International Conference Ignites Passion for Exploration in Professional Fields
In 2024, students from the College of Computer Science actively engaged in several prestigious international conferences, such as the International Conference on Machine Learning (ICML), SIGGRAPH 2024, and the Computer Vision and Pattern Recognition Conference (CVPR). These events offered students invaluable opportunities to present their research, connect with top scholars globally, and fuel their enthusiasm for exploration in their professional fields.
At these conferences, students showcased groundbreaking research on advanced topics such as machine learning, reinforcement learning, computer vision, and virtual reality. Engaging in face-to-face, in-depth discussions with leading experts, they gained valuable insights and constructive feedback from diverse academic fields. These experiences deepened their knowledge and enriched their understanding of research, helping them develop a more strategic and holistic perspective. The students below share their key takeaways and reflections from these international conferences, illustrating how this transformative experience has expanded their research horizons and fueled their future academic growth.
Title: Accelerated Policy Gradient: On the Convergence Rates of the Nesterov Momentum for Reinforcement Learning
Authors: Yen-Ju Chen, Nai-Chieh Huang, Ching-Pei Lee, Ping-Chun Hsieh
Advisor: Professor Ping-Chun Hsieh
International Conference: International Conference on Machine Learning (ICML 2024)
The Significance of the Conference:
The International Conference on Machine Learning (ICML) is one of the most influential academic conferences in the field of artificial intelligence. ICML brings together researchers and experts from around the world to showcase the latest theories, techniques, and applications in machine learning, covering subfields such as deep learning, reinforcement learning, and natural language processing. It provides a platform for collaboration and exchange between academia and industry. Research presented at ICML often represents the forefront of AI development, playing a crucial role in driving technological breakthroughs, innovative applications, and shaping the future direction of the field. Following a meticulous review process, 2,609 papers were deemed worthy of acceptance, resulting in an overall acceptance rate of 27.5%.
The Experience of Nai-Chieh Huang:
I want to express my sincere gratitude to Professor Ping-Chun Hsieh for his invaluable guidance. This paper tackles an intriguing theoretical question in reinforcement learning: Can Nesterov momentum accelerate policy gradient (PG)? Our results provide a definitive affirmative answer. We found that the objective function exhibits near-convexity around the optimal policy, a highly desirable property in optimization. This insight enabled us to demonstrate that Nesterov momentum can significantly accelerate PG. We are honored that ICML accepted our work. Participating in the conference was a great and rewarding experience—it allowed me to explore a wide range of cutting-edge research and provide opportunities to engage in meaningful, face-to-face discussions with leading experts across various fields. It was truly a highly enriching experience!
Title: Enhancing Value Function Estimation through First-Order State-Action Dynamics in Offline Reinforcement Learning
Authors: Yun-Hsuan Lien, Ping-Chun Hsieh, Tzu-Mao Li, Yu-Shuen Wang
Advisor: Professor Yu-Shuen Wang and Professor Ping-Chun Hsieh
International Conference: International Conference on Machine Learning, ICML
The Significance of the Conference:
ICML is a top-tier artificial intelligence conference. For ICML 2024, a total of 9653 submissions were received, of which 2609 were accepted, yielding an acceptance rate of approximately 27.03%.
The Experience of Yun-Hsuan Lien:
The paper presented at the 2024 ICML conference addressed a critical issue in offline reinforcement learning: the estimation of the value function. It innovatively integrated continuous-time and discrete-time reinforcement learning methods using the Hamilton-Jacobi-Bellman (HJB) equation and first-order consistency to enhance value function estimation, significantly improving model performance. Through this research presentation, we had the opportunity to discuss with many researchers at the conference. After returning to Taiwan, we will continue the discussions from the conference and initiate new international collaboration projects, further advancing our research in the field of reinforcement learning.
Title: BoostMVSNeRFs: Boosting MVS-based NeRFs to Generalizable View Synthesis in Large-scale Scenes
Authors: Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, Yu-Lun Liu
Advisor: Professor Yu-Lun Liu
International Conference: Special Interest Group on Computer Graphics and Interactive Techniques, (SIGGRAPH 2024)
The Significance of the Conference:
SIGGRAPH is a premier international conference in computer graphics and interactive techniques. It explores cutting-edge topics such as computer graphics, virtual reality, animation, visual effects, and 3D modeling. As a crucial platform for bringing together leading researchers, artists, and engineers from around the globe, SIGGRAPH drives innovation in graphics technologies through academic exchanges, technical demonstrations, and creative competitions. Additionally, it plays a pivotal role in advancing the commercialization of these technologies across diverse industries, including entertainment, design, healthcare, and education.
The Experience of Chih-Hai Su:
I want to thank Professor Yu-Lun Liu for his insightful guidance, my classmates for their collaborative efforts, and my girlfriend for her understanding and support during my intensive research period. It is a great honor to have had the opportunity to submit and present at SIGGRAPH during my time at university. Our research focused on improving 3D scene reconstruction using Neural Radiance Fields (NeRFs), and we were fortunate to have our work accepted for an oral presentation at the conference. This experience has granted me access to valuable academic resources, the latest laboratory facilities, and opportunities to connect with scholars from around the globe, thus broadening my professional network. I hope this presentation marks the beginning of my journey, and I look forward to returning to the international stage to make further contributions to the academic community.
Title: MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes
Authors: Bor-Shiun Wang, Chien-Yi Wang, Wei-Chen Chiu
Advisors: Professors Wei-Chen Chiu and Chien-Yi Wang
International Conference: IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), 2024
The Significance of the Conference:
CVPR (Computer Vision and Pattern Recognition) is one of the most influential international conferences in the field of computer vision, highly regarded by both academia and industry. It serves as a key platform for top researchers to showcase their latest findings, while also driving the development of critical technologies such as image classification, object detection, and deep learning. This year, CVPR received 11,532 submissions and accepted only 2,719 papers, resulting in an acceptance rate of just 23.6%, making it one of the most prestigious conferences in the computer vision domain.
The Experience of Bor-Shiun Wang:
I am truly honored that my research has been accepted by CVPR, marking a significant milestone in my academic journey. First and foremost, I want to express my deep gratitude to Professor Wei-Chen Chiu and co-advisor Chien-Yi Wang, whose dedicated guidance and support have given me the opportunity to present on such an international stage. During the conference, I had the privilege of engaging with cutting-edge research from various fields, and both the keynote speeches and specialized workshops provided invaluable opportunities for intellectual stimulation. What brought me the most satisfaction was successfully presenting my research and engaging in in-depth discussions with scholars, receiving many insightful suggestions and feedback.