Stability Analysis of Recurrent Neural Network Spectral Methods
组织者
演讲者
孙泽钜
时间
2023年10月20日 20:00 至 20:30
地点
Online
摘要
In this talk, we will discuss the long-term convergence results when dealing with stable nonlinear filtering problems using recurrent neural network. Our aim is to show that if the real solution to the filtering problem satisfies the ‘limited memory’ property, i.e., the effect of the observations obtained long time before on the solution of the filtering decay exponentially, then the recurrent neural network we construct can tracking the conditional expectation well in the long run. We will also focus on the difficulties we may encounter when deriving this result.