A General Global Learning Model: MEMPM

Abstract

Traditional global learning, especially generative learning, enjoys a long and distinguished history, holding a lot of merits, e.g. a relatively simple optimization, and the flexibility in incorporating global information such as structure information and invariance, etc. However, it is widely argued that this model lacks the generality for having to assume a specific model beforehand. Assuming a specific model over data is useful in some cases. However, the assumption may not always coincide with the true data distribution in general and thus may be invalid in many circumstances. In this chapter, we propose a novel global learning model, named Minimum Error Minimax Probability Machine (MEMPM), which is directly motivated from Marshall and OlKin Probability Theory [20, 24]. For classifying data correctly, this model focuses on estimating the worse-case probability, which is not only more reliable, but also more importantly provides no need for assuming specific models. Furthermore, this new model consists of several appealing features.