Subtractive, divisive and non-monotonic gain control in feedforward nets linearized by noise and delays

Jorge F. Mejías, Alexandre Payeur, Erik Selin, Leonard Maler, André LongtinFrontiers Comput. Neurosci 2014 · DOI

TL;DR

This paper presents a computational model of a neural circuit that can exhibit three distinct types of gain control: subtractive, divisive, and non-monotonic. By incorporating noise and temporal delays into a feedforward inhibitory network, the model explains how neural systems can flexibly adjust their input-output relationships, a crucial mechanism for information processing.

What Problem Does This Paper Address?

Neural systems constantly adapt how they process incoming information. This adaptation, known as "gain control," allows neurons to scale their responses based on context or stimulus properties. While subtractive gain control (a simple shift in the response curve) is well-understood, the biophysical mechanisms behind more complex gain control strategies like divisive (scaling the slope) and non-monotonic (complex, non-linear changes) have been elusive. This paper aims to bridge this gap by proposing a unified, biologically plausible model that can generate all three types of gain control.

Methods

The authors developed a spiking neural network model inspired by the electrosensory lateral-line lobe (ELL) of weakly electric fish. The core of the model consists of two populations of neurons: deep pyramidal (DP) neurons that receive sensory input and superficial pyramidal (SP) neurons that receive input from DP neurons via an inhibitory feedforward pathway.

Key methodological aspects include:

  • Leaky Integrate-and-Fire (LIF) Neuron Model: Standard model for simulating spiking neurons.
  • Feedforward Inhibition: The crucial inhibitory connection from DP to SP neurons, which is modulated by parameters like strength (G) and noise levels (σ).
  • Analytical and Numerical Analysis: The authors used both mathematical derivations (e.g., for f-I curves) and simulations to explore the network's behavior.
  • Noise and Temporal Delays: Explicitly incorporated into the model to investigate their impact on gain control and signal processing.
  • f-I Curve Analysis: The primary tool for characterizing gain control, showing the relationship between input current and output firing rate.

Datasets

Not discussed in the paper. The study relies on computational modeling and theoretical analysis, with experimental data from existing literature used for validation.

Key Insights and Ideas

  • Unified Mechanism for Diverse Gain Control: A single, relatively simple circuit architecture (feedforward inhibition with noise) can produce subtractive, divisive, and non-monotonic gain control. This suggests a common underlying principle for these different computational strategies.
  • Noise and Inhibition Strength as Key Modulators: The type of gain control observed is critically dependent on the level of neuronal noise (σ) and the strength of the inhibitory feedforward connection (G). This highlights the importance of these intrinsic neuronal properties.
  • Divisive Gain Control Emerges with Noise: Subtractive gain control is observed in the deterministic case (no noise), but the introduction of noise naturally leads to divisive gain control.
  • Non-Monotonicity from Strong Inhibition: When the feedforward inhibition is sufficiently strong, it can overcome the direct excitatory input, leading to a non-monotonic f-I curve. This is a novel finding for this type of circuit.
  • Temporal Delays Linearize f-I Curves: Surprisingly, temporal delays in the inhibitory pathway, rather than disrupting the signal, can lead to a linearization of the f-I curves by increasing variability. This is a counter-intuitive but important finding for understanding signal processing in real neural circuits.
  • Model Explains In Vivo Data: The model successfully reproduces experimental data from weakly electric fish, specifically the divisive gain control observed in SP neurons.

Key Results

  • Subtractive Gain Control: Observed in the deterministic case (σ = 0), characterized by a shift in the f-I curve.
  • Divisive Gain Control: Emerges with the introduction of noise (σ > 0) and modulated inhibitory strength (G). The slope of the f-I curve is scaled.
  • Non-Monotonic Gain Control: Occurs with strong negative feedforward inhibition (large |G|) and noise, leading to a peak in the f-I curve.
  • Linearization by Delay: Temporal delays in the feedforward pathway increase the variability of the effective input, leading to more linear f-I curves.
  • Quantitative Agreement with Data: The model's predictions for divisive gain control in electric fish SP neurons align well with experimental observations.

Limitations and Future Work

  • Simplified Neuron Model: The use of LIF neurons is a simplification; more complex neuron models could reveal additional dynamics.
  • Homogeneous Populations: The model assumes uniform properties within DP and SP neuron populations. Real neural systems exhibit heterogeneity, which the authors suggest could lead to mixed gain control regimes.
  • Ad Hoc Parameterization for Experimental Fit: While the model fits experimental data, the specific function used to describe the stimulus-dependent strength of inhibition (G) was determined in an ad hoc manner. Future work could explore more principled ways to derive this dependency.
  • Future Directions:
    • Investigate the impact of non-uniform neural properties and more complex network architectures.
  • Explore the role of synaptic plasticity in dynamically modulating G and thus gain control.
  • Further analyze the implications of delay-induced linearization for information coding.
  • Test the model's predictions for non-monotonic gain control in experimental settings.

Summary updated

BibTeX
@article{mejias2014subtractive,
  author = {Jorge F. Mejías and Alexandre Payeur and Erik Selin and Leonard Maler and André Longtin},
  title = {{Subtractive, divisive and non-monotonic gain control in feedforward nets linearized by noise and delays}},
  journal = {Frontiers Comput. Neurosci},
  year = {2014},
  doi = {10.3389/fncom.2014.00019},
}
Figure