 |
Center for Nonlinear Studies |
A basic and often implicit underlying assumption in designing traditional statistical techniques is the assumption of having independent and identically distributed (i.i.d.) training data. However, in many applications such a basic assumption is violated and, as a result, the performance of statistical techniques changes. Examples of such non-i.i.d. data arise in many applications including biology, climate research, and remote sensing.
In these situations considering a stochastic setting in which the data is generated from an underlying stochastic process (a collection of possibly dependent random variables) is more relevant. The first part of the talk will discuss the application of stochastic processes in analyzing the performance of traditional predictive techniques in situations where the data are non-i.i.d.
In the second part, we will discuss an application of the stochasticity assumption in synthesizing new statistical techniques. In these techniques we incorporate the system’s prior knowledge in the form of stochastic processes, e.g. the solution of stochastic differential equations, in classification of stochastic observations.
Host: Turab Lookman