Slashdot Mirror
When Scientific Publishing was Withheld
karvind writes "Article in Physical Review Focus reveals the silence practiced by Physical Review during WWII to delay publishing results related to fission, the splitting of an atom's nucleus accompanied by a prodigious release of energy. From the article: Because of fears that Germany would use American research to pursue an atomic weapon, the Physical Review agreed to withhold reports of significant advances. It was not until several months after an atomic bomb exploded over Nagasaki, Japan, that Phys. Rev. published the paper announcing the discovery of plutonium, the material used in that bomb. Physicist Abraham Pais later called the journal's silence on the subject 'the most important nonevent in the history of the Physical Review.'"
Firstuvall, I'd like to applaud the uncommon scientific focus of this; topics related to science in general are gee-whiz news of space exploration, not about science in its making. I would guess many slashdotters are scientists, and this brings good rest from the "SCO says they own Mickey Mouse and the patents to condoms" days.
That said, peer-reviewed outfits are still ran by humans. Neural nets have been essentially blocked by the nonparametric statistics community for a long while -- leading to the bizarre situation of having electrical engineers understand a lot about time-series prediction that the people who are actually involved with it don't -- and is only now making advance as econometricians -- who typically develop parametric statistical methods and then try to fit everything to their methods -- are adopting it, partly because of sheer job-market pressure.
And all that is in a pretty technical, numbers-in-numbers-out field.
So you pick up a peer-reviewed rag in economics -- and if economics isn't science, medicine isn't either --, and it risks having at least three types of ideological bias: a political one (generally from the more-or-less-state-intervention kind), a established-scientific-practices one (people already know their field, and getting game-theorists to accept category theory and arrow-chasing proofs is proving hard) _and_ a schools-competition one (possibly linked to political issues, since hyping up schools linked to free-market stances will harm the more-intervention camp).
Yes, you could say that physics has less politics involved. But when you're dealing with the very nature of "actual stuff", you are bumping into very deep philosophical stances that may be much harder to shake than political convictions with the scientific process only. I know many people who have come to adopt a more-free-market POV after being exposed to general equilibrium and microeconomic theory, but it's harder to convince people -- Einstein wouldn't -- that the universe is ultimately stochastic, or that our behaviour might be evolutionarily stable and a product of our genes, etc. etc.
In the end, economics has nothing like the controversy on sociobiology. Outside radical circles who have been essentially ignoring economic theory since uncertainty and assymetric information have come into play in the models, there is a very deep consensus among economists at least in the basic issues -- from Paul Krugman to Arthur Laffer.
Politics is just politics. We have our own interests, and we act to defend them. And after a while, people start to analyze what people do in the defense of their interests, and the action of special-interest groups, rent-seeking behaviour, etc. becomes clear.
Personal philosophies are a lot muddier. And physics touches the bottom of them.
For people who like this subject matter and want to read more about the history of the development of atomic bombs (including the history of early 20th century atomic physics), I can recommend The Making of the Atomic Bomb by Richard Rhodes. Solid history _and_ good writing.
I bought it after it was recommended in some other Slashdot post, and loved it.
I believe posters are recognized by their sig. So I made one.
"A Soviet scientist deduced from the Americans' silence on the topic that they were pursuing an atomic bomb. The Soviets soon followed suit."
Amateur paranoiacs cannot hope to compete with professional ones.
"Wow. Now THAT'S a lot of angry Indians." - Lt. Col. George Armstrong Custer
I agree that science is more policitized than many people think. But let me just add my two cents worth regarding pseudo-"black box" methods.
One of the reasons Neural Networks were viewed with some doubt was because of their "pseudo-black box" nature. Train it enough and you will get a model that gives you a good fit for your data, but you have no insight as to interpret the results, not least because you will almost never get the same model twice from the same data (the weights will be different every time you train them).
The neural networks idea sounded interesting because of the "cool" biological analogue it has with neurons firing in your brain (and it had interesting jargon to boot).
But if you look at its mathematical description it boils down to doing a simple regression/curve fitting with a limited nonlinear model that uses exponential functions (known in the NN community as "activation functions") like the sigmoid etc. (You can actually derive this if you write out the equations for a simple 1-2 layer neural network).
