A Hodgkin-Huxley Type Neuron Model That Learns Slow Non-Spike Oscillation

NeurIPS 1993

TL;DR

This paper introduces a novel gradient descent algorithm to automatically tune the complex parameters of Hodgkin-Huxley (H-H) neuron models using experimental membrane potential data. This method successfully models slow, non-spiking oscillations in a specific lobster neuron, revealing a previously underappreciated role for the A-current in this phenomenon.

What Problem Does This Paper Address?

Understanding neuronal behavior relies heavily on biophysically detailed models like the Hodgkin-Huxley (H-H) model. However, these models are notoriously complex, featuring dozens of variables and hundreds of parameters (e.g., maximal conductances, thresholds, time constants of ion channels). Traditionally, these parameters are painstakingly determined through difficult voltage-clamp experiments or manual "hand-tuning" in simulations. This manual process is unreliable, time-consuming, and often leads to non-unique parameter sets that can mimic the neuron's behavior without necessarily reflecting its true underlying biophysics. The paper addresses the challenge of systematically and automatically estimating these crucial H-H model parameters from readily available experimental data, specifically membrane potential trajectories.

Methods

The authors propose a gradient descent algorithm adapted for H-H type neuron models. They frame the H-H model as a network of interconnected components (sigmoid functions, multipliers, leaky integrators), similar to continuous-time recurrent neural networks. The core innovation lies in using the experimental membrane potential trajectory as a "target" and minimizing the error between the model's output and this target.

Key methodological aspects include:

  • Gradient Descent on Normalized Parameters: To handle the vastly different scales and physical dimensions of H-H parameters, they normalize them by their default values before applying gradient descent.
  • Teacher Forcing: During training, the activation and inactivation variables of the ion channels are driven by the target membrane potential trajectory, not the model's current output. This "teacher forcing" is crucial for training autonomous oscillation patterns.
  • Variation Equations: They leverage "variation equations" to efficiently calculate the error gradient with respect to each parameter. This is more efficient than traditional methods that might require perturbing each parameter individually.
  • On-line Parameter Update: Instead of batch updates, parameters are updated continuously as the model runs, using a running average of the gradient. This scheme is more stable and allows for larger learning rates.

Datasets

The paper uses two main sources of data:

  • Simulated Spiking Neuron Data: For initial validation, they generated membrane potential trajectories from a known H-H model of a spiking neuron under various current injection levels. These simulated trajectories served as the "ground truth" for training their learning algorithm.
  • Experimental Data from Lobster Stomatogastric Ganglion (DG cell): The primary application of their algorithm was to model the slow, non-spiking oscillation of an identified neuron (DG cell) from the lobster stomatogastric ganglion. They used experimental data (specifically, membrane potential trajectories in the range of -70 to -50 mV) from this real neuron as the target for training their H-H model.

Key Insights and Ideas

  • Automated Parameter Estimation is Feasible: The study demonstrates that complex H-H model parameters can be systematically estimated from relatively simple experimental data like membrane potential recordings, bypassing the need for extensive voltage-clamp experiments.
  • Parameter Space Exploration: The learning algorithm doesn't just find a single parameter set but can systematically estimate a region in the parameter space that yields the observed behavior, offering a more robust understanding of model sensitivity.
  • Novel Role of A-current: The modeling of the DG cell's slow oscillation revealed a significant role for the A-current (a potassium current with inactivation) in generating oscillations in the hyperpolarized range (-70 to -50 mV).
  • Analogy to H-H Model Dynamics: The authors draw an interesting analogy between the DG cell model's dynamics and the original H-H model. They suggest that the A-current's negative conductance in the relevant voltage range acts like the fast inward sodium current in the H-H model (providing positive feedback), while the H-current acts like the outward potassium current (providing negative feedback).

Key Results

  • Spiking Neuron Model: The algorithm successfully estimated parameters for a spiking neuron model, achieving an RMS error of less than 1.3 mV after 100 training cycles in 83 out of 100 trials, starting from random initial parameters.
  • DG Cell Model: The trained DG cell model closely matched the experimental membrane potential trajectories. The learned parameters for the A-current and H-current were found to be critical for generating the slow, non-spiking oscillations observed in the -70 to -50 mV range.

Limitations and Future Work

  • Parameter Space Unreliability: The paper acknowledges that even with a good match, the parameter space can be unreliable, with multiple settings potentially leading to similar behavior. The proposed method aims to estimate a region rather than a single point to mitigate this.
  • Generalizability: While successful for the DG cell, the algorithm's performance on other neuron types or more complex oscillatory patterns is not extensively explored in this paper.
  • Future Work: The authors suggest that this approach could be extended to more complex neuron models and networks, potentially accelerating the process of building accurate computational models of neural circuits. The identification of the A-current's role in slow oscillations also opens avenues for further experimental and theoretical investigation into its biophysical mechanisms.

Summary updated

BibTeX
@inproceedings{1993hodgkin,
  title = {{A Hodgkin-Huxley Type Neuron Model That Learns Slow Non-Spike Oscillation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1993},
}
Figure