By Veronica E. Tremblay, 2019
“The research freedom I had at KAUST allowed me to switch to studying a topic that society and humanity needs now.”
Dr. Yipeng Qin first heard of KAUST during his bachelor degree studies. A KAUST recruiter came to China with stories about an advanced research university in Saudi Arabia with large seaside villas, palm-lined avenues, and famous professors. It seemed like just good advertising, but it all turned out to be true. Qin moved to study computer graphics with a focus on geodesic computation at Bournemouth University, England. After his Ph.D., he left the U.K. thinking he might never come back.
Part of KAUST’s mission is to enhance the welfare of society. As a postdoc in the KAUST Visual Computing Center (VCC), Qin decided he needed to move away from his original research topic towards a more urgent branch of computer science. His supervisor, Professor Peter Wonka, gave Qin his full support to complete this transition. With the accessibility of KAUST’s GPU cluster resources, the collaborative spirit of Wonka’s research group, and the encouragement of his supervisor, Qin was able to take off in a new direction where results are necessary and in demand – how machines learn. With the new skillset he built as a postdoc at KAUST, he returned to the U.K. as a lecturer at Cardiff University, Wales.
“No one really understands how deep learning works, and that has legal repercussions.”
Qin’s research focuses on Generative Adversarial Networks (GANs), a promising machine-learning model for content generation. In machine learning, computers learn from training data by observation. The computer develops its own connections among the input data, processing functions, and output predictions. These connections make up the artificial neutral network: the mathematical model by which artificial intelligence (AI) makes decisions. Deep learning is simply machine learning with many (deep) levels of processing functions.
However, no one is really sure how the computer strings together the web between input and output, with deep learning often being referred to as a “black box” or “alchemy.” Researchers watch the computer develop its neural network without truly understanding how the computer draws its connections across that network. This uncertainty can have very real consequences: the safety of driverless cars is the most well-known concern.
“The modeling shows success, but why?”
By studying the training dynamics of GANs, Qin hopes to improve humanity’s understanding of AI learning and behavior. GANs can be used to generate new content – such as a face or even an image of a cat – for various applications in visual computing. When the computer generates the desired model successfully, Qin asks, “Why was the modeling successful this time?”
When the training dynamic is understood, only then can stability (guaranteed results) be engineered into the deep learning process. Currently, Qin is building on his KAUST research and studying how to guarantee modeling success despite an arbitrary choice of loss functions or noise distributions, making for a more robust and predictable AI.
Dr. Yipeng Qin was a postdoc in KAUST Visual Computing Center (VCC) Associate Director and Professor Peter Wonka’s research group from 2017 to 2019. He is now a lecturer and member of the Visual Computing research group at the School of Computer Science and Informatics, Cardiff University, Wales. His research interests include visual computing (computer graphics and computer vision), geometry processing and machine learning.
Selected publications:
1. Y Qin, X Han, H Yu, et al. “Fast and exact discrete geodesic computation based on triangle-oriented wavefront propagation.” ACM Trans. Graph., 35 (4), 125, 2016.
2. Y Qin, N Mitra and P Wonka. “Do GAN Loss Functions Really Matter?” 2018. https://arxiv.org/abs/1811.09567
3. Y Qin, H Yu and J Zhang. 2017. “Fast and Memory‐Efficient Voronoi Diagram Construction on Triangle Meshes.” Computer Graphics Forum, 36, 93-104.