Time-rescaling methods for the estimation and assessment of non-Poisson neural encoding models

Jonathan W. Pillow

Overview

This paper introduces the Conditional Renewal (CR) model, a framework that unifies stimulus-driven rate models with non-Poisson history dependencies. It provides a rigorous mathematical proof for when these models can be efficiently estimated via convex optimization. Crucially, the study reveals a significant warning for computational neuroscientists: standard "goodness-of-fit" tests can be dangerously misleading when evaluating how well a model actually captures stimulus-response relationships.

The Problem: Beyond Poisson Models

In computational neuroscience, the "gold standard" for modeling neural responses has long been the Linear-Nonlinear-Poisson (LNP) model. Its premise is simple: filter a stimulus, pass it through a nonlinearity, and generate spikes randomly according to a Poisson process. However, real neurons are rarely perfectly Poisson; they possess "memory." They exhibit refractory periods (an inability to fire twice in the same instant), bursting, and adaptation.

Researchers have typically addressed these non-Poisson behaviors in one of two ways:

  1. Generalized Linear Models (GLMs): The current firing rate depends on the history of previous spikes.
  2. Renewal Processes: The time until the next spike depends on the time since the last spike, often modeled using non-exponential distributions like the Gamma distribution.

Historically, these two approaches were rarely unified into a single, tractable framework. Furthermore, the field lacked a clear understanding of when these complex models could be reliably "fit" to data without optimization algorithms getting stuck in local optima.

The Conditional Renewal (CR) Model

To bridge this gap, the author proposes the Conditional Renewal (CR) model. In this framework, the "base intensity" (the drive to fire) is determined by both the external stimulus and the recent spiking history—similar to the GLM approach. This intensity is then used to "rescale" time. In this rescaled time, the inter-spike intervals (ISIs) are drawn from a non-exponential renewal density.

Conditional Renewal Model Schematic and Time-Rescaling

Key methodological contributions include:

  • Convexity Analysis: The paper provides a mathematical proof showing that the log-likelihood of the model is concave—meaning we can reliably find the globally optimal parameters—only if the renewal density is log-concave and non-increasing.
  • Nonparametric Estimation: To avoid assuming a specific shape (like a Gamma or Inverse-Gaussian distribution), the author uses cubic splines to estimate the renewal density directly from the data.
  • Time-Rescaling Theorem: This acts as the "glue" that allows the model to map irregular, stimulus-driven spikes into a simplified space where their statistical properties can be easily measured and compared.

The "Goodness-of-Fit" Trap

Perhaps the most striking insight from the paper is its demonstration of the "Goodness-of-Fit" trap. The author shows that a model can pass a Kolmogorov-Smirnov (KS) test—a standard metric in the field—with flying colors while being a terrible predictor of the stimulus.

Simulation Results: KS Plot vs Cross-Validation

Using simulated neural data, the study demonstrates that a model with a completely "inverted" (wrong) stimulus filter (model 'c' in the figure above) can pass the time-rescaling KS test simply by adjusting its renewal density to compensate. However, its predictive information (cross-validation bits/s) is negative, proving the KS test's insensitivity to actual stimulus-response accuracy. If you get the timing statistics right but the stimulus filter wrong, the KS test won't necessarily warn you. This suggests that cross-validated log-likelihood is a much more robust metric for model accuracy.

Application to Retinal Data

The paper validates the CR model using spiking responses from an ON parasol cell in a primate retina, stimulated with binary spatio-temporal white noise.

The CR model successfully captured the fine-grained timing of these retinal ganglion cells. Interestingly, the study found that for these specific neurons, the history-dependent rate (the GLM component) captures almost all the relevant "memory" of the system. Adding a history filter to a Poisson model provided a 14% boost in predictive power, whereas adding non-Poisson renewal properties to a model that already had a history filter provided less than a 1% improvement. While the extra complexity of non-Poisson renewal densities helps pass statistical tests, it was largely unnecessary for actual prediction in this dataset.

Limitations and Future Work

The primary limitation of the CR model lies in the convexity requirement. Because many biologically realistic renewal densities (such as those modeling absolute refractory periods) are not log-concave and non-increasing, fitting them remains a non-convex optimization problem prone to local maxima. The proof reveals a trade-off: if you want a model that is easy to optimize, your renewal density cannot account for refractory periods. This implies that real-time history filters are actually a more mathematically "well-behaved" way to model refractoriness.

Future work should explore larger datasets and different brain areas to see if the modest gains seen in the retina from non-Poisson renewal properties hold true elsewhere, or if other neuron types rely more heavily on rescaled-time history effects.

Summary updated

Abstract

Recent work on the statistical modeling of neural responses has focused on modulated renewal processes in which the spike rate is a function of the stimulus and recent spiking history. Typically, these models incorporate spike-history dependencies via either: (A) a conditionally-Poisson process with rate dependent on a linear projection of the spike train history (e.g., generalized linear model); or (B) a modulated non-Poisson renewal process (e.g., inhomogeneous gamma process). Here we show that the two approaches can be combined, resulting in a {\it conditional renewal} (CR) model for neural spike trains. This model captures both real and rescaled-time effects, and can be fit by maximum likelihood using a simple application of the time-rescaling theorem [1]. We show that for any modulated renewal process model, the log-likelihood is concave in the linear filter parameters only under certain restrictive conditions on the renewal density (ruling out many popular choices, e.g. gamma with $\kappa \neq1$), suggesting that real-time history effects are easier to estimate than non-Poisson renewal properties. Moreover, we show that goodness-of-fit tests based on the time-rescaling theorem [1] quantify relative-time effects, but do not reliably assess accuracy in spike prediction or stimulus-response modeling. We illustrate the CR model with applications to both real and simulated neural data.
BibTeX
@inproceedings{pillow2009time,
  author = {Jonathan W. Pillow},
  title = {{Time-rescaling methods for the estimation and assessment of non-Poisson neural encoding models}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2009},
}
Figure