You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
I hated “idea papers”. High level ideas without experimental evidence were not real science to me. Then I realized that many empirical results are post-rationalized randomness and overffiting. They are not evidence at all. Now I respect idea papers. They are real science.
Someone is putting an idea out there to be tested in a scientific way. They are willing to be proven wrong. They are taking a risk. Someone fitting their theory to their empirical result after the fact takes no risk. Nobody can prove the theory did not come first.
The former one tries to find causality, which is what we are after. The latter one proves nothing, just correlation.
Fitting theory to empirical results is the very essence of inductive reasoning, and hence central to science. I feel you're setting up a straw man for empirical research.
Theory and experiments form a virtuous circle. Each can be done incorrectly and criticized, but they are both necessary. When done properly, they work together to advance our knowledge. One is not more "real" science than another.
Inductive reasoning based on very few samples can also lead people astray. It’s difficult to not jump to conclusions after having observed some result. Especially if your incentive is to get a paper published.
I’m not at all arguing against empirical results or induction. A lot of amazing results have come out of it. But perhaps they need to be coupled with the right incentives to work better towards scientific progress.
My original point was that papers without empirical results are *also* science (which I didn’t believe previously), not that they are necessarily more science or better science than empirical ones.
I think BatchNorm is a good example. It was mostly an empirical results with huge success. But researchers are still finding new explanations for why exactly it works.
But that begs the question: Had we presented BatchNorm as “look, this thing works but we don’t really know why” instead of immediately proposing it decreases Internal Covariate Shift, perhaps we would have gotten to a better understanding faster.
Since it’s impossible to get a paper published with “look this works amazingly well but we don’t really know why” we probably have a lot of instances of induction based on a single observation gone wrong
Isn't the reality somewhere in between. Isn't laying hypothesis for testing, empiricism 
.
In search of a filament for his electric lamp, Edison tested 6000 specimens of Bamboo. 3 of them worked! Before that he had tried thousands of other materials.
Let us not write off empiricism for problems associated with behavioural biases. Get rid of the biases, not empiricism.
Skin in the game @nntaleb
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
