能動的推論の量子拡張とは？密度行列とイベント駆動で読み解く次世代AIF理論

https://research.smeai.org/active-inf... Active Inference Framework (AIF) is a framework that has been widely applied in fields ranging from neuroscience to artificial intelligence, as a theory that unifies perception, learning, and behavioral selection as minimizing free energy. Traditional AIF is based on Bayesian updating on probability vectors, and theoretical development has already progressed in both discrete-state space and continuous-time versions. On the other hand, in situations dealing with "sparse and event-centric" data such as neural spikes and asynchronous sensor logs, properties that are difficult to express using classical probabilistic representations alone, such as the order dependence of observations and the superposition-like ambiguity of latent representations, are beginning to be recognized as challenges. Against this backdrop, in recent years, attempts to connect AIF with the mathematical language of quantum mechanics, namely density matrices and quantum measurement theory, have been explored. This article will introduce this developing theme of "quantum extension of active inference," focusing on theoretical correspondences.