Does anyone have more info on this part? I know on their website it stated they are working on releasing a teaching tool.

So my questions are is this a web based tool or a client based that can be downloaded? Given that Google Earth went from exe to in Chrome browser, if I were a betting man I'd guess that this tool will be released on the Google Cloud/ App platform and will require Chrome and a Google account in order to be used.

If that is the case, that also means that it will need an active Internet connection, and possibly that the brains of the tool will not be local client, nor calculated in the browser itself, but instead will be all on Google's backend.

The final concern is if this tool will be truly open to the public, or if it is based on invite only system where its open to top pros and what Google really meant when it says "hopefully" it can impact go community and go fans is that by only allowing top pros to access this web based tool, it is their hope that it can indirectly help give insight and indirectly contribute to the go fans and go community by a sort of "trickle down" effect.

Even if I'm wrong and the tool will be accessible to everyone there is still the possibility that it doesn't really the sort of analysis that most of us think of when we think of a Go program analyzing a move or a position. It is possible that it would be locked down to the library of existing sgf games, so that we can use it to reference any move or any position of any game that has already been played on KGS, OGS, and/or by top pros (GoGoD etc) but it would not allow any arbitrary upload of a sgf so that not anyone will be able to upload their OWN games for move suggestions and /or positional analysis.

I say this because if that were the case, if this tool allows anyone anywhere to query any game and any position, then by mere fact of the tool suggesting a best move or giving a full positional analysis, it would in effect almost be like having a copy of AlphaGo to play with.

But if Google was okay with effectively releasing AlphaGo into the wild, Deepmind wouldn't have been quoted to state there were "compilations" to doing so, and they would have released and published the neural network weights, the training dataset of hundreds of millions of AG vs AG games, the tools and processes used to obtain the weights and the training data, and other things such as the underlining source code of AlphaGo, and other research data.

Instead all we will be getting is a measly 50 self playing games, a new paper in a few months, and a supposedly "tool" for teaching/learning.

So I'm really concerned given that it doesn't seem like Google will ever release AlphaGo in any form, client side, web access, etc that its "learning/teaching tool" might not be what we think or expect or ideally hope it might be?!

Any thoughts.

I think it's reasonable to wonder what this tool will be in practice, and I suspect you're right that it'll be very limited.

I'm not too worried in spite of that. Alphago has spurred tremendous innovation in a short amount of time, and while there might not be any team that's as good as DeepMind, other teams will eventually catch up. The same thing happened with Deep Blue. The ordinary player never got any direct chance to play against or make use of it, but now there are open source engines that play better on laptops. Maybe training neural nets will prove more of a challenge for amateurs, but I still have high hopes.

_________________
Occupy Babel!

I'm pretty sure that you can't directly play with AlphaGo. And it might only available to go organizations for professional use only to preserve the value of go business industry.

I don't think the tool will be that interesting. More importantly, I'm looking forward to the scientific paper they're going to publish so that others can start replicate and eventually improve on AlphaGo's performance.

Maybe we need a better description of what a neural net is, what training of a neural net entails (the EFFECT of training), and how, once a neural net has been trained, a second neural net of the same "geometry" can be "cloned" << already trained >>

Think of your brain. It is composed of a large number of neurons (brain cells) connected in a complex way. That is, these neurons have connections to a set of neighboring neurons, sending signals to some of these and receiving signals from some of these (the strength of the signals varying, thus zero strength implying a cut connection). If the total of the incoming signals is over some threshold, then the neuron sends, otherwise not.

OK, now imagine this sort of thing being simulated by a computer program. Each "cell" has in and out connections to some set of neighboring cells. Each cell might have a threshold value, and each of the connection in (or could be on the out) a value that modifies the signal coming or being sent on that connection. So you can think of each "cell" of the neural net having a set of values stored, one a threshold value and the others a coefficient for each connecting channel.

TRAINING the neural net is not changing the program, just adjusting all of these values until the neural net evaluates the function it is being trained to evaluate. So if one had an identical (in structure) neural net that had not (yet) been trained, could simply copy over all the values from the already trained one.


