AlphaZero vs. the human brain

[smatovic]

I think it is interesting to point out that Deepminds AlphaZero used only about
50 million weights to play chess on a super-human level.

Contrary the human brain consists of about 100 billion neurons, each neuron has
about 1000 connections to other neurons, so 100 trillion 'weights' in total.

Of course, most of these are not used for chess but for other funny things.

But I wish to point out that human neurons are clocked with up to 100 Hertz
only, resulting in an max of maybe 100 evaluations per second, 100 NPS so to
speak, AlphaZero with 50 million weights made on 4 TPUs 80 KNPS.

Maybe these numbers show us the way to go, less nps but bigger, really big,
neural networks?

--
Srdja

[towforce]

Maybe these numbers show us the way to go, less nps but bigger, really big, neural networks?

Bigger isn't always better (link).

[smatovic]

Maybe these numbers show us the way to go, less nps but bigger, really big, neural networks?

Bigger isn't always better (link).

I have no present numbers, but iirc, the policy head of LC0 alone on a depth 1
search made already about 2400 Elo.

And the smaller NNs sizes of LC0 for CPUs are easily outperformed by bigger nets
on GPU.

If your claim is that we can or have to extract somehow the magic of NNs into
some formula written that humans are able to read, I agree, and guess that will
be the next step after NNs, to un-blackbox NNs.

--
Srdja

[towforce]

If your claim is that we can or have to extract somehow the magic of NNs into some formula written that humans are able to read, I agree, and guess that will be the next step after NNs, to un-blackbox NNs.

IMO, it would be better to create a set of linear relationships between chess pieces, and use linear programming, which is extremely good at finding optimal weights in systems of linear expressions, to find a set of weightings that matches a sample set of positions with accurate evaluations. In the other thread (link), I have also offered an idea for simplifying the resulting set of expressions and weights after it has been created.

My point about NN size is that the animal kingdom has examples of complex behaviour encoded in small NNs (I gave the example of the housefly, whose brain only has around 100,000 neurons). Evolution over a long period of time has come up with excellent solutions, and I have no doubt that by using some form of optimisation, the results that a housefly's brain obtains could be achieved with even fewer neurons. A smaller NN would probably work more quickly (this, again, is the evidence of the animal kingdom: human brains are very slow at specific tasks compared to other animals, and are sub-optimal in other ways as well - see other thread for examples).

[smatovic]

As a remark, my est. 100 NPS for the huamn brain are just an upper bound, in
practice it may be one or two orders of magnitude slower, cos some neurons are
clocked with only 10 Hertz, and the network is layered...

--
Srdja

[shrapnel]

Maybe these numbers show us the way to go, less nps but bigger, really big,
neural networks?

--
Srdja

Quite Correct. My practical Experience with lc0 on InfinityChess and other Sites confirms your theory.
Using the 128x10 Networks I got beaten practically every time by the AB Engine Users, even though the Knps was very high. Slightly better experience with 320x24 Networks, but not by much.
Using the Large Nets provided by jhortos and Sergio (384x30 !),while getting only 18-20 Kn/s on my ASUS GeForce 2080 Ti, I very rarely lose.
So Yes, Large Networks are the way to go, with the caveat that the GPU should be a Top-end one to get positive results, otherwise it will only backfire, for example if you are using only a 1060 Ti.
With new and faster nVidia GPUs coming out Q1, 2020, things can only get better with lc0.

[corres]

Maybe these numbers show us the way to go, less nps but bigger, really big,
neural networks?
Srdja

Quite Correct. My practical Experience with lc0 on InfinityChess and other Sites confirms your theory.
Using the 128x10 Networks I got beaten practically every time by the AB Engine Users, even though the Knps was very high. Slightly better experience with 320x24 Networks, but not by much.
Using the Large Nets provided by jhortos and Sergio (384x30 !),while getting only 18-20 Kn/s on my ASUS GeForce 2080 Ti, I very rarely lose.
So Yes, Large Networks are the way to go, with the caveat that the GPU should be a Top-end one to get positive results, otherwise it will only backfire, for example if you are using only a 1060 Ti.
With new and faster nVidia GPUs coming out Q1, 2020, things can only get better with lc0.

Where is this big (384x30) network?
On the site of jhorthos we can see the 320x24 nets and no any bigger.
Note
Alone the bigger size is not a guarantee for better play.

[Master Om]

Maybe these numbers show us the way to go, less nps but bigger, really big,
neural networks?
Srdja

Quite Correct. My practical Experience with lc0 on InfinityChess and other Sites confirms your theory.
Using the 128x10 Networks I got beaten practically every time by the AB Engine Users, even though the Knps was very high. Slightly better experience with 320x24 Networks, but not by much.
Using the Large Nets provided by jhortos and Sergio (384x30 !),while getting only 18-20 Kn/s on my ASUS GeForce 2080 Ti, I very rarely lose.
So Yes, Large Networks are the way to go, with the caveat that the GPU should be a Top-end one to get positive results, otherwise it will only backfire, for example if you are using only a 1060 Ti.
With new and faster nVidia GPUs coming out Q1, 2020, things can only get better with lc0.

Where is this big (384x30) network?
On the site of jhorthos we can see the 320x24 nets and no any bigger.
Note
Alone the bigger size is not a guarantee for better play.

https://www.comp.nus.edu.sg/~sergio-v/t40/384x30/

[corres]

...
Where is this big (384x30) network?
On the site of jhorthos we can see the 320x24 nets and no any bigger.
Note
Alone the bigger size is not a guarantee for better play.

https://www.comp.nus.edu.sg/~sergio-v/t40/384x30/

Thanks!
Maybe is the last the most better?

[shrapnel]

Alone the bigger size is not a guarantee for better play.

Coupled with a strong GPU like 2080 or 2080 Ti or even two of them together, it IS.
Master Om gave you the link, try them, the last one is best I think.
Even the 32 Cores Players will be unable to beat you, if you use this Network, at least not easily, given at least a reasonably good Book and a 2080 or 2080 Ti.