Cultural evolution via iterated learning and communication explains efficient color naming systems

TL;DR

This work investigates how cultural evolution can produce IB-efficient color naming by combining iterated learning with communication (the NIL framework). Neural agents learn within-generation communication and across-generation transmission, guided by a reward that favors perceptually close color identifications, yielding color naming systems that are both IB-efficient and more human-like than purely efficient baselines. The study shows that IL+C yields systems near the IB limit and closer to World Color Survey patterns than IL alone, C alone, or convexity-based accounts, while also highlighting that IB-optimal solutions can be non-human. These findings support IL+C as a plausible mechanism for the emergence of human-like, efficient semantic systems and point to broader questions about how such dynamics generalize to domains beyond color and contexts richer than simple signaling games.

Abstract

It has been argued that semantic systems reflect pressure for efficiency, and a current debate concerns the cultural evolutionary process that produces this pattern. We consider efficiency as instantiated in the Information Bottleneck (IB) principle, and a model of cultural evolution that combines iterated learning and communication. We show that this model, instantiated in neural networks, converges to color naming systems that are efficient in the IB sense and similar to human color naming systems. We also show that some other proposals such as iterated learning alone, communication alone, or the greater learnability of convex categories, do not yield the same outcome as clearly. We conclude that the combination of iterated learning and communication provides a plausible means by which human semantic systems become efficient.

Figures (12)

• Figure 1: Top: Color naming stimulus grid (left), and stimuli plotted in CIELAB space (right). Bottom: 9 color naming systems displayed relative to the grid. The left column contains color naming systems from 3 languages in the WCS. Colored regions indicate category extensions, and the color code used for each category is the mean of that category in CIELAB color space. The named color categories are distributions, and for each category we highlight the level sets between $0.75-1.0$ (unfaded area) and $0.3-0.75$ (faded area). The middle and right columns contain randomly-generated systems of complexity comparable to that of the WCS system in the same row. The middle column shows random systems that are similar to the WCS system in the same row. The right column shows random systems that are dissimilar to the WCS system in the same row; at the same time, there is no other WCS system that is more similar to this random system.
• Figure 2: Efficiency of color naming, following Zaslavsky et al., 2018. The dashed line is the IB theoretical limit of efficiency for color naming, indicating the greatest possible accuracy for each level of complexity. The color naming systems of the WCS are shown in orange, replicating the findings of Zaslavsky et al., 2018. Our RM systems are shown in blue. It can be seen that the RM systems are often closer to the IB curve than the WCS systems are. The inset shows the 9 color systems of Figure \ref{['fig:maps']}, with the dissimilar random systems shown as +.
• Figure 3: Illustration of the neural iterated learning (NIL) algorithm (Ren et al., 2020). The algorithm alternates between communication within a generation, and learning that is iterated across generations.
• Figure 4: Efficiency of the (top) IL+C, (bottom left) IL, and (bottom right) C evolved color naming systems (orange dots), in each case compared with the natural systems of the WCS (blue dots). The black triangle indicates the end state of one run, shown in the inset color map. The histograms above each figure indicate the proportion of systems at the corresponding complexity level.
• Figure 5: Distribution of dissimilarity to WCS systems (minimum gNID to any WCS system), shown for IL+C and RM systems. The RM systems include both $\text{RM}_{\text{s}}$ and $\text{RM}_{\text{d}}$. Evolved IL+C systems tend to be more similar to attested WCS systems than are random but highly efficient RM systems.