Collective predictive coding hypothesis and beyond: Symbol emergence as a decentralized Bayesian inference

This talk was given at Theoretical Neurobiology meeting, which is led by Karl Friston, on 11th February 2025.

Abstract:
Humans have formed current language (symbol) systems through adaptation to the real-world environment. In recent years, large language models that achieve language understanding and generation by extensively learning from linguistic corpora through predictive learning have gained attention. However, the computational theory of intelligence that forms language itself as an extension of real-world environmental adaptation is not well understood. Recently, the speaker proposed the collective predictive coding (CPC) hypothesis as a computational model of symbol emergence systems. The hypothesis argues that our symbol systems, in a broad sense, are formed by predictive coding collectively performed by a society which is considered as a subject of predictive coding. The CPC framework can be viewed as a society-wide free-energy principle as well. From this viewpoint, symbol emergence, can be considered as decentralized Bayesian inference embodied as language games in a multi-agent system. This lecture will introduce the concepts of symbol emergence systems and collective predictive coding, discuss recent research achievements related to symbol emergence, and AI alignment to envision a future society where humans and AI robots coexist.

Transcript

1. Tadahiro Taniguchi 1) Professor, Graduate School of Informatics, Kyoto University 2) Visiting Professor, Research Organization of Science and Technology, Ritsumeikan University 3) Senior Technical Advisor, Panasonic Holdings Corporation Talk, Theoretical Neurobiology meeting 11th February 2025 Collective predictive coding hypothesis and beyond: Symbol emergence as a decentralized Bayesian inference
2.  Biography  2003-2006: PhD student, Kyoto University  2005-2008: JSPS research fellow, Kyoto University  2008: Assistant professor, Ritsumeikan University  2010: Associate professor, Ritsumeikan University  2015-2016: Visiting Associate Professor, Imperial College London  2017: Professor, Ritsumeikan University  2017: Visiting General Chief Scientist, Panasonic (Holdings) Corporation  2024-: Professor, Graduate School of Informatics, Kyoto University  2024-: Visiting Professor, Research Organization of Science and Technology,, Ritsumeikan University  2024-: Senior Technical Advisor, Panasonic Holdings Corporation  Research interest  Symbol Emergence, AI, Robotics, Cognitive Systems, Language Tadahiro Taniguchi Email： taniguchi@i.kyoto-u.ac.jp X (personal): @tanichu
3. Symbol emergence in robotics using probabilistic generative models Unsupervised lexical acquisition and spatial concept formation Application to service robotics Multimodal World modeling for manipulation Integration of imitation and reinforcement learning Multimodal object categorization
4. Why do large-scale language models seem to understand the world? Collective Predictive Coding Hypothesis [Taniguchi ‘24]  Language is formed through collective predictive coding (CPC) performed by humans. Therefore, the information of the world is coded in distributional semantics.  A symbol emergence system can be understood as a social representation learning system. Taniguchi, Tadahiro. "Collective predictive coding hypothesis: Symbol emergence as decentralized bayesian inference." Frontiers in Robotics and AI 11 (2024): 1353870. Therefore, LLMs understand the world as if they had a body, don't they?
5. Huge Neural Networks Linguistic Phenomena behind Large Language Models Linguistic knowledge Utterances & Texts Language system Distributional semantics and syntactic structure Huge linguistic resources Learning & Internalization 5 Large Language Models (LLMs) GPT-4 is OpenAI’s most advanced system, producing safer and more useful responses https://openai.com/product/gpt-4 (2023/03/15) It was understood that when human society forms language and when LLMs internalize large language corpora, LLMs can achieve a high level of practical language understanding and use.
