One of the most intriguing discoveries in modern neuroscience is that the brain's internal representations are not always stable. Even when an animal or person continues performing a task successfully, the underlying neural activity patterns can slowly change over time, a phenomenon known as representational drift. Scientists have struggled to understand why these changes occur and what they reveal about learning and memory.
In this recent study, Harvard researcher Farhad Pashakhanloo investigated whether information that a brain or artificial intelligence system learns to ignore could nevertheless influence how neural representations evolve. Using mathematical theory and computational models, he discovered that so-called task-irrelevant stimuli can act as a hidden source of learning noise. Although these inputs do not contribute to the task being performed, they can gradually reshape the neural representations of information that does matter.
The research demonstrated this effect across multiple learning systems, including biologically inspired Hebbian learning models (in which simultaneous activation of brain cells leads to pronounced increases in synaptic strength between those cells) and modern neural networks trained with stochastic gradient descent. The results showed that greater variability and complexity in irrelevant information produce faster representational drift.
These findings provide a unifying explanation for a phenomenon observed in both brains and artificial intelligence systems. Rather than viewing representational drift as random instability, researchers may be able to use it as a window into the hidden computations that drive learning. Harvard Research Computing resources enabled the large-scale simulations, numerical experiments, and theoretical validation required to test these ideas across multiple high-dimensional models and datasets.
Pashakhanloo, F. (2025), Contribution of Task-Irrelevant Stimuli to Drift of Neural Representations, Advances in Neural Information Processing Systems (NeurIPS 2025), arXiv:2510.21588 [q-bio.NC]
