The previous chapters have described various learning algorithms for pattern recognition. This chapter introduces the framework of expected loss minimization and organizes the learning algorithms described in the previous chapters in a more general and unified perspective. The square error, zero-one loss, and continuous loss are explained as loss functions. In the framework of expected loss minimization, the so-called least squares method is derived in the case of square error loss. In the case of zero-one loss, a decision function is derived from achieving the minimum error probability (Bayes error). In the case of continuous loss, a classifier with better generalization performance is obtained. Moreover, we explain the probabilistic descent method, which is a general learning method for achieving expected loss minimization. The basic idea and validity of the probabilistic descent method are summarized in the Robbins–Monro algorithm. In this chapter, the above will be explained in more detail. Sect. 14.1 describes the stochastic descent method, which is a learning method for achieving expected loss minimization. The discussion in this chapter is a preparation for clarifying the interrelationships among the learning algorithms described so far and their relationship with the Bayes decision rule in the next chapter.