Neural Noise- Representational drift can be reproduced in computational models through Hebbian learning combined with neural inhibition mechanisms.1
- The model explains why neural codes for stable behaviors still drift over time, a key issue for long-term BCI decoder stability.1 1
Weekly enrichment (2026-07-20)
- The assigned bioRxiv preprint page was inaccessible during research (blocked by a cookie/consent wall), so these bullets synthesize the convergent 2025–2026 literature on the same mechanism the title describes: representational drift generated by Hebbian learning together with neural inhibition.2 3
- A central computational result is that drift emerges from a balance of two processes: activity-dependent Hebbian-like plasticity and activity-independent stochastic synaptic changes; models reproduce observed correlation dynamics only when both are present, so Hebbian learning counterbalances stochastic turnover and prevents functional degradation.2 3
- Empirical grounding comes from chronic (multi-day) calcium imaging in mouse auditory cortex, where signal correlations between neuron pairs predicted their future noise correlations, suggesting stimulus-driven co-activation increases effective connectivity in a Hebbian manner.2 3
- Inhibition and excitatory–inhibitory (E/I) balance are the key stabilizers: a spiking E/I network with predictive plasticity showed individual neurons gradually change their preferred pattern while ensemble-level coding stays stable, with drift expressed as coordinated rotation and translation of the population state space.4
- That same predictive E/I model, extended to hippocampal CA1 place coding, reproduced experience-dependent tuning-curve drift, including the dissociation between elapsed time and intervening exposure.4
- A complementary study shows “locally balanced inhibition” (each neuron receiving inhibition proportional to its own excitatory in-degree) prevents runaway positive feedback: without it, strong Hebbian coupling collapses drift into a frozen, stereotyped code, whereas balanced inhibition preserves flexible, non-runaway reorganization and improves memory strength.5
- In sensory circuits, a mitral cell–granule cell model of the olfactory bulb reproduced gain adaptation, pattern separation/convergence, and encoding-subspace rotation using Hebbian plasticity plus structural connectivity constraints, while the relative geometry of odor responses—the low-dimensional manifold—stayed stable despite global drift.6
- The recurring theme with direct BCI relevance is that population-level representations remain decodable even as single-neuron tuning drifts, implying long-term decoders must track a slowly rotating but geometry-preserving manifold rather than assume fixed neuron-to-feature mappings.4 6