|
Abstract:
|
There is growing interest in estimating predictive models from empirical data, aka inductive inference. Most statistical learning methods formalize this problem as function estimation from finite samples. That is, the goal is to estimate a function f(x) from finite labeled samples(training data) in the form (x,y). Here, a 'good' model is expected to make predictions for future inputs. This formalization of statistical inference is known as standard inductive learning. Examples include classification and regression methods. It is known that inherent difficulty of inductive inference can be alleviated by introducing additional apriori knowledge. Under standard inductive setting, this knowledge usually refers to known properties of estimated models and distributions. In many applications, apriori knowledge is available in the form of additional data (besides labeled training samples). These situations can be better formalized using non-standard learning settings. My talk will provide motivation for new types of inference under the framework of VC learning theory, and also discuss two new inductive learning settings: Universum Learning and Learning Using Privileged Information.
|
