The one-week-long top international academic conference in the field of artificial intelligence, NeurIPS (Conference on Neural Information Processing Systems), has been held in New Orleans, United States, from December 10 to 17, 2023. A total of 3 research papers from the research group led by Assist Prof. Liu Zuozhu of ZJUI were selected, and the results mainly focused on key technologies such as privacy protection, model fairness, and long tail learning in the field of smart healthcare.
NeurIPS (Conference on Neural Information Processing Systems) is one of the top international conferences in the field of machine learning and computational neuroscience, bringing together top researchers from academic and industrial fields, and covering the latest research results in multiple fields such as machine learning and deep learning. NeurIPS is known for its high-quality papers, invited speeches and workshops, which have a significant impact on promoting scientific and technological progress in related fields.
ZJUI doctoral students Chen Ruizhe and Xiao Zikai are the first authors of the paper, and Assist Prof. Liu Zuozhu is the corresponding author of the paper. Other authors of the paper include ZJUI Assist Prof. Yang Hao (Howard), and ZJUI doctoral students Xiong Huimin and Hu Tianxiang. The related research was supported by the National Natural Science Foundation of China, Zhejiang Provincial Natural Science Foundation, and the Zhejiang University-Angelalign Inc. R&D Center for Intelligent Healthcare.
01 Fed-GraB: Improve the intelligent diagnosis performance of complex long-tail orthodontic cases while ensuring the data privacy at the same time
Assist Prof. Liu Zuozhu's group, together with professional institutions and enterprises such as Singapore University of Technology and Design, A*STAR Singapore, Sichuan University, Angelalign, etc., published the "Fed-GraB: Federated Long-tailed Learning with Self-Adjusting Gradient Balance," which provides innovative ideas and solutions for solving the long-tail distribution problem in the field of federated learning.
Federated learning can train artificial intelligence models without sharing patient data, which provides a new solution for the realization of patient privacy protection and efficient intelligent diagnosis in the field of smart medical field. However, in the case of long-tail distribution of global data, such as oral medical data covering a number of rare malocclusion cases, it is difficult for the previous federated learning methods to achieve efficient and high-quality auxiliary diagnosis and treatment under these complex distribution conditions.
▲ Framework of FedGraB
To solve this issue, the research group came up with a method called Fed-GraB, which includes a self-regulating gradient balancer (SGB) that can efficiently perform client-side gradient reweighting while preserving privacy at the same time. Through the Direct Prior Analyzer (DPA) module, the balancer can learn with the global long tail prior without violating privacy principles. Experimental results show that the performance of Fed-GraB model achieves state-of-the-art performance on several representative data sets such as ImageNet-LT and iNaturalist, especially when processing a few categories of data, its performance greatly surpasses the current optimal algorithm in the field.
Based on protecting the patients’ privacy, Fed-GraB can improve the intelligent diagnosis quality of complex long-tail cases in smart medicine. The implementation of Fed GraB is expected to boost telemedicine service innovation, promote the security of data sharing, improve the efficiency and quality of medical services, and has significant potential to develop personalized and precise treatment solutions, which can further enhance the patient experience in the healthcare industry while ensuring data privacy.
Article Link：https://arxiv.org/pdf/2310.07587.pdf
02 Fast Model Debias: achieve high-quality "debias" effect and promote the fairness and accuracy of medical treatment
Assist Prof. Liu Zuozhu's group, in collaboration with Nanyang Technological University, Stanford University, A*STAR Singapore, Angelaligns and other professional universities and enterprises, published a paper entitled "Fast Model Debias with Machine Unlearning" at the NeurIPS 2023 conference, which has made new progress in the problem of debias in deep learning models.
In the smart medical field, model debias of deep learning model is of great significance:
1) Debias can significantly improve the accuracy of model diagnosis in different populations;
2) Debias also helps to reduce misdiagnosis or missed diagnosis, thereby improving the overall quality of healthcare services when processing medical data from minority groups. When patients recognize that the health care system spare no efforts to provide fair and unbiased care, patients’ trust in the health care system increases.
