Discriminative and Generative Learning with Style Information

Conventional machine learning approaches usually assume that the patterns follow the identical and independent distribution (i.i.d.). However, in many empirical cases, such condition might be violated when data are equipped with diverse and inconsistent style information. The effectiveness of those traditional predictors may be limited due to the violation of the i.i.d. assumption brought by the existence of the style inconsistency. In this thesis, we investigate how the style information can be appropriately utilized for further lifting up the performance of machine learning models. It is fulfilled by not only introducing the style information into some state-of-the-art models, some new architectures, frameworks are also designed and implemented with specific purposes to make proper use of the style information. The main work is listed as the following summaries: First, the idea of the style averaging is initially introduced by an example of an image process based sunglasses recovery algorithm to perform robust one-shot facial expression recognition task. It is named as Style Elimination Transformation (SET). By recovering the pixels corrupted by the dark colors of the sunglasses brought by the proposed algorithm, the classification performance is promoted on several state-of-the-art machine learning classifiers even in a one-shot training setting. Then the investigation of the style normalization and style neutralization is investigated with both discriminative and generative machine learning approaches respectively. In discriminative learning models with style information, the style normalization transformation (SNT) is integrated into the support vector machines (SVM) for both classification and regression, named as the field support vector classification (F-SVC) and field support vector regression (F-SVR) respectively. The SNT can be represented with the nonlinearity by mapping the sufficiently complicated style information to the high-dimensional reproducing kernel Hilbert space. The learned SNT would normalize the inconsistent style information, producing i.i.d. examples, on which the SVM will be applied. Furthermore, a self-training based transductive framework will be introduced to incorporate with the unseen styles during training. The transductive SNT (T-SNT) is learned by transferring the trained styles to the unknown ones. Besides, in generative learning with style information, the style neutralization generative adversarial classifier (SN-GAC) is investigated to incorporate with the style information when performing the classification. As a neural network based framework, the SN-GAC enables the nonlinear mapping due to the nature of the nonlinearity of the neural network transformation with the generative manner. As a generalized and novel classification framework, it is capable of synthesizing style-neutralized high-quality humanunderstandable patterns given any style-inconsistent ones. Being learned with the adversarial training strategy in the first step, the final classification performance will be further promoted by fine-tuning the classifier when those style-neutralized examples can be well generated. Finally, the reversed task of the upon-mentioned style neutralization in the SN-GAC model, namely, the generation of arbitrary-style patterns, is also investigated in this thesis. By introducing the W-Net, a deep architecture upgraded from the famous U-Net model for image-to-image translation tasks, the few-shot (even the one-shot) arbitrary-style Chinese character generation task will be fulfilled. Same as the SN-GAC model, the W-Net is also trained with the adversarial training strategy proposed by the generative adversarial network. Such W-Net architecture is capable of generating any Chinese characters with the similar style as those given a few, or even one single, stylized examples. For all the proposed algorithms, frameworks, and models mentioned above for both the prediction and generation tasks, the inconsistent style information is taken into appropriate consideration. Inconsistent sunglasses information is eliminated by an image processing based sunglasses recovery algorithm in the SET, producing style-consistent patterns. The facial expression recognition is performed based on those transformed i.i.d. examples. The SNT is integrated into the SVM model, normalizing the inconsistent style information nonlinearly with the kernelized mapping. The T-SNT further enables the field prediction on those unseen styles during training. In the SN-GAC model, the style neutralization is performed by the neural network based upgraded U-Net architecture. Trained with separated steps with the adversarial optimization strategy included, it produces the high-quality style-neutralized i.i.d. patterns. The following classification is learned to produce superior performance with no additional computation involved. The W-Net architecture enables the free manipulation of the style data generation task with only a few, or even one single, style reference(s) available. It makes the Few-shot, or even the One-shot, Chinese Character Generation with the Arbitrary-style information task to be realized. Such appealing property is hardly seen in the literature