Saturday, June 16, 2012

Stanley on the Procedural/Declarative Distinction

In Chapter 7 of Know How (2011), Jason Stanley discusses the origins of the distinction between procedural and declarative knowledge.  He argues that the distinction has never had any implications for the question of whether or not knowing how to do something is a kind of propositional knowledge.  In other words, he says the distinction has never implied that knowing how is different in kind from knowing that.  I think he's quite wrong and his argument misrepresents the literature.

Stanley appeals to a paper by Dienes & Perner (D&P) called "A theory of implicit and explicit knowledge" (1999), in which they claim that the distinction originated in AI and was later utilized in psychology.  On page 743, they claim that the distinction "concerned how best to implement knowledge."  They continue:

Should one represent the knowledge that all men are mortal as
a general declaration “for every individual it is true that if that
individual is human it is also mortal”? Whenever knowledge of a
human individual was introduced in the database this general
information would be consulted to infer by general inference rules
that that individual must also be mortal. The alternative would be
to have a specialised inference procedure: “Whenever an
individual is introduced that is human, represent that that individual
is mortal.”
These are two different procedures for producing the same output, which is a truth-evaluable representation of an individual as mortal.  Thus, Stanley concludes that the distinction between procedural and declarative knowledge is not a distinction between truth-evaluable and non-truth-evaluable knowledge.  It is not a distinction between propositional knowledge and non-propositional knowledge, but between different ways of programming a computer to produce the same propositional knowledge states.

Though he does not seem to realize it, Stanley is contradicting D&P's analysis, which claims that procedural knowledge is unique in that it only implicitly contains what is explicit in the declarative version.  They write in the next paragraph, which Stanley ignores:
The analysis also brings out the intuitive meaning of declarative
knowledge as knowledge that declares what is the case (e.g.,
Squire 1992, p. 204: memory whose content can be declared)
because it represents explicitly that something is a fact. Non-
declarative memory can be given precision in our analysis either
as the stronger form of knowledge that does not make predication
explicit or as a weaker form of knowledge that makes predication
explicit but leaves factuality implicit.
When D&P talk about different ways of implementing knowledge, they are talking about different ways AI programmers can implement their own knowledge in the programs.  They can build machines with procedural knowledge or they can build them with declarative knowledge, or both.  In some cases, these can be different ways of deriving the same propositional knowledge states.  Stanley's thesis, however, is that that is always and only what the distinction amounts to.  He ignores the very obvious fact that procedural and declarative knowledge can be different ways of producing intentional action.  The output of the system can be a non-propositional (i.e., not truth-evaluable) performance.

This is explicit in Winograd (1975), whom both D&P and Stanley appeal to.  On page 189, Winograd writes:
It is an obvious fact that many things we know are best seen
as procedures, and it is difficult to describe them in a purely
declarative way. If we want a robot to manipulate a simple
world (such as a table top of toy blocks), we do it most
naturally by describing its manipulations as programs. The
knowledge about building stacks is in the form of a program
to do it.

Winograd goes on to explain that the assumptions of the programmer are "built in" and not explicitly represented in or consulted by the system. He argues that this, the procedural kind of knowledge, has advantages "not only to obvious physical processes like moving blocks around, but equally well to deductive processes like playing games or proving geometry theorems."  So it is clear.  Procedural knowledge and declarative knowledge are different ways of producing behavior or output, but not necessarily the same propositional knowledge states.  The implication of the distinction is that there are two kinds of knowledge relating to two different ways of producing outputs.

Stanley seems to misunderstand this.  On page 208 of Know How, he writes:  "Clearly, Winograd is thinking of the distinction as one between ways knowledge is derived, not  kinds of knowledge states."  I would say that is clearly incorrect.

Winograd clearly and explicitly says that he is talking about different forms of knowledge.  For example, on page 190, Winograd writes:  "It is theoretically possible to express second-order knowledge in a declarative form, but it is extremely difficult to do so outside the context of a particular reasoning process."  And on the next page: "By putting knowledge in a primarily procedural form, we gain the ability to integrate the heuristic knowledge easily."

Winograd is distinguishing between systems which have knowledge built in through programs, and systems which have knowledge represented in a declarative way, as general statements which must be addressed in order for the knowledge to be implemented.  These are different ways of expressing or manifesting knowledge in a machine.  That is, they are different ways of implementing the knowledge we want the machine to have.  This is a distinction between how knowledge is represented in a system.  It's a difference in the kinds of states the system has.  And it's not just a matter of systems which derive propositional knowledge.  It's much more general.

The distinction, as it originated in AI (according to D&P and as demonstrated in Winograd), is this:  machines can have knowledge in the form of procedures without declarative content, and they can also have knowledge in the form of generalized, truth-evaluable declarations which they consult in their operations.  This does seem to have implications for a discussion of whether knowledge how to do something involves states or content which is not truth-evaluable.  Proceduralists believe that all human knowledge is at root procedural, and without propositional content.  Declarativists believe that all human knowledge is at root declarative, involving statements of fact.  If knowing how is purely procedural, it does not rely on machine states, code or instructions with truth-evaluable content.  However, Stanley erroneously concludes on page 208:  "As researchers in artificial intelligence use the distinction, it is irrelevant to the question of whether knowing how to do something is a state with a truth-evaluable content."  This goes against both D&P and Winograd, which exhausts the AI literature Stanley is consulting.

I will address Stanley's discussion of the procedural/declarative distinction as it pertains to psychology in another post.