Lin Wood @LLinWood is bringing the

This is my son, Sam. He’s 5. And today he learned how shitty and harmful #ToxicMasculinity is. My rage meter is spiking right now so excuse me if this is a little raw but there are some things I want to say about BS #gender norms (a thread)

Sam is my middle child & he’s a terror. A “boy’s boy” as so many (not me) would say. He’s rough and tumble, he’s loud, he’s always dirty, loves trucks, plays sports and knee drops me from the couch. But he also loves a lot of “girl” things.

Sam has a collection of purses because he likes to carry things around. And he also loves to have his nails painted bright colors because he thinks they “look beautiful.” And he’s right – they are beautiful…

So he proudly wore his red nail polish to kindergarten this morning because Sam has absolutely no concept of nail polish only being for girls or reason to think anyone would possibly have a problem with beautiful nails.

But his classmates did have a problem. A big one. Sam was ridiculed for being a boy with nail polish. They called him names and told him to take it off. This lasted the entire day.

Important paper from Google on large batch optimization. They do impressively careful experiments measuring # iterations needed to achieve target validation error at various batch sizes. The main "surprise" is the lack of surprises. [thread]

https://t.co/7QIx5CFdfJ

The paper is a good example of lots of elements of good experimental design. They validate their metric by showing lots of variants give consistent results. They tune hyperparamters separately for each condition, check that optimum isn't at the endpoints, and measure sensitivity.

They have separate experiments where the hold fixed # iterations and # epochs, which (as they explain) measure very different things. They avoid confounds, such as batch norm's artificial dependence between batch size and regularization strength.

When the experiments are done carefully enough, the results are remarkably consistent between different datasets and architectures. Qualitatively, MNIST behaves just like ImageNet.

Importantly, they don't find any evidence for a "sharp/flat optima" effect whereby better optimization leads to worse final results. They have a good discussion of experimental artifacts/confounds in past papers where such effects were reported.

In regards to the #CapitolCoup, I’m in a unique position. I monitored forums ahead of time and then was present for the coup itself.

These are my observations based on what I saw and the evidence that’s been collected so far. I reserve the right to adjust my analysis as needed.

Let’s talk about the organization of the coup first. Seeing some people insist that the coup must have been highly planned and coordinated.

Yes and no. This was a #StochasticCoup. We’ll see little evidence of direct communication between Trump and the insurrectionists.

Within the insurrectionists, there was not one group that executed the storming of the Capitol.

Rather, there were a number of smaller groups working independently towards the same goal. I expect you’ll see little evidence that these various groups communicated with each other.

What you will see is detailed planning and coordination within these small groups.

Then there were the joyriders, those Trump supporters who weren’t directly involved in any plotting but who were more than happy to go along for the ride once the coup was in motion.

After four years, the insurrectionists were carefully attuned to Trump’s messaging.

Every tweet and statement of his was treated as marching orders and carefully dissected for meaning.

It’s why shutting down Trump’s Twitter account was so important.