Congratulations to Yijing Yang for passing her defense on June 15, 2022. Her PhD dissertation is titled with “Advanced Techniques for Object Classification: Methodologies and Performance Evaluation”. Her Dissertation Committee members include Jay Kuo (Chair), Justin Haldar, Suya You, and Aiichiro Nakano (Outside Member). All committee members were very pleased with the depth and fundamental nature of Yijing’s research. We are glad to invite Yijing here to share the overview of her thesis study. We wish Yijng all the best for her future career and life!

“Object classification has been studied for many years as a fundamental problem in computer vision. With the development of convolutional neural networks (CNNs) and the availability of larger scale datasets, we see a rapid success in the classification using deep learning. Although being effective, deep learning demands a high computational cost. Another challenge is the amount of accessible labeled data. How the quantity of labeled samples affects the performance of learning systems is an important question in the data-driven era. In this dissertation, we investigate and propose new techniques based on successive subspace learning (SSL) methodology to shed light on the above problems. It can be decomposed into four aspects: 1) improving the performance of SSL-based multi-class classification, 2) improving the performance of resolving confusing sets, 3) enhancing the quality of the learnt feature space by conducting a novel supervised feature selection, and 4) designing supervision-scalable learning systems.

Specifically, in the first two aspects, soft-label smoothing (SLS), hard sample mining, and a new SSL-based attention localization method are proposed to improve the classification performance. In the third part, a novel supervised feature selection methodology is proposed to enhance the learnt feature space, including the discriminant feature test (DFT) and the relevant feature test (RFT) for classification and regression tasks, respectively. As compared with other existing feature selection methods, DFT and RFT are effective in finding distinct feature subspaces. Intensive experiments show that they provide feature subspaces of significantly lower dimensions while maintaining near optimal classification/regression performance, robust to noisy input data, and computationally efficient. In the last part, the supervision-scalable learning systems are designed. Two families of modularized systems are proposed based on HOG features and SSL features, respectively. We discuss ways to adjust each module so that the design is more robust against the number of training samples. Experiments and analysis show that both systems demonstrate an excellent scalable performance with respect to various supervision degrees. They both outperform LeNet-5 by a large margin under extremely weak supervision and have performance comparable with that of LeNet-5 under the strong supervision condition.

I would like to thank Professor Kuo for his continued guidance and invaluable supervision. His passion, enthusiasm and the serious attitude to the research also motivate me throughout my Ph.D. study. I have learnt a lot from Prof. Kuo, including and beyond the research skills that I hope to carry forward throughout my future life and career. I would also like to thank all the MCL members for their kind help and support during these years. MCL is like a big family. Labmates here are always willing to help each other, giving advice to the research and daily life, having discussions and getting inspired from each other, and providing mental support. I really enjoyed my journey here at MCL. I wish all the best luck to our family members for a great and bright future.”