6. How can we form "language"? How did language systems come to be? (Emergence) How do language systems as a culture dynamically change? (Dynamics) 6 Why has human society been able to form such a language (without centralized representation learning system)? A) Satisfies distributional semantics. B) Has unique syntactic structures. C) In the absence of being able to see into the other person's mind, the meanings of words emerge and are shared. D) Represents intentions and objects in the real world and is useful for environmental adaptation. The properties of language formed by human society are:
7. Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Symbol emergence systems [Taniguchi+ 2016]
8. Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Representation learning Symbol emergence systems [Taniguchi+ 2016]
9. Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Symbol emergence systems [Taniguchi+ 2016]
10. Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Symbol emergence systems [Taniguchi+ 2016]
11. Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Symbol emergence systems [Taniguchi+ 2016]
12. Emergent symbol system Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Symbol emergence systems [Taniguchi+ 2016]
13. • Human children acquire functions through their own physical experiences and the integration of sensory-motor information, and they acquire language, enabling communication. • They also create symbols (language) to understand the world and cooperate adaptively and autonomously. • SER aims to realize a developmental robot that acquires, invents, and shares language based on real-world physical and social experiences. SER aims to achieve a constructive understanding of the emergence of symbols in human society. Constructive approach to Symbol Emergence Systems ⾕⼝忠⼤「記号創発ロ ボティクス」 2014 ⾕⼝忠⼤「⼼を知るた めの⼈⼯知能」2020 Symbol emergence system ⾕⼝忠⼤ 「コミュニケーションするロボットは創れるか」（NTT出版）2010 ⾕⼝忠⼤「記号創発ロボティクス」（講談社メチエ）2014 Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. Tadahiro Taniguchi, Emre Ugur, Matej Hoffmann, Lorenzo Jamone, Takayuki Nagai, Benjamin Rosman, Toshihiko Matsuka, Naoto Iwahashi, Erhan Oztop, Justus Piater, Florentin Wörgötter, Symbol Emergence in Cognitive Developmental Systems: A Survey, IEEE Transactions on Cognitive and Developmental Systems, 11(4), pp.494-516, 2019. ⾕⼝忠⼤「⼼を知るための⼈⼯知能」（共⽴出版）2020 Symbol emergence in robotics (SER)
14. Multimodal Categorization and Lexical Acquisition by Cognitive Robots [Nakamura+ 2009-] 2025/2/11 14 Takaya Araki, Tomoaki Nakamura, Takayuki Nagai, Shogo Nagasaka, Tadahiro Taniguchi, Naoto Iwahashi. Online Learning of Concepts and Words Using Multimodal LDA and Hierarchical Pitman-Yor Language Model. IEEE/RSJ International Conference on Intelligent Robots and Systems 2012 (IROS 2012), 1623-1630 .(2012)
15. Integration of Multimodal LDA with Nonparametric Bayesian Unsupervised Word Segmentation for Multimodal Object Categorization and Lexical Acquisition [Nakamura+ 2014] T. Nakamura, T. Nagai, K. Funakoshi, S. Nagasaka, T. Taniguchi and N. Iwahashi, "Mutual learning of an object concept and language model based on MLDA and NPYLM," 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2014, pp. 600-607, doi: 10.1109/IROS.2014.6942621. Visual Auditory Haptic Speech By integrating multimodal sensory information and audio signals, the robot was able to form categories, discover words from the sound signals, and obtain a grounded dictionary.
16. Multimodal Representation Learning and Probabilistic Generative Models ⾕⼝忠⼤. "分散的ベイズ推論としてのマルチエージェント記号創発.“ ⽇本ロボット学会誌 40.10 (2022): 883-888. SpCoSLAM HDP-HLM VAE MVAE General form (e.g., POMDP)  The cognitive system of a single agent can (mostly) be modeled by prediction and inference in a multimodal generative model. (In practice, self-supervised learning is sufficient).  Representation Learning, Predictive Coding, Free Energy Principle, World Models 16 Multimodal spatial concept formation and learning navigation Directly unsupervised phoneme and lexical discovery
17. World Models, Predictive Coding and Free Energy Principle in Cognitive Robotics 17  Taniguchi, T., Murata, S., Suzuki, M., Ognibene, D., Lanillos, P., Uğur, E., Jamone, L., Nakamura, T., Ciria, A., Lara, B., & Pezzulo, G. (2023). World Models and Predictive Coding for Cognitive and Developmental Robotics: Frontiers and Challenges. Advanced Robotics, 37:13, 780-806 Advanced Robotics Best Survey Paper Award 2024  Karl Friston, Rosalyn J. Moran, Yukie Nagai, Tadahiro Taniguchi, Hiroaki Gomi, Josh Tenenbaum, World model learning and inference, Neural Networks, 2021