<< neural net folks, have I given a description that might be useful for all these folks that seem to have a confused idea of what a neural net is? >>

Note that is is the LOGICAL structure of the neural net that matters, not the program implementing that neural net on particular hardware. It might take a very powerful processor to do this fast enough to play go at the specified time control, but a less powerful processor could get the same result, just be slower. I have been seeing references to "super computers". That is not really what is involved here. The implementations are making use of GPUs to speed up by parallel processing many cells at once << the total process of a "cell" going through summing up all of its (modified by coefficients) inputs and the firing or not (based on its threshold) is a fairly short/simple matter withing the capability of a GPU thread >> So the device being used by AlphaGo is more like a really powerful workstation/server class machine with a number of GPUs (very powerful video chip)s. A machine maybe an order of magnitude more powerful that a very high end laptop with a very high end video chip << a workstation class laptop, OR like one used by really serious "gamers" >>

That means it is far from clear to me that the "analyze" program would run too slow to be useful on a powerful home machine.

The new version of AlphaGo doesn't use any GPUs. It runs on a single one of these Google's new ASICs that are small enough to be held on your hand.

It's best not to think of artificial neural networks as analogous to a brain. They were originally biologically inspired, but their structure and mechanisms are quite different. The inputs and outputs of biological neurons are voltage spikes; artificial neuron activations are continuous, real-valued numbers.

In terms of the number of neurons, they are several orders of magnitude smaller than a mammalian brain (if they continue their current exponential growth, they might reach the size of human brains by 2050.) A pretty big one is comparable to something like an ant. On the other hand, they are much more densely connected, so large ones are already similar to a mammalian brain in terms of total number of connections.

It's better to think of them in terms of statistical analysis tools like linear regression. In fact, many applications of machine learning use simple, single-layer linear regression or logistic regression models. Larger neural networks like AlphaGo are just several layers of this, stacked together in various configurations.

A neural network of sufficient size can learn to approximate any function mapping one vector of numbers (or higher dimensional array; a matrix or tensor) to another. E.g. mapping a tensor of values representing an image of a cat (a height x width x 3 tensor of RGB values) to a vector representing the word "cat". Going further, certain neural networks are turning complete and can learn a very complex function - something like a program. They're implemented as a series of linear algebra operations (and a few other math functions mixed in).

The computationally intensive part is multiplication of large vectors/matrices/tensors. That's why GPUs are so good at implementing neural networks and order of magnitudes faster than general purpose CPUs, They're specialized hardware for running lots of vector/matrix operations very quickly; a nice coincidence.

Based on the descriptions that have been given about the current version of AlphaGo, it's likely that that it would run fine on a workstation-class PC (or gaming rig). It would just be somewhat slower. NVidia datacenter GPUs are not radically different from their commodity gaming cards; sometimes even the same die with different features enabled/disabled. Unfortunately, it would require pretty significant software changes to decouple it from Google infrastructure and its use of TPUs.

_________________
KGS 4 kyu - Game Archive - Keyboard Otaku

The assumption behind my post was that DeepMind is not going to release enough information to just copy the neural net, and that what we'll be left with is other teams attempting to use the bits of information they release to try and create their own programs. Everything I've heard suggests that training a policy network to the level of AlphaGo is quite expensive, though DeepMind's recent innovations have somewhat lowered the cost.

I hope I'm proven wrong, because being able to run AlphaGo for evaluation tomorrow would be amazing. That said, I believe that the third paragraph of this post is about Google...https://danluu.com/wat/.

Edit: replaced "is" with "suggests"

_________________
Occupy Babel!

Not sure what "not what it seems" mean, since there is no information about it to start with.

Hi hyperpape,

Thanks for the danluu link -- ended up reading a bit of his writings.

This ASIC is a Google TPU?

