Sifting the Signal from the Noise
Daniel A. Herrmann, Jacob VanDrunen,
May 16, 2022
Suppose you are observing a person doing things - dancing around, waving a flag, moving about, building a structure. How could you tell whether that person is signaling you. On some models of language acquisition, you already know what will count as a signal - the dance, say - and can learn from that. But what if you don't know? That's what this paper (15 page PDF) studies. The authors develop an 'attention model' to show "simple reinforcement learning agents can still learn to coordinate in contexts in which (i) the agents do not already know what the signal is and (ii) the other features in the agents' environment are uncorrelated with the signal." So you don't need to already know the elements of a language in order to learn the language.
Today: 3 Total: 3 [Direct link] [Share]
