5 ms·
If you try your hand at machine learning, you learn very quickly that 95% of your amazing bright ideas don't work. The thing you learn by reading about popular
by Eliezer 18y ago
If you try your hand at machine learning, you learn very quickly that 95% of your amazing bright ideas don't work. The thing you learn by reading about popular algorithms isn't how ingenious they are - their popularity tells you the key fact that they often actually work.
Google is an AI company, they just don't advertise the fact. It's quite possible that someone in Google already tried an approach like this, and found that it didn't really work. A lot of people had similar ideas about the Netflix Prize and it didn't work for them.
I don't mean to sound too discouraging here. It's just that brilliant new notions for data you can integrate into your AI algorithm (and search is an AI algorithm) more often than not turn out to simply... not... work. So you don't get enthusiastic about an idea like that because it sounds good - you really do have to try it. If it turns out to work great in practice, like PageRank, then you might have a Google killer on your hands. It's just that the practical performance is hard to guess by eyeballing what a good idea it sounds like.
- josefresco 18y agoGoogle is an advertising company.
- jwilliams 18y agoGoogle is an AI company, they just don't advertise the fact. Google has a strong preference for adaptive algorithms, but I don't think this makes them AI... For the most part the Intelligence part comes from users, not from machines - What Google do well is leverage this.
- ced 18y agoIf you try your hand at machine learning, you learn very quickly that 95% of your amazing bright ideas don't work. I wonder why we can't tell the good ideas from the bad ones. There has got to be some core principles that are not properly known. After all, backpropagation neural networks are exactly an idea of the "bright and ingenious" kind, but it actually works. I was very disappointed that the Stanford Machine Learning lectures were almost exclusively about "how" and not about "why".
- pchristensen 18y agoI was very disappointed that the Stanford Machine Learning lectures were almost exclusively about "how" and not about "why". Maybe they don't know why? You'd have to know the core principles to know why.
- ced 18y agoBut then... How did they come up with all these learning algorithms? Do all researchers have that 95% failure rate?
- khafra 18y agoPerhaps it's a fuzzy version of the halting problem.