The details are presented at the end of page 2. You can see on page 3 that the whole thing is a glorified “I can make neurons go to zero!” machine.

It kills me because there is real value in this work — the ability to modify neurons is a useful thing to be able to do. Showing that you can set them to specific values is worthwhile.

But that wasn’t what it was presented as. The focus was on the idea that the neurons somehow learned something. Maybe they did, and maybe time will prove me a fool by showing that this is how neurons learn in general. But it was so disheartening to spend hours carefully going over every detail, full of excitement at the possibility of machines that can learn… only to realize I could do the same thing by setting a value to zero in python, and that the complicated language seemed designed to conceal this.

I remember seeing a video on the discovery channel where they were interviewing him. He gave a dazzling account of the implications of this work.

Maybe he was predicting the upcoming ML boom. But now that I’ve worked in ML for a few years, I can safely say that this work didn’t help us get here. Maybe it helped biologists figure out how to control the signals that neurons deliver.

(I still think that it’s pretty darn cool that you can manipulate neurons at all, so hats off to whoever figured out how to do that. I just wish it was presented as what it was: the ability to set a value to zero over time via neurons.)

I really want to make a joke like “You know what else would set neurons to zero? Set them on fire and wait,” but the paper showed they could be set to arbitrary values. Which of course was the interesting and valuable part in the first place, not the plane stuff. But planes make good headlines.