Communicating artificial neural networks develop efficient color-naming systems

Abstract

Words categorize the semantic fields they refer to in ways that maximize communication accuracy while minimizing complexity. Focusing on the well-studied color domain, we show that artificial neural networks trained with deep-learning techniques to play a discrimination game develop communication systems whose distribution on the accuracy/complexity plane closely matches that of human languages. The observed variation among emergent color-naming systems is explained by different degrees of discriminative need, of the sort that might also characterize different human communities. Like human languages, emergent systems show a preference for relatively low-complexity solutions, even at the cost of imperfect communication. We demonstrate next that the nature of the emergent systems crucially depends on communication being discrete (as is human word usage). When continuous message passing is allowed, emergent systems become more complex and eventually less efficient. Our study suggests that efficient semantic categorization is a general property of discrete communication systems, not limited to human language. It suggests moreover that it is exactly the discrete nature of such systems that, acting as a bottleneck, pushes them toward low complexity and optimal efficiency.

Cognition as a Mechanical Process

“…An example is a recent demonstration that colors in a visual scene are efficiently expressed by language. This was shown in an artificial neural network [39]. This shows evidence for efficiency in the neural coding of information and that artificial models emulate this efficiency.…”

Emergent Color Categorization in a Neural Network trained for Object Recognition

“…The fact that the current categorical representation appears to emerge in the absence of color naming emphasizes that explicit color naming is not a necessity for the development of categories. This may seem to stand in contrast to many of the recent studies that use communicative concepts as a means to model the shape of categories (Chaabouni et al, 2021;Gibson et al, 2017b;Twomey et al, 2020;Zaslavsky et al, 2020). However, many of those studies derive their predictive power from combining these concepts with the nonuniformities in the utility across colors.…”

“…While the degree of importance for communication in shaping categories varies, many of these recent studies rely on communicative concepts when it comes to shaping color categories. Notably, a recent study where communicating deep neural networks played a discrimination game demonstrated that allowing continuous message passing made the emergent system more complex complex and decreased efficiency (Chaabouni, Kharitonov, Dupoux, & Baroni, 2021).…”

Conceptual Similarity and Communicative Need Shape Colexification: An Experimental Study

“…Similar adaptation and efficiency effects have also been observed in experimental settings with artificial languages (cf. Chaabouni, Kharitonov, Dupoux, & Baroni, 2021; Fedzechkina, Jaeger, & Newport, 2012; Guo et al., 2021; Nölle, Staib, Fusaroli, & Tylén, 2018; Tinits, Nölle, & Hartmann, 2017; Winters, Kirby, & Smith, 2015).…”