Functional Stability Amidst Neural Representational Drift

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 279 papers112 conceptsfrozen 2026-10-04

In classical neuroscience, we often imagine the brain as a hardwired machine: once you learn to ride a bike, a specific set of "bike-riding" neurons fires in a fixed pattern every time you hop on. However, long-term recordings have revealed a startling paradox. While your ability to ride the bike remains perfectly stable, the neural patterns representing that skill are in a state of constant flux.

This phenomenon, known as representational drift, describes the continuous shifting of neural activity patterns over days and weeks despite unchanging behavior and environments. Far from being mere biological noise, recent research suggests that drift is an inherent property of how neural networks learn, explore, and maintain information.

The Manifold of Equivalent Solutions

How can behavior remain stable if the underlying code is moving? The core intuition lies in the geometry of neural activity.

In high-dimensional neural space, a task can often be solved in many different ways. Farhad Pashakhanloo and Alexei A. Koulakov demonstrate in Stochastic Gradient Descent-Induced Drift of Representation in a Two-Layer Neural Network that learning processes like Stochastic Gradient Descent (SGD) don't settle on a single point. Instead, they find a manifold of equivalent solutions—a "flat" region in the landscape of possible synaptic weights where many different configurations yield the same perfect performance.

Once a network reaches this manifold, internal noise causes the representation to perform a "random walk" along it. Because the walk stays within the manifold, the output (the behavior) never changes, even though the internal weights and activations are drifting. Ji Xia found empirical evidence for this in Stable representation of a naturalistic movie emerges from episodic activity with gain variability, showing that while individual neurons in the visual cortex drift, the population as a whole remains stable because the fluctuations occur in a "non-coding" space—dimensions that the rest of the brain simply ignores.

Drivers of the Drift

If drift is a random walk, what is pushing the neurons around? Researchers have identified several distinct "engines" of drift:

  1. Intrinsic Excitability: Geoffroy Delamare and Denise J. Cai propose that neurons have slow, natural fluctuations in how easily they fire. In Drift of neural ensembles driven by slow fluctuations of intrinsic excitability, they show that these fluctuations bias which neurons are recruited into a memory ensemble at any given time, effectively "time-stamping" memories as the ensemble gradually evolves.
  2. Task-Irrelevant Stimuli: Even when you focus on a task, your brain processes background "noise." Farhad Pashakhanloo argues in Contribution of task-irrelevant stimuli to drift of neural representations that learning to ignore these distractors actually drives the drift of the task-relevant representations, as the network constantly adjusts its internal weights to suppress irrelevant input.
  3. Synaptic Turnover: The physical connections between neurons (synapses) are not static. Sven Goedeke and Raoul-Martin Memmesheimer suggest in Drifting Assemblies for Persistent Memory that continuous synaptic remodeling and homeostatic plasticity allow the brain to "swap out" the specific neurons in an assembly without losing the stored information.

From Volatility to Crystallization

Drift is not uniform across the brain or throughout the learning process. It appears to follow a trajectory from high volatility to relative stability.

In Volatile working memory representations crystallize with practice, Arash Bellafard and Peyman Golshani observed that neural representations in the motor cortex are initially highly unstable. However, as an animal becomes an "expert" at a task, the representation crystallizes, becoming significantly more stable. This suggests that the brain uses drift to explore the solution manifold early on, eventually settling into a more robust state once an optimal strategy is found.

Furthermore, different "anchors" can provide stability amidst the drift. Hanlin Zhu and Andreas S. Tolias discovered that while the firing rates of neurons drift, the precise timing of their spikes—the temporal code—remains remarkably consistent (Temporal coding carries more stable cortical visual representations than firing rate over time).

Bug or Feature?

Is representational drift a flaw of biological "wetware," or does it serve a purpose? Paul Masset and colleagues argue in Drifting neuronal representations: Bug or feature? that drift might be a feature of robust Bayesian learning. By constantly shifting, the brain avoids getting stuck in "local minima" (suboptimal solutions) and maintains the flexibility to incorporate new information.

This is supported by Phil Rich and Stephan Y. Thiberge, who found that behavioral errors actually accelerate drift (Error-driven changes in hippocampal representations accompany flexible re-learning). This "neural reset" allows the brain to quickly abandon a failing strategy and explore new parts of the solution manifold, facilitating rapid adaptation in changing environments.

By reframing the brain not as a static library of memories, but as a dynamic system navigating a manifold of equivalent solutions, we can finally explain how our internal world remains so fluid while our external actions remain so reliable.

Papers in this version

Showing 60 of 79 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 37 papers
  2. Version 1 · 2026-10-04 · 79 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_representational_drift_74b44f5a,
  title = {representational drift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/representational-drift/rdr_74b44f5a695a4026aa34a1c3291f53ac},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_functional_stability_amidst_neural_repre_74b44f5a_v1,
  title = {Functional Stability Amidst Neural Representational Drift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/functional-stability-amidst-neural-representational-drift/rdr_74b44f5a695a4026aa34a1c3291f53ac/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}