Convergence Analysis of Recurrent Neural Network Filtering in Discrete Time
组织者
演讲者
孙泽钜
时间
2023年09月04日 15:00 至 15:30
地点
数学系理科楼A-203
摘要
In this talk, we will discuss the convergence results of recurrent neural network (RNN) filtering in discrete time. The solution to general nonlinear filtering problems, i.e., the evolution of the conditional probability density, satisfies an infinite dimensional dynamical system, which corresponds to an RNN with infinitely many neurons. In order to implement RNN filter in practice, an efficient finite dimensional approximation is required. With the concept of Kolmogorov n-width, we can obtain asymptotic properties and also the convergence rate of RNN filters.