TL;DR
This paper proposes a computational model explaining why only some neurons recover their activity after visual deprivation, even when inhibition is globally reduced. The key insight is that the size of a neuron's "spontaneous subnetwork" dictates its homeostatic recovery, with larger subnetworks showing stronger recovery. These findings were validated experimentally, suggesting that pre-existing functional network structure plays a crucial role in homeostatic plasticity.
What Problem Does This Paper Address?
The adult brain has mechanisms called "homeostatic plasticity" that prevent neural activity from becoming too high or too low. Visual deprivation (like covering one eye) is a common way to study this. Researchers observed that after visual deprivation, while inhibition in the brain is generally reduced (a process called disinhibition), only a subset of neurons actually recover their normal activity levels. This is puzzling: if inhibition is reduced everywhere, why don't all neurons recover? Furthermore, these recovering neurons tend to be interconnected, forming functional networks before the deprivation. The paper aims to explain how these diverse homeostatic responses emerge, particularly in the context of global disinhibition, and how pre-existing network structure influences this recovery.
Methods
The researchers developed a computational model of a recurrent neural network. This network consisted of excitatory and inhibitory neurons with all-to-all connectivity. Key aspects of their approach include:
- Neuron Model: A simple firing rate model was used, where neuron output is a function of its input.
- Synaptic Plasticity:
- Excitatory Synapses: Modeled using the BCM learning rule, which is a form of Hebbian plasticity that also incorporates a homeostatic component to regulate firing rates. Synaptic pruning of weak connections was also included.
- Inhibitory Synapses: A homeostatic rule was implemented to adjust inhibitory weights, aiming to maintain a target firing rate.
- Network Development: The model simulated the development of visual stimulus selectivity and the formation of "spontaneous subnetworks" (groups of neurons with correlated spontaneous activity) through these plasticity rules.
- Simulating Visual Deprivation: The model simulated visual deprivation by simply turning off external visual input to the network.
- Measuring Recovery: Homeostatic recovery was quantified by comparing firing rates before deprivation, immediately after deprivation, and after a recovery period. The inhibition-excitatory (I/E) ratio was also measured.
- Experimental Validation: The model's predictions were tested using experimental data from adult mouse visual cortex, where calcium imaging was used to measure neuronal activity before and after monocular enucleation (visual deprivation).
Datasets
The study used experimental data from adult mouse visual cortex (layer 2/3 excitatory neurons) collected in previous work ([Barnes et al., 2015]). This data involved chronic calcium imaging before and after monocular enucleation. No new datasets were introduced.
Key Insights and Ideas
- Network Size Matters: The size of a neuron's "spontaneous subnetwork" is a critical determinant of its homeostatic recovery. Neurons in larger subnetworks show a stronger recovery.
- Disinhibition Unmasks Recurrent Excitation: Homeostatic recovery in the model is driven by disinhibition, which unmasks the underlying spontaneous recurrent excitation within subnetworks.
- Subnetwork-Specific Recovery: The model explains why only a subset of neurons recover, even with global disinhibition, by showing that recovery is not uniform but depends on the neuron's embedded functional network.
- Pre-existing Structure is Key: The strength of homeostatic recovery is linked to the proportion of "enduring functional associations" (persistent correlations) a neuron has, suggesting that pre-existing network structure, rather than the creation of new connections, is crucial.
- Synaptic Scaling is Not the Whole Story: The model demonstrates that homeostatic recovery can occur without synaptic scaling, and that synaptic scaling alone leads to uniform recovery, which is inconsistent with experimental observations of subnetwork-specific responses.
Key Results
- Model Validation: The computational model successfully reproduced the observed phenomenon of diverse homeostatic responses to visual deprivation in adult mouse visual cortex.
- Network Size Effect: Neurons belonging to larger spontaneous subnetworks exhibited significantly stronger homeostatic recovery compared to those in smaller subnetworks.
- Persistent Correlations Predict Recovery: The model predicted, and experiments confirmed, that neurons with a high proportion of persistent correlations (indicating strong enduring functional associations) showed the strongest homeostatic recovery.
- I/E Ratio Reduction: The model reproduced the experimental finding of a reduced inhibition-excitatory (I/E) ratio after visual deprivation, which is a consequence of the homeostatic plasticity mechanisms.
- Synaptic Scaling Inconsistency: A simplified synaptic scaling model led to uniform recovery across all neurons, failing to explain the observed subnetwork-specific responses and the correlation between persistent correlations and recovery strength.
Limitations and Future Work
- Subnetwork Size vs. Strength: While the model highlights subnetwork size as a determinant, the authors acknowledge that diversity in the strength of recurrent synapses within subnetworks could also lead to similar diverse responses, independent of size.
- Origin of Seeded Subnetworks: The model uses "seeded" spontaneous subnetworks, meaning their initial structure is predefined. The paper discusses plausible biological mechanisms for how such subnetworks might form (e.g., electrical coupling, common inputs from other modalities) but doesn't fully model their spontaneous emergence.
- Functional Advantage: The paper raises the question of whether subnetwork-specific recovery is functionally advantageous compared to complete recovery via synaptic scaling, suggesting it might help conserve underlying network structure. Further investigation into this functional aspect is warranted.
- Crossmodal Plasticity: The findings suggest a basis for crossmodal plasticity, where sensory modalities can influence each other. Future experiments could explore this further.
- Agnostic to Origin: The model is agnostic to the precise origin of non-visual subnetworks, leaving room for further exploration of how different sensory inputs might shape these structures.
Summary updated