2021/11/11 by Łukasz Kuciński, Tomasz Korbak, Kuciński, Łukasz +5 · 1 citation
Neuroscience · Physics and Astronomy · Social Sciences · #Artificial Intelligence (cs.AI) #Cognitive Science and Education Research #Computation and Language (cs.CL) #FOS: Computer and information sciences #Language and cultural evolution #Machine Learning (cs.LG) #Opinion Dynamics and Social Influence
paper · pdf · doi:10.48550/arxiv.2111.06464
openalex publication_date 2021/11/11 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Communication is compositional if complex signals can be represented as a combination of simpler subparts. In this paper, we theoretically show that inductive biases on both the training framework and the data are needed to develop a compositional communication. Moreover, we prove that compositionality spontaneously arises in the signaling games, where agents communicate over a noisy channel. We experimentally confirm that a range of noise levels, which depends on the model and the data, indeed promotes compositionality. Finally, we provide a comprehensive study of this dependence and report results in terms of recently studied compositionality metrics: topographical similarity, conflict count, and context independence.