backpropaganda

We embrace that as well: https://arxiv.org/abs/1404.5997

Do you know why TLO's APM is sometimes so large? Did he actually peak at 2000, or is he using a repeater or something like that?

What kind of evidence is going into this analogical reasoning? Do we also extrapolate similarly for other things? We went to the Moon in 1960s. Was Mars a month, or a year, or a decade away? Then we sent robots to Mars. Did we yet send any robots to Alpha Centauri?

Different problems have different difficulties. Solving simple problems quickly doesn't mean we'd also be able to just as easily solve the hard problems. Often the comparably simpler problems have the best reward/effort ratio and thus make quick progress, which doesn't need to be the case for hard problems.

This is incorrect. They shared neural net activations which is explicit communication.

No. Look at what's happening at a meta level and nonlinear effects. Yes, the social good projects will have some small positive effect in the world, but they also end up buying a lot of positive PR for Google, thus allowing them to continue their malpractices in their main business. Skepticism of the kind in this thread informs Google that their positive image right now is actually fragile, and that they should up their ante if they want to continue to be looked at as having a positive impact on the world, which is a critical advantage they use to gather the best talent.

This example didn't even come to my mind when writing the above comment. I think even run-of-the-mill applications of AI and ML such as feed "personalization" in Youtube, Twitter, and Facebook are pretty evil. They're clearly not optimizing for value for the user but for monetizability even if the user is being shown garbage content to get them addicted. I'm a deep learning researcher, and these applications make me seriously re-consider my choices and the impact I'm having in the world.

I hate the term "AI for social good", because it reminds me that the default use of AI today is actually far from being a social good. I wish "AI for social good" was not a thing, and default use cases of AI was for good, social or otherwise.

All the research in fairness, privacy, and other related topics at academic conferences is irrelevant if the industry continues to recklessly deploy unethical machine learning in production.

In fact, it can be argued that industry funds such research only to appear to be ethical and to give themselves cover to continue to use unethical machine learning.

Very cool. Nice trick you guys did there with the lighter colored block, and the red enemy not being able to follow across the boundary. A lot of game developers I respect and whose games I've enjoyed started programming at a young age with simple tools which allowed getting their hands dirty. While javascript is a great language to do this, you might want to consider Game Maker Studio as well.

For this particular business. I don't think he's representative. Using tools like Unity makes it much easier to deploy on Mac and Linux.

I really wish there was a platform to find top experts and ask them questions. Surprisingly enough, Twitch is one such channel. Jonathan Blow streams there some times, and I've asked him a lot of questions to which he gives detailed responses, because the question and answer is useful for all the viewers, and not a waste of his time, because it gives them a way to connect to their fans. Twitch is not meant for this though, and for instance, has no easy way to refer people to questions already answered, etc. But I think such a platform (video stream + text questions) with some modifications can provide this type of platform we're looking for.

The main bottleneck to learning new skills is access to good feedback from experts or coaches, not time.

Note that isn't tabula rasa generation. All these faces are generated from reference images. See this timestamped video for example: https://youtu.be/kSLJriaOumA?t=80. Current deep generative tech requires a carefully designed dataset of images with controlled variation. For the faces above, a dataset of celebrity faces was carefully constructed from a larger wild dataset, so as to not have any rare features, and the eyes and noses were aligned, making the dataset digestible by a GAN. This particular method then also uses multiple style images which it takes certain features from to mix up and create the mixture face.

The main problem with this tech is that the quality can not be reliably controlled. Whatever it generates has to be passed through a human checker/curator pretty much. You can't just serve up generations to the audience without checking what they are. So, imo, this tech would be very useful for the concert art stage of the creative process, being able to just create new things, or mixup existing things, but only to create an intermediate artifact, which would then be used as inspiration for the artists to create the final art.

This is also devoid of any form of higher-level reasoning. It doesn't know concepts a human artist would know, such as object permanence, intuitive physics, etc. Something as simple as a ball falling down should squish a certain way, and when hit the ground should squash another way. It also doesn't have artistic sensibilities like anticipation, etc., that have been laid out in the 12 rules of animation, and even more, there's almost no way to communicate it to these models, short of just creating a huge dataset which has those 12 rules, and just hoping, rather praying, it learns it.

The tagline is meant to emphasize that even good people are divided by politics, and the book is meant to target people who didn't find that obvious.

The difference between epistemic and aleatory uncertainty is not well understood even in the machine learning world. Most applications of machine learning to domains where uncertainty needs to be quantified use aleatory uncertainty, when epistemic uncertainty is the right one. I wish we had better ways to explain these concepts, and perhaps better terms too. Alternatives like intrinsic/extrinsic or irreducible/reducible don't really work either, because they make it seem like it's a dichotomy.

