrendously intractable Cartesian problem surrounding the ontological distinction (if any) of noumena and cognita by positing conscious cognition as the effective center of our cosmos. From Dreyfus phenomenology follows his starting observation:

[T] he meaningful objects ... among which we live are not a model of the world stored in our mind or brain; they are the world itself.

Dreyfus starting point makes for a huge economy in his version of AI: A complex model of the world does not need to be carried around in our heads when in fact our experience of that world, delivered to our brains by our senses, will probably serve far better: all we are asked to do is to react to that perceived model and not to exhaustively envisage it. The real world is in effect our 'core memory' and we are but the depository of the neural 'algorithm' that tells as how to react to the contents of that 'memory' as we move around our world. In particular, if the world itself is our model then we don’t have to internalize a model of how it reacts to our actions. In principle our actions could have vast and ramifying effects on the rest of the world and a comprehensive internal model of reality would have to include the logic required to model these knock on effects; the problem of trying to somehow cater for the possibility of these escalating effects is called the “frame problem”.

Dreyfus looks to be a fairly abrasive character and uses general terms that can be very slippery. It is easy to misinterpret him and anyone who tries to use Dreyfus' ideas and turn them into something workable is probably taking a similar risk to those who attempt to articulate the meaning of the Holy Trinity and open themselves up to charges of heresy. Dreyfus is a hard task master. He is AI's prophet of doom and as is the prerogative of prophets of the infinitely complex he tends to work apophatically; that is, he is much clearer about what human intelligence is not, rather than what it actually is. Perhaps this is a good thing because there has been so much hype and over optimism in AI that it cries out for a judgmental preacher. It is very difficult to do justice to the infinitely complex and the scientific equivalent of charges of blasphemy and idolatry as humans attempt to create images of our their own selves reminds us not to be too complacent about progress in the face of simplistic and unrepresentative models. Did I just say ‘model’? Aren’t they the things that Dreyfus says we shouldn’t be using?

My own guess is that human intelligence solves the frame problem in a plurality of ways. The ‘absorbed coping’ that Dreyfus talks about may well be found in intelligent organisms like ourselves; his view is that such organisms are dynamic systems coupled to their environment via stimuli which are not processed using representations and models, but these stimuli succeed in ‘soliciting’ the right responses without the use of representations and models. It is likely that humans have inherited this computationally economic modus operandi. And yet it seems to me that humans also appear to model the world computationally in the internal Cartesian sense: Humans can and do reflect on ‘external systems’ and can anticipate their behavior without coming into contact with them and being prompted by them. However, often this reflection may make use of pencil and paper jottings and various external contrivances that help prompt thinking, thus betraying the roots of human intelligence in organisms coupled dynamically to their environment. If the human mind does do symbolic modeling it may not actually be very good at it as a standalone system.

There is one other characteristic of the human mind suggesting that “Dreyfus is right but....”. In a connectionist model of the mind, everything is connected to everything else through pathways that may be no longer than ~ Log(N), where N is the number of neurons in the brain. Thus when attention is focused on one activity like say language translation, the whole of the mind’s accumulated connectionist experience is never far away in access terms and thus the whole domain of an individuals experience can be brought to bear on a problem. The mind has the potential to use the widest frame available to it and there may be no artificial frame or relevance boundary arbitarily drawn within the domain of one’s experience. In short the human mind may not even attempt to solve the frame and relevance problems and instead throws all its knowledge resources at the situations it meets. Learning never stops and the human mind is therefore always placing contexts within new contexts. Problem solving truly is an open ended activity.

Summary:
Human intelligence uses three modus operandi:
1. Heiddeggerian: Human intelligence uses the world itself as its own model.
2. Modelling: Human intelligence has a mental facility to model, but that facility betrays an inheritance from 'Heiddeggerian' organisms by making frequent use of 'external' world contrivances such as pencil and paper.
3. Isotropy: Human intelligence, via connectionism, does not attempt to impose any a-priori limitations on what knowledge resources are relevant to a situation. That is, it doesn't attempt to solve the frame problem.