I propose we replace UTC with Taylor Time (TT). TT+0 is obviously the same as Eastern Time, as that's where Midnights released at midnight.
Seems like a much more relevant way to determine the canonical time zone than centering it on Greenwich
I almost wish reservoir computing had won out over backprop just so that I could include this image of "The Liquid Brain" from Pattern Recognition in a Bucket at the start of all my deep learning talks
ALT Image of a water tank with k'nex motors attached captioned "Fig. 1. The Liquid Brain.
Stats brain: the std of an exponential distribution is the same as its mean
Viz brain: cool, lets take a look at the error bars on this scatter plot
Viz brain: ...
Stats brain: like I said, the std of an exponential distribution is the same as its mean
ALT A scatter plot with extremely wide error bars
Don't you hate it when you're chilling in Dreambooth and suddenly your dog starts melting?
ALT AI-generated image of a man and a dog with distorted faces
As it happens, you can use Limma to remove predictive linear signal instead of batch effects. When we did so, we found that logistic regression's performance becomes random (as expected), but the neural nets are still able to learn things. 7/
ALT Nonlinear models are able to learn from the remaining nonlinear signal, while logistic regression is not (Recount3)
ALT Nonlinear models are able to learn from the remaining nonlinear signal, while logistic regression is not (GTEx)
You're probably thinking, "Well you just said the model behavior could be unstable, so what?" But these results held across datasets and problem settings. 5/
ALT Similar performance of various models in predicting tissue data in Recount3
ALT Similar performance of various models in predicting metadata sex labels in Recount3
