Co-Creative Learning via Metropolis-Hastings Interaction between Humans and AI

Hello, welcome everyone. It is August 20th at the stroke of zero UTC on the 21st in 2025 and it's Active Inference Guest Stream 118.1. We're here with Tadahiro Taniguchi and we'll see a presentation on co-creative learning via Metropolis Hastings' interaction between humans and AI. And then I will read any questions in the live chat and ask some things about this fascinating research. So thank you for joining again, Tadahiro, and to you for the presentation. Okay, thank you very much for inviting me to this great guest swim again. So I'm Tadahiro Taniguchi from Kyoto University. And today I'll be talking about co-creative learning via Metropolis Hastings' interaction between humans and AI. And this is based on our preprint, which is recently available on archive. So this is the co-creative learning. It's a new theory of human-in-the-loop machine learning method. So I'll also discuss collective predictive coding hypothesis, which I talked in the previous guest swim. Maybe you can check it later. But as a background of this work. Okay, let me start. So here is an overview of our main contributions of this paper. The first, we formally propose co-creative learning. This is a novel paradigm where humans and AI collaboratively achieve symbol emergence. In other words, the formation of the external representation by mutually integrating their partial perceptual information to build shared representation. So this fundamentally shifts the perspective of human-AI alignment or teaching process, moving beyond traditional unilateral approach. The second, we introduce an experimental framework based on the Metropolis Hastings naming game. Through this framework, we propose the comparing empirical evidence, demonstrating that human-AI pairs can effectively engage in this co-creative learning process. So the visual of this slide illustrates the idea of mutual interaction between the human and AI centering around the shared world. Okay, so let's consider why co-creative learning is so crucial currently in this phase. So the first one is limitations of unilateral alignment or teaching. Traditionally, unilateral alignment framework or teaching, typically supervised learning based on ground truth levels provided by human annotators, have significant limitations. For instance, AI systems already excel beyond average human performance in many areas. the possessing distinct and often superior domain knowledge. Furthermore, the human and AI inherently have heterogeneous knowledge. Our ways of understanding and processing information are fundamentally different. So when, and also when embodied AI agents start exploring the world autonomously, it becomes questionable whether we should enforce our existing human knowledge and values as a sole ground truth. So we should respect, in some sense, we should respect AI's perception and or knowledge to bring our knowledge better status. Okay, another thing is the bidirectional, the influence. So actually, the large language models, for example, actively impact the human language use and decision making, as shown by recent studies. This, the mutual influence clearly highlights the limitation of the perspective of purely one-side alignment strategy, teaching strategy. Therefore, looking beyond the current generative AI error, it is important that we establish a new system theory, the principle, and the learning method to achieve true human AI symbiosis, the symbiosis, considering bidirectional alignment or teaching. That can be called co-creative learning process. So, to position co-creative learning more intuitively, from the machine learning perspective, let's briefly and roughly review existing traditional unilateral machine learning frameworks. So, in the most common, the insupervised or the reinforcement learning, the human's perspective, the human's perspective can be written like that. The XN human is the human's observation, and SN is the kind of cross-label or description of the human's perception of…