▲ Pipeline of proposed FMD
The research group proposes FMD, an all-inclusive framework for fast model debiasing. the FMD comprises three distinct steps: bias identification, biased-effect evaluation, and bias removal. The FMD construct a dataset comprising factual samples along with their corresponding counterfactual samples. By observing how the model’s predictions change with the attribute variations, any bias present can be effectively identified. The research extends the theory of influence functions, employing it to estimate the impact of perturbing a biased attribute within the training data on the model’s prediction bias measurement. Finally, they introduce an unlearning mechanism to remove the learned biased correlation, and further design an alternative strategy to unlearn biases with the counterfactual external dataset, avoiding hard requirements on access to the training data which might be unavailable in practice. Compared with the existing work, FMD does not need to conduct supervised retraining of the entire model or additional labeling of bias attributes, which reduces the need for large-scale labeling of training data in practical scenarios, and only needs to update a few parameters to achieve model debias. Experiments show that FMD can improve fairness on a variety of bias indicators at a small cost.
FMD successfully achieved unbiased deep learning models that have already been trained while protecting patient privacy (without the need to access training data). In the field of smart healthcare, the application of FMD is expected to help improve the fairness and accuracy of diagnosis and treatment, reduce bias and misdiagnosis, and bring patients a more reliable, effective, and fair medical experience.
Artilce Link：https://arxiv.org/pdf/2310.12560.pdf
03 BaCon: Solve the challenge of data imbalance and open distribution, and effectively apply algorithms to real-world scenarios
Assist Prof. Liu Zuozhu and his group, together with the Hong Kong University of Science and Technology, Angelalign and other professional universities and enterprises, published a research paper entitled "Towards Distribution-Agnostic Generalized Category Disco very" at the NeurIPS 2023 conference, providing a new solution to the challenges brought by the two inherent characteristics of data imbalance and open distribution in the actual visual world.
Data imbalance and open-ended distribution are two intrinsic characteristics of the real visual world. Though encouraging progress has been made in tackling each challenge separately, few works dedicated to combining them towards real-world scenarios. In this paper, the group formally define a more realistic task as distribution-agnostic generalized category discovery m(DA-GCD): generating fine-grained predictions for both close- and open-set classes in a long-tailed open-world setting.
▲ Illustration of distribution-agnostic generalized category discovery (DA-GCD)
The research group propose a Self-Balanced Co-Advice contrastive framework (BaCon), which consists of a contrastive-learning branch and a pseudo-labeling branch, working collaboratively to provide interactive supervision to resolve the DA-GCD task. In particular, the contrastive-learning branch provides reliable distribution estimation to regularize the predictions of the pseudo-labeling branch, which in turn guides contrastive learning through self-balanced knowledge transfer and a proposed novel contrastive loss. Experiments show that the effectiveness of BaCon is demonstrated with superior performance over all the existing baseline methods and comprehensive analysis across various datasets.
This study addresses the challenge posed by the two intrinsic characteristics of data imbalance and open-ended distribution in the real visual world and provides a more reliable and accurate solution for real-world applications such as uneven data distribution, lack of labels, and open sample diagnostics in smart medical field.
Article Link：https://arxiv.org/abs/2310.01376
Author Introduction
Chen Ruizhe is a 2021 graduate and 2023 doctoral student of ZJUI. His supervisor is Assist Prof. Liu Zuozhu, and his research focuses on trustworthy artificial intelligence and medical image processing. He has published multiple papers at conferences such as NeuroIPS, AAAI, and MIDL.
Xiao Zikai is a 2022 doctoral student at ZJUI. His supervisor is Assist Prof. Liu Zuozhu, and his research focuses on federated learning and privacy computing.
Dr. Yang Hao (Howard) is an Assistant Professor at ZJUI. His main research interests are the fundamental theory and system design of wireless communication networks. He has published over 80 papers in authoritative journals in the field of wireless communication networks, including IEEE JSAC, TWC, TCOM, TSP, and international conference papers.
His research work has received the IEEE Signal Processing Society 2022 Best Paper Award, and the IEEE ComSoc Asia-Pacific Outstanding Young Researcher Award. He is currently the editorial board member of IEEE Transactions on Wireless Communications, a top international journal in wireless communication.
Assist Prof. Liu Zuozhu is an Assistant Professor and PhD supervisor at ZJUI, and an Adjunct Professor at the Stomatology Hospital, Zhejiang University School of Medicine (Zhejiang Stomatology Hospital). He hosted multiple projects such as projects funded by the National Natural Science Foundation, Zhejiang Provincial Key Projects, and Zhejiang University Major Industrial Projects. And his research interests mainly include trustworthy artificial intelligence in medicine. Assist Prof. Liu published multiple papers in flagship journals and top-level conferences such as Patterns, IEEE Trans, ICLR, ACL, JDR, etc.