It spits out data that fits the curve, but tells you nothing about the correlations inside them. In the 1980s, people were attracted to it because of its simplicity and the fact that it seemed to be feasible way of mimicking a human's pattern matching abilities. It was all the rage back then. In the 1990s or so, people started to become aware of its weaknesses and began to look at it more circumspectly.
To give you an example, most credit card companies use Neural Networks to approve credit card applications. They pump your application data through a trained model (based on past classifications done by humans), and it spits out an "Approved" or "Not Approved" flag.
Unfortunately, you have no idea why a certain application is approved or not approved. A neural network model can't tell you that. It's only designed to give you an answer based on the its training weights, i.e. it only models the relationship between Y and X, and not the Y and X spaces themselves.
Instead, if you apply a multivariate statistical method such as PLS (via a NIPALS algorithm), the model will tell you how things are correlated (in a easy to interpret graphical fashion). It will pretty much be doing the same thing as the neural network, except that it models the X and Y spaces simultaneously, compensates for missing data by deriving from the correlation structures; all this by transforming the variables into a latent variable space that captures the maximum covariance in the data. All the equations are transparent and have a solid basis in the mathematics of linear transformations and projections.
And you get the same model each time, so it can tell you exactly why your credit card application was turned down. (Too many unpaid bills, for instance)
It is easy to become enamored of black-box methods (I know I was), but ultimately the methods that survive are the ones built on rigorous mathematical/scientific foundations. (not always possible, especially in areas like economics, but it is something to strive for)
Most ideas and theories get superseded over time, but black-box methods and theories produce the most controversies. Sometimes you can't blame the community for being a little skeptical of them.
According to a theorem usually attributed to Cybenko, any continuous nonlinear function can be represented by a linear combination of sigmoid functions of a linear combination of your parameters. In neural nets terms, a single hidden layer net with 2n+1 neurons in your hidden layer can represent _any_ continuous function.
That doesn't mean the usual neural net training algorithms are able to achieve that representation, but it's still a strong result, and it mostly justifies neural nets being increasingly looked at seriously at nonparametric (without individual input effect parameters as an usual OLS model would yield) statistics.
All in all, I do have a lot of faith in the future of nonparametric methods. They might be no substitute of empirical experiment (and that's what the parametric statistical methods that comprise econometrics strive for), but the sheer success of neural nets in spite of their lukewarm academic reception shows they can be quite useful.
Lets leave this aside, but economics is in my opinion no real science, because it uses the wrong methods and aims for the wrong goals. Economics comes all down to psychology (I would call economics the psychology of greed) but yet they try to enforce the rules after the whole system is tried to be described after mathematics. Modern mathematics failed basically in one area to be applied usefully and that is psychology. All the models which are applied to economics usually only work out in the best case because the description only cann fill one angle instead of proximity patterns. And in the end there usually comes the case where the system of formulas applied to being greedy crumbles to dust. Modern economics is the closest thing to alchemy we have nowadays.
Yes, much of the silence in scientific circles is political or, as in the case of wartime, to protect national interests.
Private interests get in the game too though. Drug studies are squelched if the sponsors don't like the results, and people wind up dying. Decades-old industry studies of smoking tobacco never saw the light of day until recently. You always hear rumors about car and energy companies not telling all they know about more efficient motors and fuels, but "those are just rumors."
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
There are two problems with this argument:
There's no point in questioning authority if you aren't going to listen to the answers.
Not really.
Plutonium was being produced in the Oklo natural nuclear reactor that was running in Gabon 2 billion years ago. It had decayed away by the time we showed up on the scene. See this, for example.
We learned of it by making it, but nature had done it long before us.
Cybenko's result was more of a reality check for the NN community than anything else. If NNs didn't have this property, there wouldn't be much use in studying them. The Weierstrass Approximation theorem, which you can find in a good real analysis book, shows that plain old vanilla polynomials of the form sum(i=1,n) a_i x^i have the same property.
Barron had a paper giving rate-of-approximation results for a certain class of functions. This starts to answer the question "how big should n be?" I'm not sure what new work has appeared along these lines. I've been out of touch with the NN community for a few years. That said, a lot of the learning people seem to be more excited about support vector machines and kernel methods these days. I guess some people group these techniques along with neural networks, but they lie on a much more solid foundation of statistics.
cheers, Rick