最新Machine Learning for Sequential Data A Review.pdfVIP

Machine Learning for Sequential Data: A Review Thomas G. Dietterich Oregon State University, Corvallis, Oregon, USA, tgd@, WWW home page: /~tgd Abstract. Statistical learning problems in many fields involve sequen- tial data. This paper formalizes the principal learning tasks and describes the methods that have been developed within the machine learning re- search community for addressing these problems. These methods include sliding window methods, recurrent sliding windows, hidden Markov mod- els, conditional random fields, and graph transformer networks. The pa- per also discusses some open research issues. 1 Introduction The classical supervised learning problem is to construct a classifier that can correctly predict the classes of new objects given training examples of old objects [19]. A typical application is in optical character recognition where the objects are images of hand-written characters and the classes are the 26 alphabetic letters. The classifier takes an image as input and produces a letter as output. This task is typically formalized as follows. Let x denote an image of a hand-written character and y ∈ {A, . . . , Z } denote the corresponding letter class. A training example is a pair (x, y) consisting of an image and its associated class label. We assume that the training examples are drawn independently and identically from the joint distribution P (x, y), and we will refer to a set of N such examples as the training data. A classifier is a function h that maps from images to classes. The goal of the learning process is to find an h that correctly predicts the class y = h(x) of new images x. This is accomplished by searching some space H of possible classifiers for a classifier that gives good results on the training