Comment by vladislav | Hacker News Reader

Comment by vladislav | Hacker News Reader

This article is using false assumptions about AI to back up its narrative.

"AI doesn’t just require top-tier talent; that talent is all but useless without mountains of the right kind of data. And who has essentially all of the best data? That’s right: the abovementioned Big Five, plus their Chinese counterparts Tencent, Alibaba, and Baidu."

Making the next quantum leaps in AI is not a question of data advantages.

1) We have just seen the world's best Go-playing AI trained completely from self-play, without access to any labeled data nor hand-engineering, trained on a few machines. This happened at Google, but could have easily been done by a small startup.

https://thenextweb.com/artificial-intelligence/2017/10/20/go...

2) Even the pioneers of deep learning are now strongly pushing back on the methodology that lots of labeled data is required to solve AI problems.

Geoffrey Hinton: "I don't think it's how the brain works. We clearly don't need all the labeled data."

https://www.axios.com/ai-pioneer-advocates-starting-over-248...

In a nutshell, the biggest challenge of pushing AI forward is tackling unsupervised learning, not having "better data".

1) That's false. AlphaGo did not just train by self-play. It trained on millions of pre-played games, and the bot also uses hand-engineered features for their MCTS hybrid (best) model. Refer to their paper for details.

2) Read the article that you are referencing please. What you are implying is not the thesis of that article nor is it what Geoffrey Hinton is saying.

What Hinton is saying is, we should throw out deep learning. What he is saying is that the current approaches to AI are fundamentally broken and aren't going to result in artificial general intelligence.

Backpropagation is not just used in supervised learning, it is also used in unsupervised learning. I happen to agree with Hinton, in that there is too much hype around the current state and successes of AI, which has mainly been in "narrow AI".

AI is a term that gets thrown around a lot these days. But there is a big difference between technology that automates the tedious tasks of daily life, and artificial general intelligence.

In the world of deep learning, data is king.

https://deepmind.com/blog/alphago-zero-learning-scratch/

>After just three days of self-play training, AlphaGo Zero emphatically defeated the previously published version of AlphaGo - which had itself defeated 18-time world champion Lee Sedol - by 100 games to 0. After 40 days of self training, AlphaGo Zero became even stronger, outperforming the version of AlphaGo known as “Master”, which has defeated the world's best players and world number one Ke Jie.

Sorry, but you are incorrect on both accounts.

1) AlphaGo Zero was indeed trained in the way I mention.

2) As directly quoted from the article, Hinton believes that a better way of learning doesn't require all that labeled data. If such a method is invented, as is required to push AI forward, big corporations would not have a data advantage, which is my original point.

AlphaGo, trained using supervised learning using games from KGS:

https://storage.googleapis.com/deepmind-media/alphago/AlphaG...

AlphaGo scratch: https://www.nature.com/articles/nature24270.epdf

I was specifically talking about AlphaGo, not AlphaGo scratch. Also, if you read the paper about AlphaGo scratch, the key innovation driving the self-learning is the use of MCTS as a policy improver, which couldn't have feasibly been done without AlphaGo and the supervised learning.

And I think Hinton is saying that we need fundamental breakthroughs for AI, and I don't think he is in favor of "traditional" modern neural network architectures. Anything that requires SGD, he doesn't like.

You can also leverage Google's apis and their expertise. These kind of articles forget about the api economy. Then we have new technologies like "serverless" that radically lower the bar to entry for new business to get into the game. Why did I work for the guy who founded Geocities instead of founding something myself? Up to my neck in studio loans and remember in 1998 you had to cough up $1 million for a starter oracle license to go big, today you can scale as you go. The only thing that will end the startup is a lack of creativity and vision, it keeps getting easier and easier to jump in.

I think the conclusion may still be sound despite the reasoning. It may not be a "data advantages" question, but it is still almost surely going to be a question of human capital -- and the only places likely able to afford the talent & operational costs are going to be "Big N" companies. There is a tremendous amount of human expertise involved in the advancements we're seeing in AI.

AlphaGo Zero is the product of incremental improvements made by the same group of computer Go experts that built the original system. _They_ learned from each experiment and incorporated that knowledge in the next version. It was not obvious _a priori_ that the AlphaGo Zero architecture or training method would succeed; if it had been, then the AlphaGo team would not have built the earlier versions. And while it runs on "only" 4 TPUs for inference, those TPUs are each about 30x as powerful as the computers that the original AlphaGo ran on, so it's more like a reduction from ~180 GPUs to 120 equivalent GPUs.

I was mostly disputing the claim about data advantages, which often gets thrown around willy nilly to favor big corporations, because it fundamentally goes against the nature of where AI is heading.

I am not disputing that having more financial resources would help anyone hire talent and build awesome infrastructure, and certainly the latest way of training AlphaGo Zero was aided by earlier experiments that relied on labeled data and extensive computational effort. However, by no means do I think big corps have a lock on these kinds of advancements. There will always be great people who would rather go the startup route, and both algorithmic and hardware advancements are drastically reducing the operational cost of training AI systems. Thus, when it comes to AI, I think very small teams will be able to get very far with the right approach.

Reply on news.ycombinator.com