Seminar on Signalling and Modelling (2 presentations)

7/04/2021 (Wednesday!), 16:00 (GMT+1, Warsaw), Zoom, click here to join the meeting

Just after Easter, the HILL seminar will be (exceptionally) on Wednesday (!) 7.04.2021, 16:00 - 18:00. Zoom link:

The first part of the meeting will be dedicated to the work of Łukasz Kuciński [1], Tomasz Korbak [2], Paweł Kołodziej [1] and Piotr Miłoś [1]. They will deliver a presentation entitled “Sufficient conditions for the emergence of compositionality in Lewis signaling games”.

[1] Institute of Mathematics, Polish Academy of Sciences [2] Department of Informatics, University of Sussex

In our talk, we will present some theoretical and empirical results regarding sufficient conditions for the emergence of compositional communication in Lewis signaling games. Theoretically, we show that in order for compositionality to be possible, the agents must possess certain inductive biases, and the environment must manifest corresponding regularities. Empirically, we present an array of experiments and ablation studies corroborating those insights.

Then we will discuss the CogSci poster by Krzysztof Główka, Julian Zubek, Borys Jastrzębski and Joanna Rączaszek-Leonardi“Compositional signals in a situated communication task: easier used than made”

This work is an attempt to create a novel model of communication emergence. To take a fresh look at recurring problems, we abandon the widespread assumption that language meaning can be modeled as straightforward mappings from signals to meanings. Instead, we model embodied and situated communication in which signal production and signal comprehension are two different abilities, requiring distinct adaptations. Under such conditions, a particular tension emerges between low signal complexity and high signal expressiveness. Our agents evolve functional communication protocols, which allows them to achieve significant task performance while keeping the signal entropy low. Because of the tendency to minimize the signal entropy, more complex compositional protocols do not emerge, but agents receiving fixed compositional signals learn to exploit them and outperform agents receiving non-compositional signals. This allows to ask novel questions about the exploitation and emergence of compositionality.

See you on Wednesday (exceptionally), 7th of April from 16:00 on Zoom: