Multi-task Multi-view Feature Learning And Its Applications

As the one of the classical machine learning problem,feature learning has been widely studied.Traditional machine learning process is based on the two stages: feature extraction and model training,while feature learning alleviate mismatch between the weakness of hand-design feature representation and tough inductive bias(usually recommend that the data is linearly almost separable)proposed by the classifier.Furthermore,when we deal with the practical problem,it is a common but sufficient strategy that training multiple related tasks simultaneously and choose the multiple complemented feature representation.So we propose two framework to solve the multi-task multi-view feature learning problem:(1)linearly weighted the feature vector with regularization constraint,simultaneously selecting feature and training model;(2)unsupervised learning a sub-space representation with space transforming,expecting to receive the most discriminative representation.By contrast,the first method has an elegant expression and cost less time on training stage,even is more robust.The second one is more complex and sufficient but slower too.There should be something that could improve the shortcoming in these methods.Deep learning method has an impressive impact to the traditional machine learning method in the real big data.Comparing with the hand-design method,the data-driven strategy could catch more intrinsic characteristics in the data.Then the explicit feature selection method is not applicable because of the extra process.But the embedded manner could apply the regularization to the fully connected layer in the neural network.From this point,our proposed method has a wildly application scope to other model learning.