-演講時間: 115年5月13日（三） 14:00~16:00

-演講地點: E6-A203教室

-演講者: 鮑興國 教授 (台灣科技大學資訊工程學系)

-演講主題: A Few Attempts to Explainable AI and their Applications

-演講摘要:

The deep learning is notorious for its black-box nature. Given the superior results from deep learning on various applications, we are still concerned with its unexpected outcomes once a while. The reason is mainly due to the uncontrollable nature of deep learning prediction. Specifically, we do not fully understand what the value means on each network neuron or layer. In this talk, we show a few of our attempts from the last few years on the topic of explainable AI and how we can understand deep network modeling as well as their predictive results. The central concept that we discuss starts from the Information Bottleneck (IB) method, which was pioneered by Naftali Tishby and his followers. Based on that, we explain how the deep learning was trained, and how neurons and layers own the representation power to summarize the input data, then can be used for downstream predictive tasks. After that, we emphasize how the IB method can be realized through the mutual information computation, where we use it to find the relationship between different parts of the deep networks. Based on the aforementioned computation, we can use it to solve various real-world problems such as noise or watermark removal, deep network pruning, and the pruning of LLMs, to name a few. Some may also be extended to IoT applications where we ask for low-power computation on edge devices.

Keywords: Explainable AI, deep learning, mutual information, model pruning, LLM.

-講者簡歷： Hsing-Kuo Pao (Kenneth) received the bachelor degree in mathematics from National Taiwan University, and M.S. and Ph.D. degrees in computer science from New York University. From 2001 to 2003, he was a post-doctorate research fellow in the University of Delaware, and later he joined in Vita Genomics as a research scientist. In 2003, he joined the department of computer science and information engineering in National Taiwan University of Science and Technology, and now a professor in the department. Prof. Pao has rich experience on bringing theoretical approaches to industry and have built solid foundation on finishing the last-mile efforts for real-world applications given the AI and cybersecurity background. His overall research topics include machine learning methodology and its applications to IoT analytics, information security, computer vision, and text mining.