CoChat
Sign Up
Cited by 39Categorization, perception, and languageLanguage and cultural evolutionSyntax, Semantics, Linguistic Variation

Communicating artificial neural networks develop efficient color-naming systems

Open access
Chaabouni, Rahma, Kharitonov, Eugene, Dupoux, Emmanuel, Baroni, Marco·2021·Proceedings of the National Academy of Sciences
DOI 10.1073/pnas.2016569118PMC PMC8000426
Open this paper in CoChat — free accountOpen PDF

Abstract

Significance: Color names in human languages are organized into efficient systems optimizing an accuracy/complexity trade-off. We show that artificial neural networks trained with generic deep-learning methods to play a color-discrimination game develop color-naming systems whose distribution on the accuracy/complexity plane is strikingly similar to that of human languages. We proceed to show that efficiency and narrow complexity crucially depend on the discrete nature of communication, acting as an information bottleneck on the emergent code. This suggests that efficient categorization of colors (and possibly other semantic domains) in natural languages does not depend on specific biological constraints of humans, but it is instead a general property of discrete communication systems.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1073/pnas.2016569118

Cites

6 of 6 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Utility-based approaches lack the explanatory power of the IB framework and instead turn to assumptions about environmental factors and fixed vocabulary sizes to account for cross-language variation.

    “Utility-based approaches, including recent studies that evaluate the communicative efficiency of utility-maximizing systems (e.g., Carlsson et al., 2023;Chaabouni, Kharitonov, et al., 2021;Kågebäck et al., 2020) lack this type of explanatory power and instead turn to assumptions about environmental factors (e.g., noise or external rewards) and fixed vocabulary sizes to account for cross-language”

    From Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity (2025) · cites this paper as Chaabouni, Kharitonov, et al., 2021

    The Discriminative Need section proposes differing cultural needs as a source of color-naming variation, and the Discrimination Game section specifies a fixed vocabulary of 1,024 words. But the paper does not establish that utility-based approaches have less explanatory power than the IB framework; its introduction says IB theory is itself agnostic about where on the efficiency curve a system should lie. The fixed vocabulary is a feature of this experiment, not presented as an explanation for cross-language variation. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • The 180 teams trained by Chaabouni, Kharitonov, et al. (2021) never learned communication more complex than 2.4 bits.

    “In addition, whereas the 180 teams trained by Chaabouni, Kharitonov, et al. (2021) never learned communication more complex than 2.4 bits, we found that VQ-VIB models consistently surpassed 2.5 bits.”

    From Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity (2025) · cites this paper as Chaabouni, Kharitonov, et al. (2021)

    The 'Effect of channel discreteness' section says the authors launched 180 REINFORCE runs and found that these agents developed less complex systems than those trained with Gumbel-Softmax. It does not report a maximum of 2.4 bits for those 180 runs, so the numerical ceiling is not verified by our copy. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Prior art in emergent communication found that training with REINFORCE was less stable and led to communication with lower communicative complexity.

    “Prior art in EC has employed both REINFORCE and supervisory methods for reference games (including, in particular, color) and found that training with REIN-FORCE was less stable and led to communication with lower communicative complexity (Chaabouni, Kharitonov, et al., 2021).”

    From Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity (2025) · cites this paper as Chaabouni, Kharitonov, et al., 2021

    In “Effect of channel discreteness” (page not specified), the paper says fully discrete REINFORCE was harder to train and had a higher failure rate, while agents trained with it developed significantly less complex color-naming systems than those trained with Gumbel-Softmax. The paper identifies Gumbel-Softmax as the comparison method, not as a “supervisory method,” so that description is not established by our copy. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • Chaabouni, Kharitonov, et al. (2021) trained 180 teams from scratch to overcome random failures of model training.

    “Chaabouni, Kharitonov, et al. (2021) trained 180 teams from scratch to overcome random failures of model training; in our results we trained 5 teams from scratch and each team converged to useful communication.”

