dbd%benden@LANL.GOV (Dan Davison) (05/24/89)
David Kristofferson says:
I would think that the BIO-MATRIX approach has some merit but it
strikes me that its case is weakened in the eyes of practicing
biologists every time one brings up this "reasoning from first
principles" example.
And Randy Smith replied:
Dan Davison clearly states (2nd and 4th
paragraphs) that since biologists cannot (now) reason from first principles,
the "Matrix substitutes reasoning by analogy for reasoning from
first principles".
Which is what I meant all along. I will look over the posting and see
if I can reword something to make this clearer. Yes, the bio-matrix
concept is about reasoning, but it's reasoning-by-analogy. This point
was developed at length in the National Academy of Sciences report
"Models for Biomedical Research: A New Perspective (NAS, ISBN
0-309-03538-4).
The Godel example that David Kristofferson invokes is commonly heard
(at least by me) in the wildest contexts. In this case, we know that
theoretic physics calculations break down at singularities (such as
before 10e-43 seconds after the big bang; this is discussed in
Hawking's "A Brief History of Time"). That breakdown is only of minor
consequence for the issue I was trying to address. The point is that
we (biologists) have transfinite calculation problems too (protein
structure, molecular dynamics) in which we can't, in the near term,
know all of the parameters involved in the calculation. In order to
work on that problem *now* we have to do our reasoning at a higher
level.
Jon Sticklen says:
By extension, the suggestion inferred that although "first principles" may
apply to disciplines like physics,that a project to gather
biological knowledge and use it as a basis for arguing from
"first principles" ...
This is the *exact* opposite of what the bio-matrix is all about.
...may be ill-founded.
Right.
Actually, within the AI community, the notion of "reasoning
from first principles" seems to be giving way to "model
based reasoning." Model based reasoning is a new sub-field
of AI which centers on starting with a model of some physical
phenomena, then, by applying reasoning methods (that are currently
the object of intense research) deriving how the model will react
given boundary conditions.
YES! Exactly. That's why this project exists--AIers seem to be coming
up with the relevant tools for the problems biologists facing (whether
they know it or not ;->).
My purpose is not to give a tutorial on model based reasoning here,
but to suggest to the members of the bio-matrix community that there
are other ways of capturing "deep understanding" (to use an AI term)
than to argue from first principles. Current model based qualitative
reasoning may well be applicable in your project.
That's the point I was trying (obviously very badly) to make.
And Peter Karp (hi Peter!) says:
I actually don't think the message that we're all refering to
(written by Dan Davison?) was meant to inspire such serious debate --
but perhaps it should have.
I had kinda hoped it would start such a discussion. I did write that
intro, with large sections taken from Harold Morowitz's section in the
Bio-Matrix Workshop Report. The original note (on the archive-server)
has the full citation.
I for one would not mind seeing a tighter definition of the goals
of The Matrix; such a definition should probably not include words
like "evangelize."
Mea culpa. Although the note was not intended as a "tighter
definition". Let's build one. See below.
And Jon Sticklen replies:
When we try to reduce one level of understanding
to another, we lose the descriptive power that the terms at the higher
level afforded, and are forced to use the vocabulary at the lower level.
Eg, mating behavior of large mammals may be reduced through a number
of steps (in principle perhaps) to the bio-molecular level and beyond.
But if we do that, then we cannot use terms like "drive to find a mate"
because such terms do not exist at the bio-molecular level. One of the
most pressing needs when trying to represent a phenomena is to search
for the appropriate level of description to capture understanding.
Going further, one of the most pressing needs of the bio-matrix project
(it seems to me) is to find a representation scheme that is robust
enough to support descriptions of different phenomena at many different
level of abstraction, and that allows each description to utilize
terms appropriate to the phenomena being described.
This is an excellent, concise summary of the problem and the way the
bio-matrix project should approach its resolution. I recall that
Peter Karp, Chris Overton, Kimberle Koile, and Rick Lathrop, among
others, had constant discussions about exactly this point at the Santa
Fe workshop. It was never clear to me how much of their work was
influenced (perhaps limited) by their knowledge representation tools
(I can't recall the names now, nuts).
