Deep Neural Networks with Non-iterative Learning Strategies for Pattern Recognition and Data Agumentation Applicable to Computer Vision and Healthcare

点击翻译按钮获取中文摘要

The past few years have witnessed a boom of deep networks, including deep auto-encoders, deep convolutional networks, deep recurrent networks, long short-term memory, generative adversarial network (GAN), etc. These methods have recently dramatically pushed forward the state of the art in diverse domains such as language understanding, robotics, game playing, healthcare, among others. However, many application areas require machine learning algorithms that have fast recognition capability. Thus, the applicability of fast learning is exceptionally high. Unfortunately, the currently well-known deep networks use iterative training strategy. With network depth increasing, deep learners require extensive computational power and can take weeks to train, making it difficult for broad applications. Thus, the following issue serves as a motivation: Can we design effective learning algorithms for deep networks in order to improve performance and increase training speed? Furthermore, the deep network itself has unprecedented parameters where a proportional number of examples must be fed. However, the recent deep network-based data augmentation algorithms such as GAN only work for images, but tabular data such as radars, ultrasonic sensors, and biological data can be applied to wide areas. Motivated by the above problems, the goal is to design effective data augmentation algorithms for tabular data generation and re-use the augmented data for deep networks training. My five-year research plan is to design a unified non-iterative learning strategy that can be used in various deep networks, resulting in lower training computational cost and higher performance applicable to data augmentation and pattern recognition. The specific objectives are as follows: 1) Designing non-iterative learning algorithms to train deep networks, including deep convolutional neural network, deep auto-encoder, and generative adversarial network applicable to computer vision and healthcare, resulting in less training computation with better performance. 2) Designing innovative augmentation algorithms that work for tabular data, including biosignal, radar, and physiological signal. Then, reusing the created synthetic tabular data to boost the recognition performance of neural networks. I believe the realization of a complete and efficient learning system is a significant achievement from both a methodological and engineering application perspective. The developed methods will solve the major bottleneck of learning speed in deep networks, which has a striking application perspective in computer vision and healthcare domains. Furthermore, this research program will provide unique opportunities to train HQPs in topics considered highly in-demand by the Canadian AI industry.