18. Symbol emergence systems [Taniguchi+ 2016] Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Representation learning by individuals 18
19. Symbol emergence systems [Taniguchi+ 2016] Tadahiro Taniguchi, Takayuki Nagai, Tomoaki Nakamura, Naoto Iwahashi, Tetsuya Ogata, and Hideki Asoh, Symbol Emergence in Robotics: A Survey, Advanced Robotics, 30(11-12) pp.706-728, 2016. DOI:10.1080/01691864.2016.1164622 Representation learning by individuals Symbol emergence by groups 19
20. Constructive studies on language evolution and symbol emergence/emergent communication  Symbol emergence/ language evolution in multi-agent agent systems  Language game (typically naming game)-based approach [e.g., Steels 2015]．  Emergent communication in multi-agent reinforcement settings [e.g., Foerster 2016]. [Luc 2015] Luc Steels. The Talking Heads experiment: Origins of words and meanings. Language Science Press, Berlin, 2015. [Foerster 2016] Foerster, Jakob, et al. "Learning to communicate with deep multi-agent reinforcement learning." Advances in neural information processing systems 29 (2016). [Lazaridou 2016] Lazaridou, Angeliki, Alexander Peysakhovich, and Marco Baroni. "Multi-agent cooperation and the emergence of (natural) language." arXiv preprint arXiv:1612.07182 (2016). [Classical approach] 90's - 00's mainly For example, "Talking head experiment" [Steels 2015] modes the process in which agents form categories and sharing labels through language games in the real world. Also, Vogt, Spranger, Belpaeme, and many other researchers contributed to this field Lu Lu Lu = Lu = [Modern DL-based revival] 2016- Two influential papers written by Lazaridou et al. and Foester et al. reboot the trend of studies of emergent communication based on deep learning. Deep reinforcement learning and Lewis signaling game, e.g., referential game, provides the basis of the studies.
21. Metropolis-Hastings naming game [Taniguchi+ 2023] Outline 1. Perception: Speaker and Listener agents (Sp and Li) observe the d-th object and infer their internal representations (assuming joint attention). 2. Communication: Speaker tells the name of the object probabilistically. The Listener determines if it accepts the naming with a certain probability depending on its belief state. 3. Learning: After the communication is performed, the Listener updates its internal parameters for representation learning and naming. 4. Turn taking: Speaker and Listener alternate their roles and go back to 1.  Taniguchi, T., Yoshida, Y., Matsui, Y., Le Hoang, N., Taniguchi, A., & Hagiwara, Y.. Emergent communication through Metropolis-Hastings naming game with deep generative models. Advanced Robotics, 37(19), 1266-1282. (2023)  Yoshinobu Hagiwara , Hiroyoshi Kobayashi, Akira Taniguchi and Tadahiro Taniguchi, Symbol Emergence as an Interpersonal Multimodal Categorization, Frontiers in Robotics and AI, 6(134), pp.1-17, 2019 The ratio of how much the other person's name and the name you assumed yourself match your own beliefs.
22. MH Naming Game is a Decentralized MCMC Bayesian Inference Agent A Agent B Agent A Agent B (2) Sample from a proposal distribution (3) Judge by an acceptance rate , where (1) Sample from the posterior distribution (4) Update parameters ) Decomposition Composition If the acceptance decision is made with the probability the naming game is equivalent to Metropolis-Hastings algorithm, mathematically. It is guaranteed that signs are shared, and categories are formed among agents to approximate posterior distribution 𝑃 𝑤, 𝑧, 𝜃, 𝜙 |𝑜 (given the two agents' observations).
23. Experiment 1: MNIST dataset  Conditions  The MNIST image data and those rotated by 45 degrees were given to Agent A and B, respectively. (10,000 images of 0-9 with 1000 each)  Comparative methods 1. MH naming game: Proposed method. 2. No communication: The two agent categorize the observations independently. 3. All acceptance: The two agents always accept the proposal of the other. 4. Gibbs sampling (topline)︓Direct centralized inference of Inter-GMM+VAE.  +MI︓Mutual learning of VAE and GMM.
24. Experimental Results Clustering Sharing sign Symbol emergence is NOT ONLY for emergent communication, BUT ALSO for better representation learning.
