Neuroscience and machine learning have long drawn inspiration from each other, and a new study deepens this connection by showing that the brain, like modern AI models, benefits from unsupervised learning.
In this research, scientists recorded activity from up to 90,000 neurons in the visual cortex of mice as they were exposed to visual stimuli. Surprisingly, neural changes associated with learning were present even when the mice received no instruction or reward. This suggests that unsupervised exposure alone can reshape neural circuits and prepare the brain for future learning.
The medial higher visual areas (HVAs) were particularly active in this unsupervised learning, developing representations of the stimuli that resembled those formed during actual task performance. Meanwhile, the anterior HVAs showed signals tied to reward prediction, indicating that these regions may kick in only during supervised or reinforcement learning.
Importantly, the study also showed that this unsupervised pretraining accelerated later supervised learning in behavior tests, echoing how pretraining in AI systems or image classifiers boosts performance on specific tasks.
The findings challenge the common assumption that neural plasticity in the sensory cortex is driven solely by task learning. Instead, the brain seems to form useful internal models of the world simply through observation, much like how self-supervised learning enables machines to learn representations without explicit labels.
As researchers continue to explore the biological basis of learning, this study opens exciting paths, like linking unsupervised brain plasticity to synaptic mechanisms, and drawing parallels between cortical learning and artificial self-supervised methods. It’s a strong reminder: sometimes, just watching the world can be a powerful teacher.