Feature Learning With Complicated Environment

The problem of extracting features from given input data is of critical importance for the successful application of machine learning[8].Feature learning seeks an optimal transformation from input data into a(typically real-valued)feature vector that can be used as an input for a learning algorithm.Classic feature learning methods include dictionary learning,dimensionality reduction,manifold learning,distance metric learning,neural network and so on.Distance metric learning,as a subfield of feature learning,can learn a task-specific distance metric,thus improving the performance of many similarity/dissimilarity based methods,such as kNN,which has gained much attention in recent years.Meanwhile,development in machine learning has encouraged governments and companies to deploy them in automated systems,such as autonomous vehicles and automatic trading systems,which requires algorithms to ensure robust for both the knowns and unknowns conditions in complicated environments[9][10].In this paper,we consider the complexities in reality,and conduct the following research on distance metric learning in different situations:1.In complicated environment,possible noises or disturbances on instances will make change on their relationships,so as to affect the metric to learn.In this paper,we learn metric and noises of instances jointly and obtain a more robust metric.2.Linkages between objects are diverse and can be derived from multiple perspectives.This paper proposes a framework which considers multiple metric relationships between samples.A combination operator is introduced to flexibly represent both spatial and semantic linkages.Besides,the framework can degrade to many other metric learning methods.3.In resource constrained situations,this paper considers combining metric learning and feature selection.Feature selection can filter those task-irrelevant ones,thus reducing computational cost and effectively improving the algorithm's efficiency.In experiments,we validate this method's priority from two aspects:feature selection and metric learning.