    From Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity (2025) · cites this paper as Chaabouni, Kharitonov, et al. (2021)

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 38

  1. [3]M. D. Hauser, The possibility of impossible cultures. Nature 460, 190– 196 (2009). 19587759
  2. [4]T. Regier, C. Kemp, P. Kay, “ Word meanings across languages support efficient communication” in Handbook of Language Emergence, B. MacWhinney, W. O’Grady, Eds. (John Wiley & Sons, 2015), pp. 237– 263.
  3. [5]G. Zipf, Human Behavior and the Principle of Least Effort (Addison-Wesley, Boston, MA, 1949).
  4. [9]C. Sims, Rate–distortion theory and human perception. Cognition 152, 181– 198 (2016). 27107330
  5. [11]B. Berlin, P. Key, Basic Color Terms: Their Universality and Evolution (University of California Press, Berkeley, CA, 1969).
  6. [15]Y. LeCun, Y. Bengio, G. Hinton, Deep learning. Nature 521, 436– 444 (2015). 26017442
  7. [16]D. L. Yamins, H. Hong, C. F. Cadieu, E. A. Solomon, D. Seibert, J. J. DiCarlo,, Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proc. Natl. Acad. Sci. U.S.A. 111, 8619– 8624 (2014). 24812127 PMC4060707
  8. [18]R. Futrell, “ Neural language models as psycholinguistic subjects: Representations of syntactic state” in Proceedings of the 2019 North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Long and Short Papers) (NAACL, Minneapolis, MN, 2019), vol. 1, pp. 32– 42.
  9. [19]C. Manning, K. Clark, J. Hewitt, U. Khandelwal, O. Levy, Emergent linguistic structure in artificial neural networks trained by self-supervision. Proc. Natl. Acad. Sci. U.S.A. 117, 30046– 30054 (2020). 32493748 PMC7720155
  10. [20]S. Steinert-Threlkeld, J. Szymanik, Ease of learning explains semantic universals. Cognition 195, 104076 (2019). 31756684
  11. [22]A. Baronchelli, T. Gong, A. Puglisi, V. Loreto, Modeling the emergence of universality in color naming patterns. Proc. Natl. Acad. Sci. U.S.A. 107, 2403– 2407 (2010). 20133744 PMC2823871
  12. [23]V. Loreto, A. Mukherjee, F. Tria, On the origin of the hierarchy of color names. Proc. Natl. Acad. Sci. U.S.A. 109, 6819– 6824 (2012). 22509002 PMC3344991
  13. [24]J. Carr, K. Smith, J. Culbertson, S. Kirby, Simplicity and informativeness in semantic category systems. Cognition 202, 104289 (2020). 32502868
  14. [26]S. Scholte, Fantastic DNimals and where to find them. Neuroimage 180, 112– 113 (2016). 29288865
  15. [27]R. M. Cichy, D. Kaiser, Deep neural networks as scientific models. Trends Cognit. Sci. 23, 305– 317 (2019). 30795896
  16. [28]T. Regier, N. Khetarpal, P. Kay, Color naming reflects optimal partitions of color space. Proc. Natl. Acad. Sci. U.S.A. 104, 1436– 1441 (2007). 17229840 PMC1783097
  17. [30]D. Lewis, Convention: A Philosophical Study (Harvard University Press, Cambridge, MA, 1969).
  18. [31]E. Gibson,, Color naming across languages reflects color use. Proc. Natl. Acad. Sci. U.S.A. 114, 10785– 10790 (2017). 28923921 PMC5635863
  19. [32]E. Kharitonov, R. Chaabouni, D. Bouchacourt, M. Baroni, “ EGG: A toolkit for research on emergence of language in games” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): System Demonstrations (2019), pp. 55– 60.
  20. [33]F. Ratliff, On the psychophysiological bases of universal color terms. Proc. Am. Phil. Soc. 120, 311– 330 (1976).
  21. [34]P. Kay, T. Regier, Resolving the question of color naming universals. Proc. Natl. Acad. Sci. U.S.A. 100, 9085– 9089 (2003). 12855768 PMC166442
  22. [35]D. Roberson, I. Davies, J. Davidoff, Color categories are not universal: Replications and new evidence from a stone-age culture. J. Exp. Psychol. Gen. 129, 369 (2000). 11006906
  23. [36]D. Roberson, J. Davidoff, I. R. Davies, L. R. Shapiro, Color categories: Evidence for the cultural relativity hypothesis. Cognit. Psychol. 50, 378– 411 (2005). 15893525
  24. [37]N. Zaslavsky, C. Kemp, N. Tishby, T. Regier, Color naming reflects both perceptual structure and communicative need. Topics Cognit. Sci. 11, 207– 219 (2019). 30457215
  25. [39]C. Hardin, “ Basic color terms and basic color categories” in Color Vision: Perspectives from Different Disciplines, W. Backhaus, R. Kliegl, J. Werner, Eds. (de Gruyter, Berlin, Germany, 1998), pp. 207– 218.
  26. [41]E. Kharitonov, R. Chaabouni, D. Bouchacourt, M. Baroni, “ Entropy minimization in emergent languages” in Proceedings of 37th International Conference on Machine Learning: Virtual Event (Proceedings of Machine Learning Research, 2020), pp. 2718– 2728.
  27. [42]A. Lazaridou, A. Peysakhovich, M. Baroni, “ Multi-agent cooperation and the emergence of (natural) language” in International Conference on Learning Representations 2017 (OpenReview.net, 2017).
  28. [43]S. Havrylov, I. Titov, “ Emergence of language with multi-agent games: Learning to communicate with sequences of symbols” in Conference on Neural Information Processing Systems, I. Guyon,, Eds. (Advances in Neural Information Processing Systems, 2017), pp. 2149– 2159.
  29. [44]C. Hockett, The origin of speech. Sci. Am. 203, 88– 111 (1960). 14402211
  30. [45]M. Studdert-Kennedy, L. Goldstein, “ Launching language: The gestural origin of discrete infinity” in Language Evolution, M. Christiansen, S. Kirby, Eds. (Oxford University Press, Oxford, UK, 2003), pp. 235– 254.
  31. [46]N. Chomsky, New Horizons in the Study of Language and Mind (Cambridge University Press, Cambridge, UK, 2000).
  32. [47]T. Verhoef, S. Kirby, B. de Boer, Iconicity and the emergence of combinatorial structure in language. Cognit. Sci. 40, 1969– 1994 (2016). 26706244
  33. [49]D. P. Kingma, J. Ba, Adam, “ A method for stochastic optimization” in 3rd International Conference for Learning Representations (San Diego, 2015).
  34. [50]C. J. Maddison, A. Mnih, Y. W. Teh, “ The concrete distribution: A continuous relaxation of discrete random variables” in International Conference for Learning Representations, 2017 (OpenReview.net, 2017).
  35. [51]E. Jang, S. Gu, B. Poole, “ Categorical reparameterization with Gumbel-Softmax” in International Conference on Learning Representations 2017 (OpenReview.net, 2017).
  36. [52]J. Schulman, N. Heess, T. Weber, P. Abbeel, “ Gradient estimation using stochastic computation graphs” in Conference on Neural Information Processing Systems, C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, R. Garnett, Eds. (Advances in Neural Information Processing Systems, 2015), pp. 3528– 3536.
  37. [56]V. Mnih,, “ Asynchronous methods for deep reinforcement learning” in Proceedings of the 33rd International Conference on Machine Learning, M.-F. Balcan, K. Q. Weinberger, Eds. (JMLR.org, 2016), vol. 48, pp. 1928– 1937.
  38. [57]N. Slonim, “ The information bottleneck: Theory and applications,” PhD thesis, Hebrew University (Jerusalem, Israel, 2002).

