About Me

I am Bojian Yin, an Associate Professor at the Institute of Automation, Chinese Academy of Sciences. My research lies at the intersection of deep learning, brain-inspired intelligence, and foundational AI. I study the mathematical mechanisms that make intelligent learning efficient, adaptive, and robust.

At its core, I believe the problem of AI is a problem of learning. Intelligence is not a fixed artifact produced by one enormous training run; it is the capacity to keep acquiring, revising, and reusing knowledge. Yet today's models still learn slowly, forget what they have seen, and demand enormous resources. I search for the mathematical principles that could let machines learn the way brains do, continually, efficiently, and robustly, and I turn those principles into practical algorithms and systems. This pursuit has led to two first-author papers in Nature Machine Intelligence. But this is only a beginning, and the research continues.

Research Interests

My research focuses on the mathematical and brain-inspired principles of learning, and on turning them into robust, generalizable algorithms for large language models and agentic systems. My work spans several interconnected areas:

• Learning Algorithms Beyond Backpropagation. I develop local, online, and forward-mode learning rules that let large models train and adapt efficiently, including continual and test-time learning in dynamic environments. This line includes Forward Propagation Through Time and Stochastic Variational Propagation as scalable, biologically grounded alternatives to backpropagation.
• Memory and Continual Inference. I study persistent and associative memory together with reset-free continual inference, enabling long-context LLMs and long-horizon agents to keep learning and reasoning over time without forgetting or restarting. See Never Reset Again for a mathematical framework for continual inference in recurrent models.
• Efficient Sequence Modeling and Reasoning. I design efficient recurrent and state-space architectures for long-range sequence modeling, and adaptive inference-time computation that scales with problem difficulty. Recent work includes sparse selective-update RNNs for long-range modeling.
• Mathematical Foundations of Learning and Generalization. A core theme across my research is understanding, mathematically, when and why these mechanisms learn and generalize, work that began in spiking and recurrent networks, including two papers in Nature Machine Intelligence, and now extends to generative and agentic AI.

Feel free to reach out if you're interested in collaborating or just chatting about learning algorithms, brain-inspired AI, or LLMs and agents.

Publications

1. Accurate online training of dynamical spiking neural networks through Forward Propagation Through Time. B. Yin, F. Corradi, S. M. Bohte. Nature Machine Intelligence, 2023.
2. Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. B. Yin, F. Corradi, S. M. Bohte. Nature Machine Intelligence, 2021.
3. Efficient Sparse Selective-Update RNNs for Long-Range Sequence Modeling. B. Yin, F. Corradi. arXiv preprint, under review, 2026.
4. Stochastic Variational Propagation: Local, Scalable and Efficient Alternative to Backpropagation. B. Yin, F. Corradi. arXiv preprint, under review, 2025.
5. Using the structure of genome data in the design of deep neural networks for predicting amyotrophic lateral sclerosis from genotype. B. Yin, M. Balvert, R. A. van der Spek, B. E. Dutilh, S. M. Bohte, J. Veldink, A. Schonhuth. Bioinformatics, 2019.

Miscs

I have worked across academia, national research institutes, and neuromorphic hardware startups, with experience in algorithm design, mathematical modeling, FPGA and embedded deployment, SDK optimization, and interdisciplinary collaboration.

CV

Download CV.