Thanks for the share, interesting to say the least:

There’s the company that’s incredibly secretive about infrastructure. For example, there’s the team that was afraid that, if they reported bugs to their hardware vendor, the bugs would get fixed and their competitors would be able to use the fixes. Solution: request the firmware and fix bugs themselves! More recently, I know a group of folks outside the company who tried to reproduce the algorithm in the paper the company published earlier this year. The group found that they couldn’t reproduce the result, and that the algorithm in the paper resulted in an unusual level of instability; when asked about this, one of the authors responded “well, we have some tweaks that didn’t make it into the paper” and declined to share the tweaks, i.e., the company purposely published an unreproducible result to avoid giving away the details, as is normal. This company enforces secrecy by having a strict policy of firing leakers. This is introduced at orientation with examples of people who got fired for leaking (e.g., the guy who leaked that a concert was going to happen inside a particular office), and by announcing firings for leaks at the company all hands. The result of those policies is that I know multiple people who are afraid to forward emails about things like insurance updates for fear of forwarding the wrong email and getting fired; instead, they use another computer to retype the email and pass it along, or take photos of the email on their phone. Normal.

Now that I think of it, is it even possible for anyone anywhere, outside of Google, be it a private person, or a corporate entity, to PURCHASE a TPU from Google? Or is Google keeping it all under wraps and never sharing the hardware and merely renting the use of the hardware as a hidden /remote service and basically welcoming other's to pour money, time, and development energies into pumping up Google's platform and helping it reach larger economies of scale by time-sharing it all in a way yet still retaining all the secrets and never giving out the actual hardware?

If this is the case, it does not bode well for the future of AI and all the talk of "democratiazation" of AI when in reality given the trends, Google's alone stands unique in holding the keys to the kingdom and the rest of us just live in it.

Did you try to Google the answer?

I duckduckgo'd it,

https://www.quora.com/Where-can-I-buy-T ... p-learning

Apparently my worst absolute fears have come true... its isn't possible to purchase a TPU, only rent.

Is there any particular reason to believe that's about Google? The author's resume lists quite a few companies, not just the big G: https://github.com/danluu/tex-resume

And for what it's worth, my impression is that Googlers are often quite chatty. Or at least, leaks about upcoming Google products and rumors about what Google is planning seem fairly common. Some of these may be intentional 'leaks' or just false rumors, but they often seem genuine.

Plus I can't recall any other ex-Googlers talking about the company publicly executing leakers, word of which would definitely get around.


First of all, it seems strange to me that you're trying to demonize Google for providing a cloud computing platform, a service that's quite common and fairly useful. And something which Amazon, Microsoft, and other companies also sell. Your hostility seems quite overboard.

As for TPUs, there are other companies selling machine learning hardware: http://www.anandtech.com/show/11367/nvi ... -announced

And while Google claims TPUs are faster for certain tasks, you can write code that runs on other hardware as well. You aren't locked in.

Between this and your calls for DeepMind open source AlphaGo you seem to be in the habit of making excessive demands. I mean, I love open source too, but you don't need a board game AI and you're not entitled to its source code. I think you should take a step back and check your aggression at the door.

I'm glad that you have such a comfortable life free from war, poverty, famine, disease and tragedy that your worst fears are the inability to purchase a high spec processing unit recently developed by a private company that we wouldn't even know existed if they hadn't publicised it.

I'm more interested in the exchange of human resources that specialized in the development of AlphaGo, for example Dr. Aja Huang. I don't believe that they are all struck by a long term non-disclosure agreement. We'll probably find some of them working for other Go software / (hardware?) companies or startups.

_________________
Couch Potato - I'm just watchin'!

Indeed, but also Google/DeepMind treat their employees well and will likely give them nice things to encourage them to stay for a while. In fact DeepMind have hoovered up rather a lot of the top AI talent around the world: if you look at the authors of various Go bot papers, such as early 'policy' neural networks from university research groups, they then ended up working at DeepMind. Or David Silver, who I think knows Demis back from his game studio days and came up with the RAVE technique in MCTS, quit UCL to join. Also Aja came to England a few years ago specifically to work at Google.

Microsoft was once in computer go race too

https://www.microsoft.com/en-us/researc ... way-to-go/

But since Thore Graepel move to DeepMind I think that project got cancelled

Just for the historical record, they also used the GoGoD database (free but with permission).