Yann LeCun

Factual · AI capabilities

“In reinforcement learning, you give the machine a single scalar, you tell the machine you did good, you did bad, and you only tell this to the machine once in a while.”

Yann LeCun, interviewee · 2021 · Lex Fridman #258 transcript: Dark Matter of Intelligence and Self-Supervised Learning (Enacast)

This quote could not be matched word for word to the source page.

In reinforcement learning the machine receives only a single scalar feedback, and only occasionally.

Mostly true True in substance, wrong in a detail.

The second judge said mixed; the verdict shown is the first judge's.

What would count as true or false

Judge as of 2021 against the standard definition of reinforcement learning in authoritative textbooks (e.g. Sutton & Barto) and survey literature. True: these sources describe standard RL feedback as a scalar reward signal, and the claim's description (a single scalar, given sparsely or only occasionally) fits standard RL as generally described. Mostly true: the scalar-reward part is correct, and 'only once in a while' describes a common or typical case (sparse or delayed rewards) but is not a defining feature, because many standard setups give a reward at every step. Mixed: one of the two parts (single scalar; only occasionally) holds clearly and the other is generally wrong. Mostly false: RL feedback is scalar only in a narrow subset of formulations, or the claim misdescribes it in a way that distorts the field. False: authoritative sources show RL does not typically use scalar reward feedback.

Evidence

1. [PDF] Reinforcement Learning: An Introduction ****Draft**** - UMIACS Richard S. Sutton and Andrew G. Barto (draft of Reinforcement Learning: An Introduction, second edition in progress; MIT Press)

   “On each time step, the environment sends to the reinforcement learning agent a single number, a reward.”

2. [PDF] Reinforcement Learning: An Introduction ****Draft**** - UMIACS Richard S. Sutton and Andrew G. Barto (draft of Reinforcement Learning: An Introduction, second edition in progress; MIT Press)

   “Reinforcement learning also is not restricted to problems in which behavior breaks down into separate episodes, like the separate games of tic-tac-toe, with reward only at the end of each episode.”

3. 3.1 The Agent-Environment Interface Richard S. Sutton (incompleteideas.net, online draft of Sutton and Barto's reinforcement learning book) · 1 January 1997

   “Rewards, on the other hand, are always single numbers.”

4. Reward is enough Artificial Intelligence (Elsevier), review article by Richard S. Sutton · 24 May 2021

   “In addition to their generality, rewards also provide intermediate feedback, potentially at every time-step, on progress towards the goal.”

Reasoning

Sutton and Barto's textbook drafts describe the RL reward as a single number (E1, E3), which supports the scalar part of the claim. The same sources say the agent receives a reward on each time step (E1), and that RL is not restricted to settings with reward only at the end of an episode (E2). Sutton (2021) adds that rewards give feedback "potentially at every time-step" (E4). So the scalar part holds, but "only once in a while" describes a common case rather than a defining feature, which matches the criteria for mostly true.

Cite this claim

Yann LeCun, “In reinforcement learning the machine receives only a single scalar feedback, and only occasionally.” (said 2021). Verdict: mostly true. Reliability API, claim c_ad7582694fdf, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_ad7582694fdf