Ying Wu College of Computing Researchers Win Awards at Top AI Conference
Research on trustworthy AI and machine learning earned a Best Paper award for Ying Wu College of Computing assistant professor Hai Phan at the prominent Association for the Advancement of Artificial Intelligence annual conference in Washington last month.
Phan's work, XRand: Differentially Private Defense Against Explanation-Guided Attacks, aims to ensure explainability, fairness, privacy and robustness in artificial intelligence algorithms. He is teaching this subject in the college's new M.S. in Artificial Intelligence program, which builds on the existing graduate certificate in AI.
"Current machine learning models are ‘fair’ at best, and AI will not be completely reliable without a viable solution,” Phan said. He explained that most existing machine learning models have been able incorporate a combination of explainability, privacy and robustness, but the addition of fairness involves complex correlations that are significantly more challenging. Exploring this unproven territory is what earned his paper the honor of being one of only 12 selected out of 8,777 submissions.
“AI has the power to create a better society – as long as it is safe, ethical and trustworthy. My work is a step towards making sure this is achieved, and recognizing my paper as such is a great honor. Fully comprehending how and why this is important is what students can expect in my class,” he said.
Five other NJIT computing faculty also presented at the conference, which is the most prestigious in its field: Cristian Borcea (Complement Sparsification: Low-Overhead Model Pruning for Federated Learning); Jing Li, Hua Wei and Guiling Wang, (Safelight: A Reinforcement Learning Method Toward Collision-free Traffic Signal Control); and Pan Xu (Equity Promotion in Public Transportation). All six research papers support the objective of using artificial intelligence to solve real-world problems, with an emphasis on security and safety.
Borcea, an expert on mobile computing and sensing, focused his most recent research on making federated learning practical on resource-constrained mobile and internet-of-things devices. Federated learning is a collaborative deep learning paradigm that preserves user privacy. However, because the traditional federated learning neural networks are dense, they require significant memory, computation and communication overhead. Borcea’s solution reduces the size of these networks through a novel distributed algorithm, such that they result in less overhead and battery power consumption on mobile and internet-of-things devices, while achieving good prediction accuracy and privacy protection.
Improving traffic safety through signal control is the subject of the collaborative paper by Li, Wang and Wei. Statistics demonstrate that roughly one-quarter of road accidents in the U.S. happen at intersections due to problematic signal timing. The research team formulated a solution for a safety-enhanced residual reinforcement learning method, called Safelight, which employs multiple novel optimization techniques to achieve both efficiency and safety. Extensive experiments were conducted using both synthetic and real-world benchmark datasets. Results showed that their method can significantly reduce collisions while increasing traffic mobility.
Initial research by Wei was done in part through a grant from the National Science Foundation. Also playing a key role in this work were Ph.D. student Wenlu Du, who presented the paper at the conference, and colleagues Jingyi Gu, Junyi Ye, and alumna Xiaoyuan Liang. Liang was among the earliest Ph.D. graduates in deep learning. Wang founded the new M.S. degree program.
Xu’s research concerns social inequities in public transportation for underserved communities. His proposal optimizes a model integrating increased transport infrastructure, including additional bus lines and ride-hailing services to connect needy residents to railway systems. The solution would allocate a given limited budget to different candidate programs such that the overall social equity is maximized, which is defined as the minimum covering ratio among all pre-specified protected groups of households based on factors such as income and race. This is done through the design of a rounding algorithm for approximating the solution. The algorithm was tested against baselines on real public data sets collected in Chicago.
Learn more about the new graduate programs in artificial intelligence offered at our Newark and Jersey City locations:
