Feedback Inhibition Shapes Emergent Computational Properties of Cortical Microcircuit Motifs

Zeno Jonke, Robert Legenstein, Stefan Habenschuss, Wolfgang Maass

This research paper, published by Zeno Jonke and colleagues from the Graz University of Technology, explores how the brain’s "wiring" and learning rules allow it to make sense of a messy, overlapping world. By modeling a specific motif of the cerebral cortex—the interplay between excitatory pyramidal cells and inhibitory PV+ cells—the authors demonstrate that biologically realistic "soft" inhibition is the key to unlocking advanced computational properties like blind source separation and sparse assembly coding.

TL;DR

By replacing idealized "Winner-Take-All" models with biologically grounded divisive inhibition, the authors show that cortical circuits can autonomously learn to disentangle overlapping sensory inputs. This mechanism allows the brain to represent complex scenes using sparse "assemblies" of neurons without succumbing to a combinatorial explosion of required cells.

What Problem Does This Paper Address?

In theoretical neuroscience, many models rely on a "Winner-Take-All" (WTA) architecture. In a WTA circuit, the most active neuron suppresses all others, ensuring a clear, unambiguous output. However, real biology is rarely that "hard." Experimental data shows that in the cortex (specifically Layer 2/3), inhibition is "divisive" or "soft"—multiple neurons can, and do, fire simultaneously in response to a stimulus.

The challenge is understanding what we gain from this "softness." If multiple neurons fire at once, how does the brain prevent a "combinatorial explosion," where it would need a unique neuron for every possible combination of things we might see or hear? The authors investigate whether this realistic inhibition, combined with Spike-Timing Dependent Plasticity (STDP), allows a circuit to perform Blind Source Separation: the ability to identify individual "sources" (like a single voice in a crowded room) from a mixture of signals.

Methods

The authors developed Model $\mathcal{M}$, a data-based microcircuit model consisting of 400 excitatory neurons and 100 inhibitory neurons.

Key technical features include:

  • Stochastic Spike Response Model: Neurons are modeled as stochastic processes where the firing rate depends exponentially on the membrane potential.
  • Divisive Inhibition: Unlike WTA models that subtract activity, this model uses a divisive normalization where inhibition acts as the denominator in the firing rate equation, matching in-vivo observations.
  • STDP: Synapses from input neurons to excitatory neurons are plastic, meaning they strengthen or weaken based on the precise millisecond-timing of spikes.
  • Connectivity: Connection probabilities (e.g., the 57.5% chance of an E $\to$ I connection) were pulled directly from experimental data from the Petersen Lab.

Datasets

The model was evaluated using three primary synthetic benchmarks designed to mimic sensory processing:

  1. Orientation Selectivity: 2D pixel arrays of bars at 180 different orientations.
  2. Spatio-temporal Patterns: 150ms long, time-varying firing rate patterns generated via Ornstein-Uhlenbeck processes.
  3. Földiák’s Superposition-of-Bars: A classic benchmark for blind source separation. The network is shown an 8x8 grid where up to three random horizontal or vertical bars are superimposed. The network must learn to represent the individual bars, even though it only ever sees them in "mixed" form.

Key Insights and Ideas

  • From WTA to k-WTA: The "soft" inhibition transforms the circuit into a k-WTA (k-Winners-Take-All) system. Instead of one winner, an "assembly" of roughly 17 neurons responds to a feature. This distributed representation is more robust and computationally powerful than a single-neuron code.
  • Autonomous Unmixing: Through STDP, the network naturally learns to "unmix" signals. If it frequently sees Bar A and Bar B together, it doesn't create a "Bar A+B" detector. Instead, it develops a "Bar A" assembly and a "Bar B" assembly that fire together.
  • The Role of I-I Connections: The paper provides a hypothesis for why inhibitory neurons are connected to each other. These connections sharpen the timing of the circuit, allowing inhibition to track excitation with a tiny 3ms lag, which is essential for stable learning.
  • Avoiding Combinatorial Explosion: Because the code is modular (using different assemblies for different features), the network can represent a near-infinite variety of scenes using a limited pool of neurons.

Key Results

  • Orientation Tuning: After learning, excitatory neurons developed sharp, Gaussian-like tuning curves for specific bar orientations, while inhibitory neurons remained broadly tuned—exactly what is seen in biological "orientation pinwheels."
  • Blind Source Separation: In the overlapping bars task, the model achieved a high F1-measure of 0.87, indicating it could reliably signal the presence of individual bars within a mixture.
  • WTA Comparison: Standard WTA models failed this task; they were unable to develop a modular code and instead tried to learn specific combinations of bars, which is inefficient and non-scalable.
  • Temporal Dynamics: The model successfully replicated the ~3.5ms lag between excitation and inhibition observed in-vivo, suggesting the model's architecture captures the fundamental physics of cortical motifs.

Limitations and Future Work

  • Fixed Inhibitory Weights: While the input-to-E synapses were plastic, the inhibitory synapses (E $\to$ I, I $\to$ E, I $\to$ I) were kept at fixed, data-derived values. The authors note that adding inhibitory plasticity (iSTDP) is a logical next step.
  • Single Motif Focus: The study looks at a single microcircuit motif. How these motifs behave when stacked into deep, hierarchical networks remains an open question.
  • Experimental Verification: The authors propose that their findings could be tested in live animals by using optogenetics to silence specific I-I connections and measuring if the excitation-inhibition lag increases as predicted.

Summary updated

Abstract

Cortical microcircuits are very complex networks, but they are composed of a relatively small number of stereotypical motifs. Hence, one strategy for throwing light on the computational function of cortical microcircuits is to analyze emergent computational properties of these stereotypical microcircuit motifs. We are addressing here the question how spike timing-dependent plasticity shapes the computational properties of one motif that has frequently been studied experimentally: interconnected populations of pyramidal cells and parvalbumin-positive inhibitory cells in layer 2/3. Experimental studies suggest that these inhibitory neurons exert some form of divisive inhibition on the pyramidal cells. We show that this data-based form of feedback inhibition, which is softer than that of winner-take-all models that are commonly considered in theoretical analyses, contributes to the emergence of an important computational function through spike timing-dependent plasticity: The capability to disentangle superimposed firing patterns in upstream networks, and to represent their information content through a sparse assembly code.<b>SIGNIFICANCE STATEMENT</b> We analyze emergent computational properties of a ubiquitous cortical microcircuit motif: populations of pyramidal cells that are densely interconnected with inhibitory neurons. Simulations of this model predict that sparse assembly codes emerge in this microcircuit motif under spike timing-dependent plasticity. Furthermore, we show that different assemblies will represent different hidden sources of upstream firing activity. Hence, we propose that spike timing-dependent plasticity enables this microcircuit motif to perform a fundamental computational operation on neural activity patterns.
BibTeX
@article{jonke2017feedback,
  author = {Zeno Jonke and Robert Legenstein and Stefan Habenschuss and Wolfgang Maass},
  title = {{Feedback Inhibition Shapes Emergent Computational Properties of Cortical Microcircuit Motifs}},
  journal = {Journal of Neuroscience},
  year = {2017},
  doi = {10.1523/jneurosci.2078-16.2017},
}

Loading related content...

Figure