Cited by

Sorted by citations

Authors
Chaabouni, R., Kharitonov, E., Dupoux, E., & Baroni, M.
Affiliations
  • Centre National de la Recherche Scientifique (FR)
  • Meta (United States) (US)
  • Facebook AI Research, 75002 Paris, France;
  • Institució Catalana de Recerca i Estudis Avançats (ES)
Corresponding author
Chaabouni, Rahma
ORCID
Venue
Proceedings of the National Academy of Sciences
Publisher
National Academy of Sciences
Published
2021-03-15
Volume
118
Issue
12
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
DOI
10.1073/pnas.2016569118
PMC
PMC8000426
OpenAlex
W3136558904
Cited by
39
References
6
Citations by year
2026: 6 · 2025: 5 · 2024: 6 · 2023: 6 · 2022: 8 · 2021: 8
FWCI
6.77
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • HAL (submittedVersion, other-oa)
  • Repositori digital de la UPF (Universitat Pompeu Fabra) (submittedVersion, cc-by)
Funders
Agence Nationale de la Recherche (ANR-10-IDEX-0001, ANR-17-EURE-, ANR‐19, ANR-10-IDEX0001-02., ANR-17-EURE-0017, ANR-10-IDEX, ANR-19-P3IA-0001, ANR10IDEX000102 PSL, ANR-17, ANR-10, IDEX-0001-02)
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Categorization, perception, and language
  • Social Sciences › Cultural Studies › Language and cultural evolution
  • Social Sciences › Arts and Humanities › Language and Linguistics › Syntax, Semantics, Linguistic Variation
Keywords
color naming, artificial neural networks, semantic classification, communication efficiency, accuracy-complexity trade-off, language evolution

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review