Signatures of criticality arise from random subsampling in simple population models

Marcel Nonnenmacher, Christian Behrens, Philipp Berens, Matthias Bethge, Jakob H. MackePLOS Computational Biology 2017 · OpenReview

TL;DR

This paper challenges the prevailing idea that neural populations are "optimized" to operate at critical points, a concept borrowed from physics. The authors demonstrate that common analysis techniques, when applied to simple neural models with realistic correlations, can artifactually produce these "criticality signatures" due to random data subsampling. This suggests that observed criticality might not reflect a special computational state but rather a consequence of how data is analyzed.

What Problem Does This Paper Address?

A significant line of research in computational neuroscience has explored the hypothesis that neural populations, particularly in the retina, are poised at a "thermodynamic critical point." This idea stems from observations in physics where systems near critical points exhibit unique properties beneficial for information processing. Studies analyzing neural data have reported "signatures of criticality," such as diverging specific heat, which have been interpreted as evidence for this optimized state. However, the authors question whether these signatures truly reflect an underlying biological principle or are an artifact of the analytical methods used. Specifically, they investigate whether simple, non-critical models, when analyzed with standard techniques, can produce these same signatures.

Methods

The core of the paper involves building and analyzing simplified models of neural population activity.

  • Phenomenological LN Cascade Model: A feed-forward model where neurons have linear receptive fields, followed by a non-linear activation and stochastic spiking. Crucially, this model includes shared Gaussian noise, introducing correlations between neurons without relying on complex recurrent connections.
  • Maximum Entropy Models (K-pairwise): These models are fitted to neural data (simulated or real) to capture statistical properties like firing rates and pairwise correlations. The authors use a "K-pairwise" model, which extends standard pairwise models to also capture population spike counts.
  • Temperature Parameter: A concept borrowed from statistical physics is introduced, allowing the authors to perturb the fitted models and observe how their statistical properties change.
  • Specific Heat Calculation: This thermodynamic quantity is calculated from the variance of log-probabilities of spike patterns, serving as a key indicator of criticality.
  • Beta-Binomial Model: A simplified "flat" model where all neurons share the same firing rate and correlation, and only the population spike count distribution matters. This model allows for analytical tractability.
  • Rao-Blackwellisation: An algorithmic optimization technique used to speed up the estimation of model parameters in maximum entropy models, particularly for large populations.

Datasets

  • Simulated Retinal Ganglion Cell (RGC) Activity: The primary dataset used for analysis was generated from a custom simulation of RGCs, incorporating realistic receptive field properties, shared noise, and natural image stimuli.
  • Published RGC Data: The authors also reference and compare their findings to previously published empirical data from retinal ganglion cell recordings (e.g., from Tkačik et al. and Mora et al.).

Key Insights and Ideas

  • Artifactual Criticality: The most striking finding is that simple feed-forward models with realistic correlations, when analyzed using standard techniques (like random subsampling and maximum entropy fitting), exhibit signatures of criticality. This suggests that these signatures may not be unique to optimized critical states but can arise from the analysis process itself.
  • Correlations are Key: The presence of neural correlations, whether from shared stimulus drive, common input noise, or recurrent connectivity, is shown to be a crucial factor in generating these criticality signatures.
  • Subsampling Matters: The way neural data is subsampled for analysis significantly impacts the observed statistics. Random subsampling of a correlated population can artificially inflate measures of criticality.
  • Challenging the Optimization Hypothesis: The results directly challenge the interpretation that observed criticality in neural systems implies an evolutionary optimization for operating at a critical point for enhanced computation.
  • Flat Models Capture Signatures: Even highly simplified "flat" models (where all neurons are statistically identical) can reproduce criticality signatures, highlighting the power of correlations and analysis methods over complex neural mechanisms.

Key Results

  • A simple feed-forward LN cascade model of RGC activity, incorporating shared noise, reproduced the observed signatures of criticality (diverging specific heat, peak near unit temperature) when analyzed using standard methods.
  • The specific heat divergence rate in flat models was found to increase monotonically with the strength of correlations.
  • The specific heat divergence observed in experimental data across different stimulus conditions (checkerboard, natural images, full-field flicker) was qualitatively reproduced by the model, correlating with the strength of neural correlations induced by each stimulus.
  • The authors developed and implemented efficient algorithms (using Rao-Blackwellisation) for fitting complex maximum entropy models to large neural populations, making these analyses more feasible.

Limitations and Future Work

  • Model Simplification: While the LN cascade model is simple, it captures key aspects like receptive fields and shared noise. However, it doesn't include recurrent connections or more complex neuronal dynamics, which could also contribute to observed phenomena.
  • Interpretation of "Criticality": The paper focuses on specific statistical signatures (diverging specific heat) derived from equilibrium thermodynamics. It acknowledges that other definitions of criticality exist (e.g., based on temporal dynamics) and that their findings might not apply to those.
  • Future Directions: The authors suggest that future work should focus on developing data-analysis methods better adapted to the specific properties of biological neural data, moving beyond equilibrium thermodynamics. They also imply that understanding the functional consequences of these observed statistical properties, whether critical or not, remains a key challenge.

Summary updated

Abstract

Author summary Understanding how populations of neurons collectively encode sensory information is one of the central goals of computational neuroscience. In physics, systems are often characterized by identifying and describing critical points (e.g. the transition between two states of matter). The success of this approach has inspired a series of studies to search for analogous phenomena in nervous systems, and has lead to the hypothesis that these might be optimized to be poised at ‘thermodynamic critical points’. However, translating concepts from thermodynamics to neural data analysis has been a challenging endeavour. We here study the data analysis approaches that have been used to provide evidence for criticality in the brain. We find that observing signatures of criticality is closely linked to observing activity correlations between neurons– a ubiquitous phenomenon in neural data. Our study questions the experimental evidence that neural systems are optimised to exhibit thermodynamic critical behaviour. Finally, we provide practical, open-source tools for analyzing large-scale measurements of neural population activity using maximum entropy models.
BibTeX
@article{nonnenmacher2017signatures,
  author = {Marcel Nonnenmacher and Christian Behrens and Philipp Berens and Matthias Bethge and Jakob H. Macke},
  title = {{Signatures of criticality arise from random subsampling in simple population models}},
  journal = {PLOS Computational Biology},
  year = {2017},
  url = {https://openreview.net/forum?id=rnDHh2yh_h},
}
Figure