In relation to this article, it says that when you build a model that fits the data, it's not enough. It's a further requirement that the model makes testable predications that turn out to be correct.
Models that make accurate predications are valuable. Models that aren't then tested by checking their predictions are, by definition, untested. That, to my mind, makes them potentially dangerous.
I'd be interested to know why you describe making falsifiable models is just "a learning algorithm."
Thanks.
the problem with the article is that it lumps in many 'models' with wholly untested ones undeservedly, saying And no, that does not mean simply training your model on half of your data set and showing that you can effectively explain the other half of your data. That is hypocritical because this is more or less precisely what scientists, following the "scientific method", are doing, albeit many times slower than a computer with a parametric model.
> I'd be interested to know why you describe making falsifiable models is just "a learning algorithm."
what happens in some piece of the world over some duration of time can be represented as a function from that piece of the world's initial state to its final state. in the sciences, we try to estimate such functions. the scientific method is a very simple algorithm for doing so: generate a hypothesis, compute N "testable predictions" f(x_i)\approx y_i, and then accept your hypothesis if they all turn out to be correct.
When you build a model and then do science (i.e. Hypothesis/Test/Explain loop) on the model instead of doing science on Nature, you learn a lot about your model. But you must be honest and admit that you have learned nothing about Nature. What you might have done is come up with hypotheses. But it's not enough. It's not science yet - it's just a complicated hypothesis. It doesn't count until you test it on Nature.
The point (and it's an obvious one, well-understood by most complexity researchers) is that your models need to enable you to falsify them by doing real, repeatable experiments on Nature, not on models. There's no point building a model of a complex system if the only experiment that can falsify it is so complex that it can't be done.
A lot of time is wasted this way - in many ways it's the nature of the beast (it's called "complexity" for a reason). Nonetheless, this is frustrating to scientists who think the resources spent on learning facts about models might be better spent on learning facts about Nature - the more radical likely being willing to trade all the modelling (and modelers?) for even a single properly verified fact.
> this is frustrating to scientists who think the resources spent on learning facts about models might be better spent on learning facts about Nature - the more radical likely being willing to trade all the modelling (and modelers?) for even a single properly verified fact.
there is no getting rid of models. neither science nor any other human activity i am aware of is capable of verifying facts about the physical world.
it would be essentially the same thing if you took half the data, made a model, tested it on the other half, got good results, and stopped.
but thats not going to happen. your fist model will stink. you'll refine it and try again. and you'll keep doing that. unfotunatly, this mean your model ends up being dependent on all the data.
the easiest way to prevent this is to build a model and the test it on real world data that occurs after you build the model. if it doesn't work, tweak the model, and then find new data again.
this is exactly what the scientific community as a whole is always doing.
That doesn't mean that because you didn't manually tune each and every weight in the classifier that its behaviour is not solidly governed by scientific principles.
Science and the application of the scientific method are what got us there in the first place, and has given us the keys to understanding these systems. Genetic Algorithms are a another area where you could easily be tempted to think that you are not doing science, the same thing happens there.
Science is the key to understanding. The scientific method is the ultimate arbiter between what we know to be true, and what we can not prove to be either false, fantasy or outright lie.
I feel that you think that somehow these computational models somehow supplant the scientific method, but you are forgetting one crucial bit here, those models themselves will need to be understood, and only the scientific method will allow you to do so.
It's different for the simple reason that to really verify a hypothesis, you don't just test it on the other half of the same data. That's just step 1.
What you need to do is work out the further implications of your hypothesis, and test those. When doing so, one of three things will happen:
1) You remain an astronaut. In other words, your work is so far out that none of its implications falsify or corroborate any existing theories, and all require their own experiments. This can be good (i.e. ground-breaking work), but for your hypothesis to become accepted as anything other than an interesting diversion, you will need to wait for the rest of science to catch up, or to keep building out implications until you can connect what you're doing to existing theories.
2) The implications of your hypothesis directly corroborate an existing theory. This is useful, but a bit boring. It means you've basically discovered yet another implication of an existing theory. Experiment some more, and then publish.
3) The implications of your hypothesis contradict existing hypotheses or theories. This is exciting! This is where the real science lies. Now you get to design a series of experiments to figure out who is right, who is wrong, and why. You debug your hypothesis and/or the existing science 'til it works, possibly overturning established wisdom along the way.
4) A combination of (2) & (3) occur. Woah! This is really exciting. You've discovered that the implications of one existing theory falsify those of another existing theory, uncovering a fundamental flaw in our understanding so far. Well done! You've got a lifetime of debugging ahead of you, but your contribution is likely to be truly important.
> there is no getting rid of models.
True; and models are very useful. But there's a big difference between models used as an explanatory / predictive tool (i.e. Newtonian physics), and models used as a substitute for Nature when doing experiments (i.e. what's done in Computational Neuroscience, my field). The latter attempts to duck the issue of doing real experiments that are complicated / expensive by running the experiment on a computer.
Now, this is certainly useful - but to be science, it must be accompanied with a solemn understanding that "What you learn here don't mean jack." It's just a model, and probably a bad one. God doesn't use 64-bit floating points. The only possible use is to help you design a better experiment to try in the real world. But if you never get to doing that, then you've wasted everybody's time, because you got stuck on step 0 (Hypothesize).
> There is something fundamental about the scientific method. Namely, it's the learning algorithm; the only one that actually actually works (so far). Everything else is just variations on or approximations of this algorithm.
i believe that what i call the scientific method is merely a variation on/approximation of bayesian inference.
Could you enumerate those ?
The are very interesting though, that's for sure.
in this case, the development of inference is certainly a consequence of the development of the scientific method. but the validity of the latter is a mathematical consequence of the validity of the former. bayesian inference is more fundamental.
another good book that informs my personal views on this matter: http://www.amazon.com/Probability-Theory-Logic-Science-Vol/d...
1. Define the question
2. Gather information and resources (observe)
3. Form hypothesis
4. Perform experiment and collect data
5. Analyze data
6. Interpret data and draw conclusions that serve as a starting point for new hypothesis
7. Publish results
8. Retest (frequently done by other scientists)
The idea of an absolute sacrosanct Scientific Method is contrary to what scientific research is about. It defines science to be an academic method.
So, assume that I have my personal system of doing research. I don't define the problem. I don't form formal hypotheses; I don't perform experiments (formal academic exercises) instead I use trial and error; I take frequent stratetegic naps to develop my ideas; I analyze the idea any way I like; I don't publish my results; instead I note my results meticulously on my lab books. I don't care anything about academic rules and ignore academics and their science fetish.... instead I make discoveries. I ignore the paperwork, and career considerations. . . I make discoveries. You recognize Edison's method. Are you saying that what I am doing is not science because I did not follow your step by step "scientific method?"
The scientific method is an attempt to brand academic physics as science and the rest as stamp collecting. This marketing gimmick actually has been working. People associate academic physics with science.