Why and how to study the impact of perception on language emergence in artificial agents
Year: 2021
Type: article
Abstract: The study of emergent languages in deep multi-agent simulations\nhas become an important research field. While targeting\ndifferent objectives, most studies focus on analyzing properties\nof the emergent language—often in relation to the agents’\ninputs—ignoring the influence of the agents’ perceptual processes.\nIn this work, we use communication games to investigate\nhow differences in perception affect emergent language.\nUsing a conventional setup, we train two deep reinforcement\nlearning agents, a sender and a receiver, on a reference game.\nHowever, we systematically manipulate the agents’ perception\nby enforcing similar representations for objects with specific\nshared features. We find that perceptual biases of both sender\nand receiver influence which object features the agents’ messages\nare grounded in. When uniformly enforcing the similarity\nof all features that are relevant for the reference game,\nagents perform better and the emergent protocol better captures\nconceptual input properties.
Authors: Xenia Ohmer, Michael Marino, Michael Franke, Peter König
Language: English
Subfield: Cultural Studies
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
Funder: Deutsche Forschungsgemeinschaft
OpenAlex ID: W3185761978
