How Collective Predictive Coding Hints at the Future of Science with AI Shiro Takagi (takagi4646@gmail.com) 2024/10/24

Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science CPC-MS Interpretates science as a symbol emergence system that performs collective predictive coding 2

1. AI Scientist in Future Science 2. Towards Automated Science System What perspective will CPC-MS offer us for shaping the future of science? 3

1. AI Scientist in Future Science

Lu et al. (2024) The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery https://sakana.ai/ai-scientist/ 5

Lu et al. (2024) The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery An AI Scientist that conducts every step from idea generation, experimentation, paper-writeup to peer review 6

7 Lu et al. (2024) The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Lu et al. (2024) The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery This is just the beggining of the age of AI Scientist! 8

What perspective will CPC-MS offer us on designing future science with the involvement of AI Scientists? 9

In CPC, science is a multi-agent system 10

An AI scientist is just one of the agents with distinct traits 11

Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science 12 This is the graphical model of CPC-MS Unique parameter for each agent

Human & AI scientists would likely have different biases Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science 13

Thus, different internal representation & observation Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science 14

This would lead to diverse global representation sampling -> good for social objectivity (explained by the last speaker) w w w Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science 15

Concerns about the lack of shared common ground <- prerequisite for social objectivity Taniguchi et al., (2024) Collective Predictive Coding as Model of Science: Formalizing Scientific Activities Towards Generative Science 16

Global scientific rerpresentations (Shared explicit knowledge) Proposition from AI/human is highly likely to be rejected by human/AI because their internal representation would look different Human Scientist AI Scientist AI/human would likely reject human/AI’s proposal -> from distribution update perspective, this means poor convergence efficiency 17

The entry of AI Scientists into the scientific community brings diversity Promotes “objective” scientific inquiry Enables more extensive sampling Loss of common ground among scientists Worsened distribution convergence While the diversity introduced by AI Scientists enhances objectivity, it also creates a trade-off by raising communication challenges between human and AI scientists (the human & AI scientists alignment probelm) → It is crucial to discuss the design of science that incorporates AI scientists 18

2. Towards Automated Science System

1900s Robot AI Dendral BACON Adam AlphaFold AI Scientist MLAgent ChemCrow Coscientist MOOSE prompt2model ... Automated Theorem Proving SciML Physics Informed ML 2000s 2012 Laboratory Automation Scientific Workflow Program Synthesis Scholarly Document Processing Automated Experimental Design Literature Based Discovery Symbolic Regression ... Computer ML DNN Nobel Turing Challenge AI for Science 4thScience Curious Agent AI Feynman Geometric DL Galactica Bayes for Science Neural Operator ReviewRobot PaperRobot MLR-Copilot AlphaGeometry data2paper ... WINGS ... ... ChatGPT 2022 Scientific Claim Verifi. Mahoro Solevent SemNet ... DISK 3rdScience [Wang+ 2023] 2017 Transformer AutoML MLOps AM Logic Theorist Automatic Statistician Eve 20 Long history of research automation

Langley et al. (1987) Scientific Discovery: Computational Explorations of the Creative Process Schmidhuber (1991) Curious Model-Building Control Systems 21

Kitano (2021) Nobel Turing Challenge: creating the engine for scientific discovery King et al. (2009) The Automation of Science Waltz and Buchanan (2009) Automating Science 22

Lu et al. (2024) The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery https://sakana.ai/ai-scientist/ 23

AI Scientist Studies related to research automation have mainly focused on automating a researcher’s process or aimed to realize AI “scientist”, but science is collaborative social process and not confined into single agent... 24 AI Scientist AI Scientist AI Scientist AI Scientist

AI Scientist Global scientific rerpresentations (Shared explicit knowledge) AI Scientist AI Scientist CPC-MS is a mathematical/probabilistic model of the entire scientific community and provides a starting point for the automation of the entire scientific community beyond realizing an AI scientist 25

Conclusion CPC-MS enables discussions on the potential impact of AI scientists on science from a probabilistic perspective It also lays the foundation for designing and implementing new forms of science, such as future automated science 26