3 ms·

> Learning to rotate an object in simulation without randomizations requires about 3 years of simulated experience It's interesting to me that this is about th

> Learning to rotate an object in simulation without randomizations requires about 3 years of simulated experience It's interesting to me that this is about the same amount of time it takes humans to develop similar levels of motor control. I don't know enough about AI or neuroscience to say whether it's likely to be a coincidence or not, though.

Ajedi32 8y ago
Interesting observation. I suspect it's probably coincidence though. Other tasks which humans are able to learn (such as [playing Dota][1]) have taken OpenAI much longer to master. OpenAI Five spends 180 years of training per day, per hero in order to learn Dota, and it still isn't at the level of professional players (though that may change soon). Though I suppose you could argue that Dota benefits more from high-level reasoning, whereas basic motor control is a more intuitive skill. (And therefore better suited for this type of AI.) [1]: https://blog.openai.com/openai-five/ https://blog.openai.com/openai-five/
- mortenjorck 8y ago
  Also, the time referenced in the article is presumably three years of non-stop training – given that an infant has a calendar packed with other things like sleeping, crying, and other non-motor activities, total human time logged on learning fine motor control is probably half that, if not less.
andreyk 8y ago
Basically coincidence; the models here are far simpler and do not at all mimic the human nervous system.
- red75prime 8y ago
  I don't doubt that it is coincidence. But it's interesting that vertebrates' brains seems to be using specialized structure (cerebellum) for motor coordination. And here we have far simpler artificial system without any evolutionary priors.
mxwsn 8y ago
It's not really meaningful. Humans learn with entirely different stimuli and experience (can you imagine subjecting a person to learning this object manipulation task in the same way as this robot?). In addition, Alpha Go Zero demonstrated an order of magnitude better efficiency in training time than Alpha Go due to only algorithmic differences. Humans are about converged to 3 year training time for babies learning these skills, but I doubt that these learning algorithms are as efficient as they will ever be.

Ajedi32 8y ago

Interesting observation. I suspect it's probably coincidence though. Other tasks which humans are able to learn (such as [playing Dota][1]) have taken OpenAI much longer to master. OpenAI Five spends 180 years of training per day, per hero in order to learn Dota, and it still isn't at the level of professional players (though that may change soon). Though I suppose you could argue that Dota benefits more from high-level reasoning, whereas basic motor control is a more intuitive skill. (And therefore better suited for this type of AI.) [1]: https://blog.openai.com/openai-five/ https://blog.openai.com/openai-five/

mortenjorck 8y ago
Also, the time referenced in the article is presumably three years of non-stop training – given that an infant has a calendar packed with other things like sleeping, crying, and other non-motor activities, total human time logged on learning fine motor control is probably half that, if not less.

andreyk 8y ago

Basically coincidence; the models here are far simpler and do not at all mimic the human nervous system.

red75prime 8y ago
I don't doubt that it is coincidence. But it's interesting that vertebrates' brains seems to be using specialized structure (cerebellum) for motor coordination. And here we have far simpler artificial system without any evolutionary priors.

mxwsn 8y ago

It's not really meaningful. Humans learn with entirely different stimuli and experience (can you imagine subjecting a person to learning this object manipulation task in the same way as this robot?). In addition, Alpha Go Zero demonstrated an order of magnitude better efficiency in training time than Alpha Go due to only algorithmic differences. Humans are about converged to 3 year training time for babies learning these skills, but I doubt that these learning algorithms are as efficient as they will ever be.