Current model-free RL methods would likely need millions of hours of practice, with numerous accidents, for an autonomous car to learn to drive.
Unresolvable The public record could not settle it.
Factual: Statements about the present or the past, checked against the record: did they hold up?
What would count as true or false
Judged as of 2019, as a hedged expert estimate. True: research or credible expert estimates from about 2019 indicate that model-free RL methods have such poor sample efficiency that learning to drive in the real world would take on the order of millions of hours of experience (or a comparably vast amount), and would involve many crashes or accidents, because learning by trial and error needs failures. Mostly true: the evidence supports an extremely large requirement with many accidents, but puts the order of magnitude somewhat differently (for example, hundreds of thousands of hours). Mixed: the evidence supports poor sample efficiency, but it is unclear whether the scale or the claim about accidents holds. Mostly false: the evidence shows the requirement is much smaller than claimed, or that it applies only to some narrow setups. False: evidence from that time shows model-free RL could learn real-world driving with modest practice and few accidents.
Set before the research began.
Evidence (3)
-
E1 Learning to drive in a day Wayve · 25 June 2018
“A crucial point to note is that DeepMind’s Atari playing algorithms required millions of trials to solve a task. It is remarkable that we consistently learnt to lane-follow in under 20 trials.”
-
E2 Learning to Drive in a Day Wayve (Kendall, Hawke, Janz, Mazur, Reda, Allen, Lam, Bewley, Shah), arXiv:1807.00412 · July 2018
“From randomly initialised parameters, our model is able to learn a policy for lane following in a handful of training episodes using a single monocular image as input.”
-
E3 My_Bibliography_for_Research_on_Autonomous_Driving/sections/7_model_based_rl.md at master · chauvinSimon/My_Bibliography_for_Research_on_Autonomous_Driving GitHub (personal bibliography notes by chauvinSimon) · 3 July 2019
“The millions of training steps are required to converge, which is impractical for real applications, and the safety of the learned driving policy is not guaranteed.”
Reasoning
E3 says millions of training steps are needed to converge but does not say this refers to model-free RL, and it gives no figures in hours or accidents. E1 and E2 report fast learning of lane following but do not say the method was model-free RL, so the evidence does not show how model-free methods would perform at real-world driving.
Cite this claim
Yann LeCun, “Current model-free RL methods would likely need millions of hours of practice, with numerous accidents, for an autonomous car to learn to drive.” (said 2019). Verdict: unresolvable. Reliability API, claim c_1687ab52e310, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_1687ab52e310
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Yann LeCun
-
Mostly true
“For starters, transformers are equivariant to permutations.”
Transformers are equivariant to permutations, as ConvNets are equivariant to translations.
-
Mostly true
“If it were the case, we would have AI systems that could teach themselves to drive a car in 20 hours of practice, like any 17 year-old.”
-
Mostly true
“The architecture is good, the architecture of the encoder is good, but the fact that you train the system to reconstruct images does not lead it to produce to long, good generic features of images.”
-
Mostly true
“LLMs can do none of those or they can only do them in a very primitive way and they don’t really understand the physical world.”
-
Mostly true
“No, there’s one thing that autoregressive LLMs or that LLMs in general, not just the autoregressive one, but including the bird style bidirectional ones, are exploiting and its self supervised…”
Verdicts are made automatically; how verdicts are made.