25. Experiment 2: Fruits 360 dataset  Conditions  Fruits 360 image dataset and those rotated by 25 degree were given to Agent A and B, respectively. (Total 2350 images in 10 categories)  The MH naming game performed unsupervised categorization via symbol emergence at the same level as Gibbs sampling (topline) for the two-agent integrated system.  The symbol emergence between two agents leads representation learning and categorization better performance in unsupervised and decentralized manners.  Result Clustering Sharing sign
26. Agent A Agent B Semiotic communication as Inter-personal cross-modal inference  Semiotic communication by sampling: By sampling 𝑤 ∼ 𝑃 𝑤 𝑜 by the Speaker agent and 𝑜 ∼ 𝑃 𝑜 𝑤 by the Listener agent (i.e., ancestral sampling), the Listener can recall the image represented by 𝑤 .  Inter-personal cross-modal inference: From the viewpoint of original integrated PGM, this is considered as cross-modal inference. However, the "modalities" belong to different agents. Inter-personal cross-modal inference Recalled (i.e., reconstructed) images from each sign
27. Collective Predictive Coding Hypothesis Sign Agent 1 Agent 2 Agent K Communication (Sampling and acceptance) Inference Representation learning Collective representation learning Internal representation Observation Decomposition Inference Taniguchi, T. (2023, August 15). Collective Predictive Coding Hypothesis: Symbol Emergence as Decentralized Bayesian Inference. https://doi.org/10.31234/osf.io/d2ty6 • Could symbol/language emergence in human societies be viewed as collective predictive coding? • Language games that promote symbol emergence can be viewed as decentralized Bayesian inference. • The social version of free energy principle. 27
28. Okumura, R., Taniguchi, T., Hagiwara, Y., & Taniguchi, A. (2023). Metropolis-Hastings algorithm in joint- attention naming game: Experimental semiotics study. Frontiers in Artificial Intelligence, 6. Metropolis-Hastings algorithm in joint-attention naming game: Experimental semiotics study [Okumura+ 23]  We tested if people follow the acceptance probability calculated in the MH Naming Game when they play a game similar to the MH Naming Game.  We obtained positive results.
29. Inukai, J., Taniguchi, T., Taniguchi, A., & Hagiwara, Y. (2023). Recursive metropolis-hastings naming game: symbol emergence in a multi-agent system based on probabilistic generative models. Frontiers in Artificial Intelligence, 6. Recursive Metropolis-Hastings Naming Game: Symbol Emergence in a Multi-agent System based on Probabilistic Generative Models [Imukai+ 2023]  The MH Naming Game was extended to an N-agent setting.  Performing communication between randomly selected pairs is considered an approximation of the RMHNG (Recursive MH Naming Game).  It was demonstrated that RMHNG successfully infers the posterior distribution collectively.
30. Application of MHNG to Multi-Agent Reinforcement Learning [Ebara+ 2024] Symbol emergence based on generative models can be extended to emergent communication in multi-agent reinforcement learning by means that reinforcement learning can be formulated as control as inference from the perspective of generative models. 30 Ebara, H., Nakamura, T., Taniguchi, A., & Taniguchi, T. (2023, December). Multi-agent Reinforcement Learning with Emergent Communication using Discrete and Indifferentiable Message. In 2023 15th International Congress on Advanced Applied Informatics Winter (IIAI-AAI-Winter) (pp. 366-371). IEEE. (Competitive Paper Award)
31. Compositionality and Generalization in Emergent Communication using Metropolis-Hastings Naming Game [Hoang+ 2024] 31 Le Hoang, N., Matsui, Y., Hagiwara, Y., Taniguchi, A., & Taniguchi, T. Compositionality and Generalization in Emergent Communication Using Metropolis-Hastings Naming Game. In 2024 IEEE International Conference on Development and Learning (ICDL), (2024)  Approximate expansion of MHNG into a discrete token sequence using a VAE based on GRU as the language model
32. Action Perception Internal representations (World model) Environment (World) From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence) 𝑃 𝑍 |𝑋 Internal representations (latent states) Observations (sensor-motor information)
33. Action Perception Internal representations (World model) Environment (World) From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence) 𝑃 𝑍 𝑋
34. Action Perception Internal representations (World model) Language (Emergent symbol system) Environment (World) Utterance Interpretation Constraint Organization From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence) 𝑃 𝑊, 𝑍 𝑋 External representations (language)