My statement of the goals has been examined and found lacking. OK,
let's try to build something better. Perhaps some of Chris Overton's
introductions to the bio-matrix (see the 1989 summer meeting
announcement) would be more appropriate?
dan davisonovert@PRC.UNISYS.COM (05/24/89)
I wasn't going to respond to any of this now, although I find it very interesting, because I am working on a grant proposal due June 1, but after reading Dan's message I feel like putting my 2 cents worth in. I take "reasoning from first principles" to mean the capability of making strong predictions about the behavior of a system based on the knowledge of a set of rules which embody a theory about the system. In the physical sciences, these rules usually are stated concisely in the form of mathematical equations. With few exceptions, no set of simple rules and no equally powerful formal language is available to describe or support reasoning in the biological sciences. One strong rule in biology is the central dogma, but then it, like everything else in biology, has an exception. [In fact, the only good rule in biology is that there aren't any unbreakable rules.] In general, biological disciplines have only weak theories and little predictive power as compared to the physical sciences. This is why "reasoning from first principles" has not proven especially successful in biology. But I keep hoping that some day, just maybe, we can have a reasonable set of rules to describe biological system. The biological sciences encompass disciplines from the molecular level to the eco-systemic level. I prefer to think of each level in the hierarchy of organization as having its own rules and formal language, and that the lower levels in the hierarchy impose constraints on the higher levels. Of course, interactions take place across widely separate levels, but that too can be taken into account. Then each level would have its own first principles from which to reason, rather than as Jon Sticklen suggests, the lowest level (whatever that is) having the "first principles" which manifest themselves at the higher levels. It's certainly not the case that there aren't first principles to be found in biology (sorry for the double negative), it's just that their character and their complexity will be different from those in the physical sciences. And it will require a different approach to uncovering them then that used in the physical sciences. [In part, this is the justification for the bioMatrix project.] This is because biological systems are extraordinarily complex self-regulating systems and because the existing examples are historical accidents. So what's this got to do with reasoning from first principles? Well, the fact that the systems are complex implies that the number of rules needed to describe them is going to be large. We're not going to see something like e=mc^2 as a description of vertebrate development. Second, while all biological systems are related through evolution, the fact that the systems are historical accidents means that if you tell me exactly how a drosophila works, I've still got a long way to go before I understand how a mouse works. I like to think about this by analogy to natural languages---another group of complex, historical systems. For example, given a complete formal description of Chinese, what predictions could you make from first principles about the syntax and semantics of English if all you have to work with are as many examples of spoken English sentences as you care to hear? Not many. On the other hand, when parallels can be identified between mouse and fly systems, for example by functional or sequence similarity, then strong inferences can be made. This forms the basis for reasoning by analogy in biology. Peter Karp and Jon Sticklen would like a clearer statement of the goals of the bioMatrix. Peter has been involved in this as long as I have so he should be able to articulate his own perspective on the goals of the bioMatrix. Let me say though that there seem to be as many views of what the bioMatrix is as there are people involved in it. Biologists in general perceive it as a grand tool to give them better database access. Computer scientists see it as an opportunity to test out whatever system they are building in an interesting domain. My personal view is that the point of the bio-matrix is to try to aid in the development of appropriate theories of biology at each level of the hierarchy of organization. The general approach advocated is an information intensive one which requires multiple components to make it work. The components include a means to access the ever growing body of information in biological databases, a means to properly organize and represent the ground data as well as higher level abstract concepts (ie laws and generalizations of biology), and, of major importance, methodologies for reasoning about the information. Reasoning by analogy is one especially powerful method in biology, but other potentially important reasoning methods, such as qualitative reasoning, may also prove important.
mike@TOME.MEDIA.MIT.EDU (Michael Hawley) (05/24/89)
Brrr. I hope you people know, we're in the middle of the AI winter (which puts me at ground zero). I don't want to snow on the parade, but this fallout about first principles is making me shiver. I have always been convinced that the most bangs for the buck, by FAR, will come when you computerize a hundred gigabytes of biological information and make it searchable. Formalizing the AI of biology -- and theorizing about reasoning on biological principles -- well, I wouldn't want to point fingers or anything, but at this stage I would much rather work with a few dozen system experts than a few dozen expert systems. More fun, too. So what will it take to siphon mainstream biological literature into useful computer form? Perhaps there is some way I can help. I think I may have been the first person ever to "grep" the Origin of Species, and I would sure like to catalyze the computerization of biological literature and journals somehow. I would like to do nothing better than twist Steve Jobs' arm to this end. Mike Hawley