Your intuition is strongly assuming transitivity: "A beats B handily and B beats C handily, therefore this is clear evidence that A beats C handily". I already gave an example where this is not true, and I don't know why you're still doubling down on the same wrong intuition.

> DeepMind didn't find any when training in different ways

Intransitivity may not occur within AG versions but could occur when compared to humans. You need very different types of strategies for intransitivity to occur.

> Are they supposed to run a tournament just to show that a massive improvement that crushes all previous versions can still beat puny humans?

I think so, yes. We do this with Chess. Why should Go be different?

Deepmind is aware of this problem. See gwern's links if you are interested.

Because Zero was not trained on human games and Master was? Do we have any model which was trained only from self-play that can beat humans? I don't think so.

Thanks a lot for the links. They look quite interesting. It does seem that Deepmind is aware of this, and are working on evaluating this.

Transitivity might be true within AlphaGo versions, but that doesn't give me any confidence that it would also hold when a human is in the equation. If a group of policies more or less occupy the same state space, they are likely to be transitive, but if they occupy disjoint state spaces, I don't think we can be sure of transitivity.

I think you're wrong. The version that defeated Ke Jie and 60 masters online was AlphaGo Master, which afaik did train on human games.

We do know that AlphaZero's chess playing style is different. AlphaZero doesn't care as much about material, while Stockfish and humans (presumably learning from Stockfish) do. This is because Stockfish uses a material heuristic to search for moves.

Consider rock-paper-scissor, and agent X which plays X every move. Agent rock beats agent scissor all the time, and agent scissor beats agent paper all the time. But we cannot conclude that agent rock will beat agent paper all the time. Chess and Go might have this intransitivity property. The fact that one game exists which is not transitive means the burden of proof that Chess and Go are transitive rests on the authors.

My own suspicion is that AlphaZero should be weaker against humans compared to AlphaGo Lee since the former is not trained on human games while the latter is.

Comparing AlphaZero to AlphaGo Lee seems problematic. I don't think transitivity holds in this case, i.e. AlphaZero can beat AlphaGo Lee almost surely, and AlphaGo Lee can beat Lee Sedol almost surely, but it could be possible that AlphaZero is not able to beat Lee Sedol at all. This is because the state spaces reached in computer-computer games are probably very different from the state space reached in human-computer games. I could be wrong, but at the very least this should be discussed by Deepmind.

How are you judging success here? The paper has no experiments at all, not even toy.

Awesome! It seems such thin and/or light fonts are okay to read on retina displays (which is where I assume they were tested) but not on other displays. I happened to be reading on a non-retina display.

+1. The code is easy to read, but the main text is hurting my eyes.

Autoencoders lead to other ideas. Understanding AEs and VAEs help you understand GANs, CycleGAN, PixelCNN, Pix2pix, and so on. AEs teach that you can do other things with neural networks, apart from supervised learning. VAEs teach that neural networks can parametrize more complex distributions.

At a competent level? None exist. Being better than random doesn't count as competency. As others have said in this thread, it wouldn't even pass as median performance.

> is only helping to advance the state of reinforcement learning

Zero new algorithms or ideas were introduced by OpenAI Five. We just learn that model-free RL doesn't scale and we already knew that from Atari and robotics benchmarks.

That's a great example. It would be great if you could write a blog post on OpenAI Five. There's a LOT of misinformation on this and could use a treatment like this: https://www.alexirpan.com/2018/02/14/rl-hard.html

Great test. The bots won't be able to pass. Both fast reaction times and lack of cohesive strategy will give them away in the first 5 minutes.

This is the kind of misinformation OpenAI is so great at. Otherwise smart people like tlb are led to believe that this was the full game even though it's not even close to the full game. Thanks for helping me make my point.

> The bots were routinely pulling off coordinated team behavior that players couldn't figure out, but that worked. This has to qualify as complicated.

This is just not true. Do you know the game? Are you speaking as a player? Or are you telling us what's written in the OpenAI blog post? There is nothing a player couldn't figure out. The caster team lost because of the broken game and because few of them were rusted (Merlini hadn't played for months). The bots were garbage at the TI, and got beaten without any problem by the pro teams.

> Already mentioned that the hero pool has been opened up. Along with the removal of the other restrictions (invincible courier, items) this is basically pure DotA.

The hero pool is still 18 heroes. Dota 2 has over 110 heroes. Can you please try to think what makes you say something so wrong with so much confidence?

> The bots beat a team of 5 casters (granted, with some of the older restrictions in place) who are individually in the top 1% of DotA players by MMR.

Casters don't count as serious players. Would you regard any sports commentators as good players? Would you put 5 of them randomly in a team and say that's representative of the best players of that sport?