35. Language (Emergent symbol system) From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence)
36. Internal representations (World model) Language (Emergent symbol system) Utterance Interpretation Constraint Organization From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence) Autonomous subject Distributed nodes (Modality)
37. Action Perception Internal representations (World model) Language (Emergent symbol system) Environment (World) Utterance Interpretation Constraint Organization From Predictive Coding (World Model Learning) to Collective Predictive Coding (Symbol Emergence) Distributed (external representation (/representation) learning by humans
38. 38 https://x.com/hayashiyus/status/1831309992638759210 Taniguchi, T., Takagi, S., Otsuka, J., Hayashi, Y., & Hamada, H. T. (2024). Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science. arXiv preprint arXiv:2409.00102. Hayashi, Y. AI Alignment network Formulation of CPC based on the free energy principle [Taniguchi+ 2024] 1. Ordinary variational free energy × Number of agents (Representation learning, predictive coding, world model learning) 2. Collective regularization term (Alignment of external representation w conditioned by internal representation z and, symbol emergence)
39. Peirce's semiotics and arbitrariness http://visual- memory.co.uk/daniel/Docum ents/S4B/japanese/ ss) セミオーシス・記号過程 対象 サイン 解釈項 C.S.パース 「記号論」の祖 プラグラティズムの思想家 https://en.wikipedia.org/wi ki/Charles_Sanders_Peirce Isn't it possible that while we rely on the plasticity within individual brains, we become a collective social intelligence by exploiting the plasticity of symbolic systems? From neuro plasticity to semiotic plasticity
40. https://en.qualia-structure.jp/ https://www.youtube.com/watch?v=2BNkicNcnLI&t=80s Tadahiro Taniguchi, Noburo Saji C02: Symbol Emergence Systems Theory for Qualia Structure PL: Nao Tsuchiya
41. Qualia structure alignment through language emergence 41 Taniguchi, T., Oizumi, M., Saji, N., Horii, T., & Tsuchiya, N. (2024). Constructive Approach to Bidirectional Causation between Qualia Structure and Language Emergence. arXiv preprint arXiv:2409.09413.
42. Structural influence in qualia and language 42 Taniguchi, T., Oizumi, M., Saji, N., Horii, T., & Tsuchiya, N. (2024). Constructive Approach to Bidirectional Causation between Qualia Structure and Language Emergence. arXiv preprint arXiv:2409.09413.  How does distributional semantics affect internal representation formation?  How does distributional semantics emerge through language emergence?
43. Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science [Taniguchi+ 2024]  Collective predictive coding, which encodes a model of the world as explicit knowledge through distributed interactions, seems to be an appropriate model for scientific activity.  CPC-MS = Collective Predictive Coding as Model of Science 43 Taniguchi, T., Takagi, S., Otsuka, J., Hayashi, Y., & Hamada, H. T. (2024). Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science. arXiv preprint arXiv:2409.00102. Scientific knowledge
44. MH naming game and scientific discussion  Scientific discussion, especially the peer review process, is structurally similar to the MH (Mutual Hypothesis) naming game.  The overall process of scientific activities can possibly be viewed as decentralized Bayesian inference, in the same way as the MH naming game.  This leads us to the idea of CP as a model of science (CPC-MS). Taniguchi, T., Takagi, S., Otsuka, J., Hayashi, Y., & Hamada, H. T. (2024). Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science. arXiv preprint arXiv:2409.00102.
45. Towards the creation of symbiotic intelligence based on symbol emergence systems and collective predictive coding  Symbiotic AI alignment  Humans are inevitably affected by the language use and generation of AI (e.g., “delve into”).  LLM is a “summary” of the language of humans with bodies, and in a sense, knowledge is grounded.  Along with the discussion of AI alignment, a discussion of symbiotic alignment of the overarching human-AI coupling system is necessary. Desirable AI as a partner to humans Human-AI Symbiotic Multi-Agent Society
46. Information 僕とアリスの夏物語 ⼈⼯知能の,その先へ (岩波科学ライブラリ) 2022/1/15 横澤⼀彦・編「⼼をとらえるフ レームワークの展開」 （東京⼤学出版会）2022/10/11 6章 記号創発ロボティクス(⾕⼝) ⾕⼝忠⼤, 河島茂⽣, 井上明⼈ (編集)『未来社会と「意 味」の境界: 記号創発システム論／ネオ・サイバネ ティクス／プラグマティズム』勁草書房、2023/8/30 Email： taniguchi@i.kyoto-u.ac.jp X (personal): @tanichu