Astrocyte Regulation Confines Representational Drift to Null-space
This is my NEUR 1440 paper, where I explored a neuro-glial hypothesis for one of the central puzzles in neural decoding: how behavior can remain stable while the neural representations supporting it keep drifting.
Full Manuscript
The rendered pages below preserve the full PDF layout and include all figures.
Page 1 of 10.
Page 2 of 10.
Page 3 of 10.
Page 4 of 10.
Page 5 of 10.
Page 6 of 10.
Page 7 of 10.
Page 8 of 10.
Page 9 of 10.
Page 10 of 10.
Reading Notes
Why Representational Drift Matters
Representational drift describes the observation that neural activity patterns associated with the same stimulus, cognitive state, or movement can change continuously over days or weeks. A neuron that appears tuned to one movement direction today may rotate its preferred direction, change its modulation depth, or drop out of the functional ensemble later.
For intracortical brain-computer interfaces, this drift is not just a biological curiosity. Most decoders are trained on neural activity from a limited calibration window. When the neural distribution changes, a fixed decoder gradually becomes mismatched to the user, which creates the familiar burden of recalibration.
The interesting paradox is that animals do not usually forget a learned behavior every time the representation drifts. Something in the circuit must allow neural activity to change while preserving the behavioral readout.
The Null-space View
One useful way to think about this problem is through neural population geometry. A high-dimensional neural population can produce a much lower-dimensional behavioral output, which means many changes in neural activity are behaviorally silent. These silent directions form an output-null space.
The null-space hypothesis suggests that representational drift is tolerated as long as it remains mostly in this output-null space. Individual neurons may change, but the population-level readout stays stable.
That explanation is elegant, but it leaves a physiological question open: what actually constrains the drift so that it does not leak into behaviorally relevant coding dimensions?
The Astrocyte Hypothesis
In this paper, I propose that astrocytes may act as biological stabilizers of the neural manifold. Astrocytes tile local cortical territory, monitor large numbers of synapses, clear neurotransmitters, buffer extracellular potassium, and release gliotransmitters that modulate plasticity. These properties make them well positioned to sense and regulate population-level activity rather than only single-neuron firing.
The core hypothesis is that astrocytes help constrain representational drift in two complementary ways:
- They anchor individual neurons by damping excessive changes in excitability or synaptic gain.
- They constrain the direction of population drift by adjusting local ensemble excitability so that the behavioral readout remains stable.
In this view, astrocytes form a slow feedback loop around local neural ensembles. They do not prevent all drift; instead, they help shape drift into dimensions where it does not disrupt behavior.
Excitability as a Drift Driver
A second part of the paper connects this idea to the excitability-drift hypothesis. If slow fluctuations in intrinsic excitability determine which neurons are recruited into a functional ensemble, then regulating excitability becomes a way to regulate representational drift.
Astrocytic glutamate transporters such as GLT1 are especially relevant here. When astrocytic regulation is weakened, neuronal excitability can remain too high. I frame this with a machine learning analogy: high excitability behaves like a high learning rate. It may accelerate learning, but it can also make the learned trajectory noisy and unstable. Healthy astrocytic regulation may act more like a learning-rate scheduler, allowing plasticity early while supporting stable convergence later.
A Possible Experiment
To test this hypothesis, I propose a retention experiment inspired by astrocyte manipulation studies in motor learning. After mice learn a motor task to stable performance, one group would undergo astrocytic disruption, such as GLT1 manipulation or chemogenetic silencing, while a control group receives a sham manipulation.
If astrocytes help keep drift in the output-null space, then disrupting them after learning should cause faster behavioral decay during later recall tests. The prediction is not simply that learning becomes different, but that an already learned representation becomes less robust because neural drift is no longer properly constrained.
Implications for iBCIs
The engineering motivation is a neuron-astrocyte hybrid BCI architecture. Fast neuronal spiking activity would still drive real-time motor decoding, while slower astrocytic signals could provide contextual state information about excitability, fatigue, plasticity, or ensemble recruitment.
In that framework, astrocytes would not replace neurons as the main control signal. Instead, they could serve as a biological meta-learner: a slow stream that helps the decoder infer which neural features are currently reliable and how the population state is changing over time.
The broader point is that long-term robust neural decoding may require moving beyond a purely neuron-centric view. If astrocytes help stabilize the cortical manifold, then understanding neuro-glial dynamics could be useful not only for neuroscience, but also for building iBCIs that remain stable over months or years.