1/ A ∞-wide NN of *any architecture* is a Gaussian process (GP) at init. The NN in fact evolves linearly in function space under SGD, so is a GP at *any time* during training. https://t.co/v1b6kndqCk With Tensor Programs, we can calculate this time-evolving GP w/o trainin any NN

Tired of word clouds? Want to do better sentiment analysis? Not sure how to look at the words underneath your measures?

Our long overdue paper on generalized word shift graphs is finally here!
https://t.co/lIBXvbMJWX
https://t.co/vSL1REYT8V

So what are they?

1/n

If we have two texts, there are many ways we can compare them. Weighted averages are a particularly useful measure because they're flexible and interpretable

Proportions, Shannon entropy, the KLD, the JSD, and dictionary methods can all be written as weighted averages

2/n

But weighted avgs are also slippery. When we try to compress complex phenomena like happiness, surprise, divergence, or diversity into a single number, it can be unclear what we're measuring

If the measure goes up, what does that mean? Why did it do that? Can we trust it?

3/n

Very often, that's the end of the line and we're left with an uneasy feeling in the pit of our stomach that our weighted avg is actually picking up a data artifact or some other unintended peculiarity

Word shift graphs help us address those concerns

4/n

First, word shifts look under the hood of weighted averages to see what's going on

All weighted averages are a sum of contributions from individual words. We can pull out those words, and rank which ones contribute the most to the difference between two texts

5/n

Cheat sheet that summarizes #DataScience in 10 pages
(Links in the comments below 👇)

2/ Link to the cheatsheet by Maverick

On Bayesianism, the Many Worlds Interpretation, and personal identity.

Some thoughts worked out in a letter to a friend, which is the kind of thing you do when off Twitter for a glorious week. (🧵)

“Chance is ignorance”—the Bayesian story; all probabilities represent states of mind, not states of the world. One *could* put (some) chances “in the world”, but let’s take Occam’s Razor seriously...

That the probability of a fair coin coming up heads is 50% simply means that marginalizing (tracing, as the physicists say) over the hidden facts leaves you, nearly, maximally ignorant of the outcome.

Quantum uncertainty (access below!) poses an apparent challenge to this story. There seems to be nothing to be ignorant about when it comes to (say) electron spin—there is nothing “inside” the

The electron is a simple object, in other words. So where does the uncertainty come from? One could follow David Wallace’s wonderful interpretation in terms of chaotic dynamics and decoherence, but let’s see if we can take another route...

An amazing new project from @bearpelican was just released: https://t.co/DBov6sZTVS . A beautiful design; you can auto-generate a melody from chords, chords from a melody, and more.

It's technically brilliant, combining BERT, seq2seq, and Transformer XL
https://t.co/jF3mO5aXiu

It's also a wonderful example of leveraging and customizing the fastai framework in a deep & thoughtful way.

Here's the full set of blog posts diving in to this