Watch the video in real media (5.3MB) or Windows Media Format (2.7MB).

(slowly, to tune of "black coffee")
Well, I've got a midnight deadline, and it's already six
I gotta write this paper, gotta send it into NIPS
Got to get some numbers from my experiments
Been sampling for hours and it still don't make no sense

(straight fast 1-4-5 blues)
Well, I implemented backprop, I coded it up
I plugged in all the data, had another cup
I tweaked all the constants, when I had the urge
I had another look and thought my brain had diverged

'Cause it's going 'round in circles, and there's nothing I can do
I'm just looking for a fixed point - I've got the non-convergence blues

They said I should try kernels, said it would be fun
They said it would converge with probability one
I plugged in all the data and still came up short
I couldn't find a region of compact support

And it keeps going 'round in circles...
So I called Dr. Gauss, but despite good intention
He was locally embedded in some other dimension
I tried Reverend Bayes - he said we'd be fine
If I didn't require polynomial time

And I'm just going 'round in circles...

(slowly again)
So if you're in Vancouver, and you go to NIPS
It'll really behoove ya to follow these tips:
If you've got no numbers, the best advice yet
Is to go get yourself another data set
Or you'll be going 'round in circles, with them non-convergence blues