Collective Predictive Coding and Active Inference: A New Perspective on Emergent Communication

Hello, welcome everyone. It is the 22nd of April 2025. We're in active inference gueststream 104.1 with Tadhiro Tanaguchi. We will be hearing a presentation and a discussion on collective predictive coding and active inference. a new perspective on emergent communication. So, thank you for joining and looking forward to the presentation. Okay, thank you very much. I thank you very much for giving me this great opportunity and speak at this great live stream. So, I'm Taniguchi from Kyoto University, Japan. Okay, today I will talking about collective practical coding and active inference a new perspective on emergent communication. So maybe collective maybe uh for many of you the corrective predictive coding maybe a new term and I would like to uh introduce the this our theory. Okay. So before starting let me introduce myself. I'm and actually the uh I have been working on the simple emergence for a long time and but in more pro sense the AI and robotics and especially cognitive developmental robotics is my background. Okay. Last year I moved to Kyoto University and I'm now the professor at uh Koto University in Japan. So if you have a chance to visit is a beautiful city, please stop by our laboratory. Okay. Anyway, uh let me start. So before uh starting uh let me let me share the key key points of today's talk. So I will propose the collective predictive coding CPC in short as a dynamics underlying symbol emerging system. I will explain what is symbol emerging systems later as well and present a formulation of uh rangage emergence based on probabilistic genative models showing its relationship between to the free energy principle and active influence. So also position CBC as a natural generalization of FP to the group or societal level and presents the CPC hypothesis as an interpretation of why large models appear to understand the structure of the physical world and also introduce new perspectives uh derived from or inspired uh by CBC. BC such as CBC as a model of science and quad process theory. Maybe the topic will maybe be too much but uh I would like to introduce our idea. I hope it is exciting. So and before diving into today's main topic maybe the CPC as I mentioned CPC will be a uh new term for you and let me introduce the corrective predict coding CPC hypothesis briefly but this uh hypothesis extends the idea of predictive coding to social and linguistic phenomena. So it states that simple systems including language are formed through collective the predictive coding by humans. So in other words simple emergence or language emergence can be understood as a social representation learning process. So as a result the structure of the world may be encoded in distributional semantics uh in language. So the hypothesis also suggests that this is why appear to understand the world they inherit this structure through vast amounts of distributed human genetic text. Okay. So and the main topic is CPC and the free energy principle and so on. But uh let's let me choose the LM as a starting point of today's talk maybe because everybody knows recently. Okay. So um we have but what are they actually based on? So they are completely reliant on human language data. Okay. So let's take a moment to reflect on our linguistic lives our language human language. So actually as I showed on this slide on this slide. So each individual we uh possesses a unique set of uh linguistic and the world knowledge but no one has complete knowledge or master complete knowledge of language or knowledge of the world. So and actually we dynamically learn even when we when we were children or infant. So we directly learn and adapt our linguistic knowledge through communication and including reading, writing and conversation. So rockage system itself is not a concrete uh fix system. So that is always changing and that is at some have say above ourselves. So nobody has complete the language system. Importantly uh through uh communication the language system as…