Capturing multiple timescales of adaptation to second-order statistics with generalized linear models: gain scaling and fractional differentiation

Kenneth W. Latimer, Adrienne L. Fairhall

Abstract

Single neurons can dynamically change the gain of their spiking responses to account for shifts in stimulus variance. Moreover, gain adaptation can occur across multiple timescales. Here, we examine the ability of a simple statistical model of spike trains, the generalized linear model (GLM), to account for these adaptive effects. The GLM describes spiking as a Poisson process whose rate depends on a linear combination of the stimulus and recent spike history. The GLM successfully replicates gain scaling observed in Hodgkin-Huxley simulations of cortical neurons that occurs when the ratio of spike-generating potassium and sodium conductances approaches one. Gain scaling in the GLM depends on the length and shape of the spike history filter. Additionally, the GLM captures adaptation that occurs over multiple timescales as a fractional derivative of the stimulus variance, which has been observed in neurons that include long timescale after hyperpolarization conductances. Fractional differentiation in GLMs requires long spike history that span several seconds. Together, these results demonstrate that the GLM provides a tractable statistical approach for examining single-neuron adaptive computations in response to changes in stimulus variance.

Introduction

Neurons are not static transducers; they are dynamic adaptors. When the "volume" of their input (the stimulus variance) changes, neurons adjust their sensitivity, or gain, to maintain efficient information processing. This adaptation occurs across multiple timescales, ranging from milliseconds to several seconds. Traditionally, researchers have faced a trade-off between biophysically detailed models (like Hodgkin-Huxley), which are realistic but difficult to fit to data, and statistical models (like Generalized Linear Models), which are tractable but often seen as too simple to capture complex computations like power-law adaptation.

This paper investigates whether a simple GLM can perform the same adaptive computations as complex biophysical neurons, specifically focusing on gain scaling and fractional differentiation.

Methods

The researchers employed a "model-to-model" approach. They first simulated spike trains using a biophysically detailed Hodgkin-Huxley (HH) model, varying the ratios of sodium ($G_{Na}$) and potassium ($G_K$) conductances to create different "virtual neurons." They then fit a GLM to these simulated spikes. The GLM predicts spiking based on a stimulus filter (weighting incoming signals) and a spike history filter (capturing how past spikes influence future probability).

To test adaptation, the models were fed Gaussian white noise with shifting standard deviations ($\sigma = 1.0$ to $2.0$) and periodically modulated variance envelopes. Principal Component Analysis (PCA) was used to determine if the GLM's internal parameters mapped onto the HH model's physical properties.

Key Results

Emergent Gain Scaling

The study found that gain scaling in the GLM is an emergent property of the spike history filter. When the filter is sufficiently long (at least 50ms), the model naturally scales its response to stimulus variance, mimicking cortical neurons. The GLM successfully replicated the gain scaling observed in HH models, particularly when the $G_{Na}/G_K$ ratio was near 1.0.

Figure 1: Demonstration of gain scaling in biophysical simulations. Panels A-D illustrate strong gain scaling where stimulus distributions (C) and input-output relationships (D) align after scaling. Panels E-G show a regime with poor gain scaling.

Fractional Differentiation and Long-Term Memory

Beyond short-term gain scaling, the GLM also captured "fractional differentiation," where a neuron responds to the rate of change of variance. This behavior, matching experimental observations of an order $\alpha \approx 0.2$, required exceptionally long spike history filters spanning up to 10 seconds. This suggests that long-term adaptation is essentially a manifestation of long-term memory within the neuron's spiking history.

Mapping Statistics to Biophysics

PCA revealed that the GLM's filters are not arbitrary; they correlate strongly with the biophysical $G_{Na}/G_K$ ratio. This indicates that the GLM effectively captures the essence of underlying ion channel dynamics. Furthermore, the GLM disentangles the stimulus filter from spike history effects, showing that the complexity often seen in Spike-Triggered Averages (STAs) is frequently a side effect of history rather than the stimulus processing itself.

Discussion and Future Work

While the GLM proves remarkably capable, this study utilized synthetic datasets and single-compartment "point neurons." Real-world synaptic inputs change membrane conductance in ways not fully captured by injected current noise, and complex dendritic trees may perform additional adaptive computations. Future work should explore how these single-neuron adaptive properties scale within larger networks and under more realistic input conditions. Nonetheless, these results establish the GLM as a powerful, tractable tool for studying the adaptive computations of the brain.

Summary updated

Abstract

Abstract Single neurons can dynamically change the gain of their spiking responses to account for shifts in stimulus variance. Moreover, gain adaptation can occur across multiple timescales. Here, we examine the ability of a simple statistical model of spike trains, the generalized linear model (GLM), to account for these adaptive effects. The GLM describes spiking as a Poisson process whose rate depends on a linear combination of the stimulus and recent spike history. The GLM successfully replicates gain scaling observed in Hodgkin-Huxley simulations of cortical neurons that occurs when the ratio of spike-generating potassium and sodium conductances approaches one. Gain scaling in the GLM depends on the length and shape of the spike history filter. Additionally, the GLM captures adaptation that occurs over multiple timescales as a fractional derivative of the stimulus variance, which has been observed in neurons that include long timescale after hyperpolarization conductances. Fractional differentiation in GLMs requires long spike history that span several seconds. Together, these results demonstrate that the GLM provides a tractable statistical approach for examining single-neuron adaptive computations in response to changes in stimulus variance.
BibTeX
@inproceedings{latimer2020capturing,
  author = {Kenneth W. Latimer and Adrienne L. Fairhall},
  title = {{Capturing multiple timescales of adaptation to second-order statistics with generalized linear models: gain scaling and fractional differentiation}},
  booktitle = {Frontiers in Systems Neuroscience},
  year = {2020},
  doi = {10.3389/fnsys.2020.00060},
}
Figure