Unsupervised Pretraining in Biological Neural Networks

On August 24, 2025, at 9:00 AM Beijing Time, Biolà hosted the 15th issue of the Bioneer First-Author Forum, featuring Dr. Lin Zhong, a research scientist at the Janelia Research Campus of the Howard Hughes Medical Institute (HHMI). The session was delivered in English. As the first author of the featured study, Dr. Zhong presented “Unsupervised Pretraining in Biological Neural Networks,” published in Nature in 2025.

A central question in perceptual learning is whether neural plasticity is driven primarily by supervised, task-dependent learning or can emerge through unsupervised experience alone. Combining large-scale neural recordings with tightly controlled virtual-reality experiments in mice, Zhong and colleagues recorded activity from tens of thousands of neurons in visual cortex and showed that the majority of experience-dependent neural plasticity occurred without explicit task supervision.

The study further demonstrated that such unsupervised pretraining substantially accelerated subsequent task-specific learning. These findings suggest that, much like pretraining in artificial neural networks, biological neural systems may first acquire useful representations through passive or task-independent experience and then leverage these representations when learning specific behavioral objectives. The work provides a framework for linking principles of biological learning with contemporary ideas in machine learning.

We sincerely thank Dr. Lin Zhong for sharing the discoveries and scientific thinking behind this work with the Biolà community, and for discussing how large-scale neural recordings can reveal the learning algorithms implemented by biological neural networks.

Citation: Zhong, L., Baptista, S., Gattoni, R., Arnold, J., Flickinger, D., Stringer, C., & Pachitariu, M. (2025). Unsupervised pretraining in biological neural networks. Nature, 644(8077), 741–748. https://doi.org/10.1038/s41586-025-09180-y