Top-down vs bottom-up approaches in science?

Question

I am a student of a natural science but very interested in philosophy. During my studies, I have noted a perceived difference in how various disciplines approach the explanation of data they obtain. I would roughly separate this in two classes:

top-down: In some disciplines, scientists seem to first try to recognize patterns in the data, then try to find mechanisms which generate similar patterns. I would label this kind of approach a top-down approach, since the identification of structures in the data precedes attempts to explain its origin. Examples are biology and chemistry, or public opinion polling, where various kinds of regression are often the first data analysis. Edit: I expect most black-box machine learning exercises like convolutional neural networks also fall into this category.

bottom-up: In other disciplines, scientists seem to start by shifting well-known building blocks around until their predictions fit the observed data. I would label this a bottom-up approach, since the recreation of the data-generating process often precedes (or scarcely profits from) attempts to understand patterns in the data. Examples are (I would say) hydrogeology, meteorology, or criminology.

Thinking some more, the difference may be the complexity of the systems under investigation. For disciplines in the top-down category, the processes under investigation can often be studied in isolation, for disciplines in the bottom-up category this separation is generally not possible.

Would you agree with such a distinction? Do you know of any work which has explored similar ideas?

Answer 1

Science is presented as this orderly thing with inescapable conclusions. But it is certainly not done that way. That is only the way the story is told to impress people the scientists are trying to chat up at the pub.

Scientists will look at whatever part of the problem they can and play with it. They will jiggle the abstract theory parts around. They will filter the data. They will say "what if this experiment is flawed this way?" They will hold up an idea that seems weird just to see what other ideas it shakes loose. They will take a theory and predict what the data would be in an experiment then go look for an experiment that gives that kind of data. (Not all of these methods are considered valid.)

They will have emotionally charged arguments with their colleagues, occasionally with unfortunate side effects and outcomes. The story is (though I can't find any newspaper report of it) that one prof hit another over the head with a full pot of coffee, producing terrible injuries as you might imagine. The reason was some disagreement over the correct way to do a calculation in an empirical study of multi-quark systems to predict the behavior of an experiment at CERN.

What I'm trying to tell you is, the process of science, the way scientists do things, is often incredibly messy.

But when the narrative is finally told, it is told all-the-way-down. Meaning, the abstract theory part is made as consistent as possible internally. The data is vetted and examined and reproduced as carefully and critically as possible. And the two are connected as clearly as possible.

AND THEN the process starts all over again. Somebody suggests a new experiment. Or re-examines an old experiment. Or does the calculation differently and gets a new answer. Or they tweak this or mod that or add the other. Again, chewing on whatever part of the problem they can get their own mind around, or that is most interesting to them.

You see, science knows it does not know everything; otherwise it'd stop.

Answer 2

As in everything in existence, there is a time and place for each.

Top-down approach is used when a phenomena is well understood (or is considered well understood as in theory we make for it well fit the data).

Bottom-up approach is used when we want to have that theory. Bottom-up approach is harder than the top-down approach.

In top-down approach you are made familiar with the players from the get-go. Rules are presented to you as already known and firm. You have a very solid base before you go further and investigate something, and that something is a small thing, not much important to the whole tree, some branching at the 5th or 10th level say.

Bottom-up approach is needed (there is no way around it) when you dont even know all the major players, yet alone fundamental rules of the game. You start with a small set, whatever the data is in front of you, and make a theory about it. Then you go up and you dont necessarily know you are going up. You modify your theory for this additional data. So on you go.

You have to keep your mind open. There are no fundamental rules that you know till you are sufficiently up in the tree. You dont know you are sufficiently up in the tree till you reach the root or a root.

Another important distinction between the two approaches is, top-down study is taught to you. Its your secondary research. As in, somebody else already did the primary research (observe, experiment the natural data) and you are just reading his book or listening to his lecture. If you dont have that facility then you are bound to do bottom-up analysis.

Research is hard, especially when its primary.

Answer 3

The distinction between 'Principle' theory and 'Constructive' theory is one that Einstein made: https://plato.stanford.edu/entries/einstein-philscience/

The central question is whether we can ultimately construct all knowledge on known axioms, or if knowledge ultimately relies on Principle theory (i.e. assumption of the phenomenon of consciousness).

I think that this might also be a concept that the Russell's Paradox / Gödel's incompleteness theorem touches on (within the field of mathematics).

Would be interested to hear whether this was something you pursued further?