Recently, research completed by Zhang Zheng, an undergraduate in Foundations of Mathematical Science at DLI, was officially accepted by The Forty-Third International Conference on Machine Learning (ICML 2026), a top-tier international conference in artificial intelligence. Titled “BTSP-CAM: A Brain-Inspired Geometric Memory for Class-Incremental Learning”, the work addresses the stability–plasticity dilemma in continual learning and proposes BTSP-CAM, a brain-inspired geometric memory method. Zhang Zheng is the first author of the paper; Professor Xu Qi of the School of Computer Science and Technology and Professor Guo Qingjie of DLI supervised the research.
Continual learning is an important research direction in artificial intelligence. It aims to enable models to keep acquiring new knowledge as humans do, while retaining their ability to recognise and understand what they have already learned. Existing deep learning models, however, are prone to “catastrophic forgetting” when continuously exposed to data from new classes. To address this, the research team drew inspiration from the neural mechanism of behavioural time-scale synaptic plasticity (BTSP) in the hippocampus, introducing the ideas of random memory allocation and pattern separation into pre-trained-model-driven class-incremental learning. The resulting algorithm requires neither the storage of historical samples nor back-propagation updates to the backbone network.
The method externalises model plasticity from the pre-trained backbone into a binary synaptic memory matrix, writing new knowledge through local random bit flips, and completes rapid consolidation using a CA1-like competitive layer together with a closed-form ridge regression readout. Experimental results show that BTSP-CAM achieves competitive learning performance in the exemplar-free setting; as a lightweight plug-in, it also steadily improves the performance of a range of existing continual learning methods built on pre-trained models.
DLI has long attached importance to cultivating students’ innovative spirit and research capability. Adhering consistently to a student-centred educational philosophy, the College supports initiatives such as the “1+X Innovation Cultivation Programme”, the “Outstanding Undergraduate Basic Research Project” and the “Outstanding Undergraduate Innovation and Entrepreneurship Project”, in order to nurture top innovative talent with an international outlook, an innovative spirit and practical ability.
About the Conference:
The International Conference on Machine Learning (ICML) is one of the oldest, largest and most influential international academic conferences in machine learning and artificial intelligence, and is recommended as a Class A international conference by the China Computer Federation (CCF). Together with NeurIPS and ICLR, it is known as one of the “three major conferences” in artificial intelligence — the most selective, the highest in standard and the most influential. The 43rd International Conference on Machine Learning, ICML 2026, will be held in Seoul, Republic of Korea, in July 2026.
About the Student:
Zhang Zheng is a 2023-cohort undergraduate at DLI. He has been invited to serve as a reviewer for journals in CAS Zone 1 and for CCF-A conference papers, has won four national-level awards in national discipline competitions for university students under the Ministry of Education as a team leader, and currently leads one national-level innovation and entrepreneurship training project.
