Levels of explanation

A specific theory in functionalism may posit certain structures and processes and it may make predictions, say, about behaviour, reaction time or error patterns. W.r.t. structures and processes, several levels of explanation are possible.

Daniel Dennett's (1978) levels of increasing abstraction for the description of an information-processing device:

1. The Physical (physical architecture) (need not go lower to atoms etc. as this level is already acceptable to physical science)

2. The Design (program implemented by this architecture)

3. The Intentional (``ascribing to the system the possession of certain information and supposing it to be directed by certain goals'' (Dennett (1978) Brainstorms. Cambridge: MIT Press; p.6).)

   Intentional = Directed upon

   Intentional content of a mental state = its semantic properties i.e. propositions/ symbols/ representations

   Intentional systems (Dennett, 1971; 1978): Have rationality and purpose, beliefs and desires, i.e. have at their disposal goals, procedures and strategies. You would try to outwit a program just as you would try to outwit a person.

   Intentional systems are supposed to emerge at a certain level of complexity. From a methodological viewpoint complex systems are better studied at an intentional level.

Marr in 1982 identified essentially the same levels in his classic analysis of levels of description for an information-processing device: (More on Marr's work in Perception in Semester 2)

1. The Physical

2. The Algorithmic

3. The Computational (most abstract level, involves an analysis of the problem-situation, of goals, strategies)

Stillings et. al., 1989 pp. 328-332 example of a chess-playing computer - following Dennett:

1. Hardware strategy (transistors, chips, cogs) - holds promise of accurate predictions but may be impracticable, non-intuitive and would not work for a different hardware implementation of a chess playing machine.

2. Program strategy (subroutines, addresses) - Would assign the same explanation to two physically dissimilar but computationally equivalent computers. But two programs using the same strategies though written in different programming languages would constitute different ``explanations'' for chess playing. Besides it may still be too complicated and would fail in case of hardware malfunction.

3. At the most abstract and most intuitive level is the mentalistic strategy (plans, goals, desires, beliefs, knowledge) - easy to test and implement, will generalise to other hardwares, but unable to handle both hardware and software bugs. Predictive power is lost, truth is traded for convenience.

The three levels identified by Dennett/Marr are strictly independent. For example, when discussing the computations used by a device, we are not interested in the algorithms it employs or its physical structure.

Which of the levels is Psychologically the most important (for understanding an information-processing system)?

For Marr the target of explanation is the algorithmic level, but he claims that this level is best approached from the top-down: by looking at the problem-situation confronting the system. Marr argues that the nature of the algorithms underlying cognition are more fully constrained by the computational problems they confront, than by the physical structure of the system in which they are implemented: given that the algorithmic level is not open to direct observation, the most promising approach for an understanding of the algorithmic level is in a top-down manner. Dennett agrees that we ultimately desire an explanation at the algorithmic level. To stop at the intentional level is to leave intelligence and rationality unexplained.

Explanations at these two higher levels have had considerable success, fruitful research programs have emerged, but ontological considerations show that the problem of mind-body dualism is unresolved.

The above descriptions of levels neglect the role of Anthropology - the cultural or sociological level of explanation is presumably considered to be outside the scope of cognitive science. Gardner's proposal for the three levels is an exception:

1. Cultural/ Sociological (neglected by the others, highlighted in SitCog - see sec. )

2. Representational

3. Biological, specifically, Neurological (has gained importance in the last 15 years, since Gardner's 1984/ publ. '85/ epilogue '87 MNS)