This NeurIPS 2021 paper by David Liu and Máté Lengyel introduces a powerful new statistical framework for understanding one of the most persistent mysteries in neuroscience: why do neurons fire so inconsistently?
Overview
The authors present the Universal Count Model (UCM), a semi-parametric framework that can model arbitrary spike count distributions without the restrictive "Poisson assumption." By leveraging sparse Gaussian processes, the UCM captures how neural variability is independently modulated by both observed behavior and hidden latent factors, revealing that variability is often as structured and informative as the firing rate itself.
The Problem with the "Poisson Assumption"
In classical neuroscience, the "firing rate" (the average number of spikes over time) is king. However, neural activity is notoriously "noisy"—even if you show a mouse the exact same stimulus twice, a neuron might fire five times in the first trial and zero in the second.
Most existing models treat this variability as a nuisance or assume it follows a simple parametric form, like a Poisson distribution (where variance equals the mean). But real neurons are often "over-dispersed" (more variable than Poisson) or "under-dispersed" (more regular). Furthermore, current methods often require many repeated trials of the same stimulus, making them useless for studying natural, ongoing behavior. The gap the authors address is the lack of a flexible, data-efficient model that can characterize arbitrary variability and its dependence on complex, non-stationary covariates.
The Universal Count Model (UCM)
The UCM is a "universal" model in the sense that it can approximate any discrete probability distribution for spike counts (from 0 to a maximum $K$).
Key technical components include:
- Sparse Gaussian Processes (GPs): These map input covariates (like an animal's head direction or speed) to the parameters of the spike count distribution. GPs allow for flexible, non-linear relationships without assuming a specific functional form.
- Basis Expansion & Softmax: The model uses a basis expansion ($\phi$) and a linear-softmax mapping to transform GP outputs into a vector of probabilities for each possible spike count.
- Latent Variables: The model can infer unobserved "latent" trajectories ($Z$) that might represent internal states like attention or arousal, which help explain why neurons might appear correlated.
- Scalable Inference: By using stochastic variational inference (SVI), the model can be trained on the massive datasets generated by modern high-density neural probes.
Key Insights and Results
The authors validated the UCM using both synthetic data and real-world electrophysiological recordings from the anterodorsal thalamic nucleus (ADn) and the postsubiculum (PoS) in freely moving mice. These neurons are "tuned" to the direction the animal's head is pointing.
The Poisson Straightjacket
The authors found that head direction cells defy simple parametric models. Specifically, these neurons tend to be under-dispersed (more regular) at high firing rates and over-dispersed at low rates—a nuance that Poisson or Negative Binomial models cannot capture.
Variability is a Signal
A major insight from the UCM is that behavioral covariates (like animal speed) can modulate a neuron's variability just as strongly as they modulate its mean firing rate. This suggests that variability isn't just noise; it's a regulated part of the neural code.
Explaining Away Correlations
By introducing a 2D latent factor, the model was able to "explain away" the noise correlations between neurons. Interestingly, while these latent factors explained why neurons fired together, they did not explain away the non-Poisson nature of the individual neurons. The inferred latent factors varied on a timescale of ~1 second, which sits right in the "gap" between fast sensory processing and slow behavioral changes, potentially representing internal cognitive drift.
Superior Fit
In cross-validated log-likelihood tests, the UCM significantly outperformed standard Poisson and heteroscedastic Negative Binomial (hNB) models. Using a modified Kolmogorov-Smirnov test, the authors showed that the UCM was the only model capable of capturing the true statistics of the thalamic head direction cells.
Limitations and Future Work
The authors acknowledge a few areas for growth:
- Hyperparameter Selection: The choice of the number of GPs ($C$) and the specific basis expansion ($\phi$) is currently empirical. Future work could involve automatic relevance determination (ARD) to select these more systematically.
- High-Dimensional Inputs: For very high-dimensional inputs (like raw video frames in sensory studies), the model might require "deep kernels" or more complex architectures to remain scalable.
- Temporal Priors: While the model uses Markovian priors for latents, more sophisticated temporal priors could better capture the complex dynamics of neural circuits.
By moving beyond the "mean firing rate" paradigm, this work provides a principled toolkit for researchers to explore how the brain represents uncertainty and dynamics through the structure of neural noise.
Summary updated