How to do backpropagation in a brain [video] | Hacker News Reader

How to do backpropagation in a brain [video]

youtube.com

27 pointsmodeless11 comments

I'm not quite clear on this: Is this a theory of how backpropagation might be constructed using known neural mechanisms as building blocks, or is this actual observation of how the brain really trains neurons?

Very much a theory, which is inspired by recent successes in training artificial neural networks. It is consistent with our observations so far but certainly not proven by any means.

The title on youtube is simply "Deep Learning - Geoffrey Hinton" and the original url (not the shortened, or time-marked) is:

https://www.youtube.com/watch?v=kxp7eWZa-2M

From the above link, it's a recent talk:

> Published on Jul 29, 2015

> Filmed at the University of Cambridge on June 25th, 2015

I linked specifically to this part of the video because it's really a standalone talk, as he says. "How does the brain do backpropagation?" is the title of this talk (as he says it verbally; the slides have a slightly different paraphrasing, "How to do backpropagation in a brain").

The earlier part of the video is on a different topic; mostly explaining the various successes of deep learning over the past several years. Geoff doesn't start talking until about 6 minutes in, so if you want to see that part of the video start here: https://youtu.be/kxp7eWZa-2M?t=5m55s

The problem is most url shorteners like "youtu.be" are banned on HN due to abuse, so your submission was initially "[dead]" when it was first created. The only reason for putting the right URL and original title in my comment was to help the mods (who have since changed both the url and title).

Ah, thank you for that. I didn't realize site-specific shorteners like youtu.be were banned.

1. we don't know how the brain does complicated thing X

2. we have discovered a way to program a computer to do complicated thing X

3. maybe the brain works like this too

This is faulty reasoning and I see it again and again. This is not how to proceed according to the scientific method

Why not? You could accept your point 3. as a theory, yet to be proven. Your point 2. is simply a necessity to show that the theory makes some kind of logical sense.

No just no.

They've come up with a hypothesis by making observations and guessing. If the hypothesis is also falsifiable, then it is definitely how the scientific method works, since we can now run experiments to determine if the hypothesis is false. The scientific method doesn't say much about how hypotheses are formed, just in how they are evaluated.

Now we can argue about whether this can be falsified or not, and that is legitimate. Bad science would be not taking that step or doing that step badly.

yes exactly. My criticism is only if point 2 is used as evidence of point 3. I see this a lot.

If point 2 is used only as a means to generate a hypothesis, which is then tested empirically, then absolutely this is fine

thanks for the clarification

Keep in mind that artificial neural network are not meant to simulate biological neural networks. They are just inspired by a simplified (and obviously flawed) computational model of the brain from way back in the stone age of neuroscience.

Finding parallels between what ANNs evolved into today and the little we know about the brain may be interesting, but don't think, even for a moment, that you have a model for "backpropagation in the brain".

P.S.: I like the implied criticism of the Human Brain Project at 50:10 - "You need a computational theory (not a billion euros) to recognize what you discovered!". We really need to work on theoretical models of computation before we throw money and hardware at simulations. No emergent behavior is going to jump out and save us that step.

Reply on news.ycombinator.com