Learning excitatory-inhibitory neuronal assemblies in recurrent networks

Owen Mackwood, Laura Naumann, Henning Sprekeler

TL;DR

This paper proposes a computational model where both the input and output synapses of inhibitory interneurons learn simultaneously. This dual plasticity is crucial for forming stimulus-specific "excitatory-inhibitory (E/I) assemblies" in neural circuits, mirroring observations in the mouse visual cortex. These assemblies, in turn, enable competition between similarly tuned excitatory neurons, shaping how the brain processes information.

What Problem Does This Paper Address?

The brain's intricate connectivity and how it arises from simple learning rules is a major puzzle in neuroscience. Specifically, in the primary visual cortex (V1) of mice, researchers have observed that connections between excitatory neurons (pyramidal cells) and inhibitory neurons (PV interneurons) are not random. Instead, synapses tend to be stronger between neurons that respond to similar visual features, forming "E/I assemblies." This suggests a more sophisticated, stimulus-specific form of feedback inhibition than previously thought, especially in areas like V1 that lack clear topographic organization. The challenge is to understand the underlying plasticity mechanisms that could create these assemblies.

Methods

The authors developed a computational model of a recurrent neural network composed of excitatory (Pyr) and inhibitory (PV) neurons. The core of their approach lies in simulating activity-dependent synaptic plasticity rules for both the excitatory inputs to PV interneurons and the inhibitory outputs from PV interneurons. They employed a homeostatic learning objective: to keep the net synaptic current onto Pyr neurons at a target level across various stimuli. This objective was implemented using gradient-descent-like rules, which they approximated for biological plausibility. To test the necessity of each plasticity mechanism, they performed "knock-out" experiments, disabling plasticity in either input or output synapses. They also simulated single-neuron perturbation experiments to observe the network's response dynamics.

Datasets

Not discussed in the paper. The study relies on simulated data generated by their computational model.

Key Insights and Ideas

  • Synergy is Key: The most striking finding is that both input and output plasticity of PV interneurons are essential for forming E/I assemblies. Neither mechanism alone is sufficient.
  • Homeostatic Drive: A simple homeostatic objective – maintaining a target firing rate for excitatory neurons – can drive the formation of complex, stimulus-specific inhibitory circuits.
  • Feature Competition: Once formed, these E/I assemblies create a "feature competition" where similarly tuned excitatory neurons inhibit each other, a phenomenon observed experimentally.
  • Beyond "Blanket Inhibition": The model demonstrates that inhibition can be highly specific, contradicting the older "blanket of inhibition" hypothesis, even in networks without topographic maps.

Key Results

  • When both input and output plasticity of PV interneurons were active, the model successfully reproduced the stimulus-specific E/I assembly structure observed in mouse V1, including correlations between synaptic strength and response similarity.
  • Knocking out output plasticity prevented the formation of E/I assemblies and stimulus-specific inhibition.
  • Knocking out input plasticity also prevented E/I assembly formation, leading to untuned inhibitory inputs to Pyr neurons.
  • Perturbing a single excitatory neuron in the trained network led to the suppression of similarly tuned neurons, consistent with experimental findings.

Limitations and Future Work

  • Simplified Network: The model uses only one type of interneuron (PV) and does not include feedforward inhibition, which could influence the results.
  • Biological Plausibility of Plasticity Rules: While the authors use biologically inspired approximations, the exact mechanisms for implementing these rules in real neurons remain an open question, particularly for the excitatory input plasticity to interneurons.
  • Future Directions: The authors suggest exploring the role of feedforward inhibition, incorporating more diverse interneuron types, and investigating how these plasticity rules might differ in developing versus adult brains. They also propose that the observed feature competition could be beneficial for efficient sensory representation.

Summary updated

Abstract

Understanding the connectivity observed in the brain and how it emerges from local plasticity rules is a grand challenge in modern neuroscience. In the primary visual cortex (V1) of mice, synapses between excitatory pyramidal neurons and inhibitory parvalbumin-expressing (PV) interneurons tend to be stronger for neurons that respond to similar stimulus features, although these neurons are not topographically arranged according to their stimulus preference. The presence of such excitatory-inhibitory (E/I) neuronal assemblies indicates a stimulus-specific form of feedback inhibition. Here, we show that activity-dependent synaptic plasticity on input and output synapses of PV interneurons generates a circuit structure that is consistent with mouse V1. Computational modeling reveals that both forms of plasticity must act in synergy to form the observed E/I assemblies. Once established, these assemblies produce a stimulus-specific competition between pyramidal neurons. Our model suggests that activity-dependent plasticity can refine inhibitory circuits to actively shape cortical computations.
BibTeX
@inproceedings{mackwood2021learning,
  author = {Owen Mackwood and Laura Naumann and Henning Sprekeler},
  title = {{Learning excitatory-inhibitory neuronal assemblies in recurrent networks}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.59715},
}
Figure