In the past decade, there has been a growing interest in methods for estimating (learning) models from data. In this talk, I will first discuss and contrast three principal learning methodologies, i.e. classical statistics, statistical learning theory and data mining. My discussion will focus on conceptual and methodological issues underlying each methodology, rather than on technical aspects and math. I will also emphasize the distinction between the problem formulation and constructive learning algorithm. Unfortunately, existing statistical formulations (such as classification, regression and density estimation/clustering) may not be appropriate for many real-life applications. Therefore, I will introduce and discuss several new learning formulations. In particular, I describe multiple model formulation, where available data is generated by several (unknown) statistical models. Hence, existing learning methods (for regression and classification) based on single model formulation, are no longer applicable. I will also describe constructive learning methodology for multiple model estimation, and present empirical comparisons illustrating potential advantages of this approach.