Pragmatic reasoning, perceptual biases & emergent meaning in visually-grounded communication between neural agents

Please bwaldon [at] stanford.edu (email) for the Zoom link.

Multi-agent simulations are a standard tool for studying language evolution. More recently communicating agents modelled as artificial neural networks, so-called neural agents, have come into focus. In this talk, we reiterate previous arguments from the “pre-neural” literature that the emergence of a conventional pragmatic code is shaped by pragmatic reasoning and perceptual biases inherent in the agent architecture. We then look at simulation results using neural agents in visually-grounded communication scenarios showing that lexical meanings acquired by pragmatic neural agents from reinforcement learning show a mutual exclusivity bias (thus addressing a challenge put forward by Ghandi & Lake 2020, NeurIPS). Using relational label smoothing (Marino et al. 2021) we manipulate perceptual biases of neural language agents and find that not only do experimentally induced perceptual biases influence the emergent meaning, but also the converse holds: perceptual biases can be passed on through language alone.