Analog Computation at a Critical Point: A Novel Function for Neuronal Oscillations?

NeurIPS 1990

TL;DR

This paper proposes that regularly firing neurons, when operating at a critical point, can act as analog computers. They demonstrate how the spatial dimensions of external stimuli can be encoded in the temporal correlations of individual neurons, offering a novel mechanism for visual processing.

What Problem Does This Paper Address?

The paper tackles the question of how biological neural systems, particularly the visual cortex, might perform complex computations, such as recognizing object dimensions, using local interactions. Traditional models often struggle to explain how systems with short-range connections can achieve long-range coordination and process global spatial information. The authors are motivated by the observation of oscillatory firing in the visual cortex and propose that operating at a "critical point" – a state of matter where systems exhibit long-range correlations – could be the key. This critical state allows for extreme sensitivity to external perturbations and a unique way to encode information.

Methods

The authors build a theoretical model by mapping regularly firing neurons to a simplified statistical mechanics system: the 2D XY model. In this model, neurons are represented as unit-length planar spins, and their interactions are described by an effective Hamiltonian derived from simulating small clusters of Fitzhugh-Nagumo neurons. The core idea is that the dynamics of these spins at a critical point exhibit "dynamic scaling," which allows spatial properties of input signals to be translated into temporal correlations within individual neurons. They then use numerical simulations of these spin systems and Fitzhugh-Nagumo neuron networks to validate their predictions.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and numerical simulations rather than empirical datasets.

Key Insights and Ideas

  • Neurons as Critical Systems: The paper posits that regularly firing neurons can be viewed as systems operating at a critical point, similar to phase transitions in physics. This critical state is crucial for their proposed computational mechanism.
  • Spatial Information in Temporal Correlations: The most striking insight is that the spatial characteristics of an "object" (like its dimensions) can be encoded in the temporal correlation functions of individual neurons. This means a single neuron can "sense" the size of a stimulus it's part of.
  • Dynamic Scaling as a Computational Tool: Dynamic scaling, a phenomenon at critical points, allows for a trade-off between spatial and temporal information. This enables the system to represent spatial extent through temporal patterns.
  • Robustness of Critical Computation: Unlike systems requiring delicate fine-tuning, networks operating at a critical point can perform these computations robustly, making it biologically plausible.
  • Mapping Neurons to Spins: The paper provides a rigorous connection between the dynamics of realistic spiking neurons (Fitzhugh-Nagumo model) and the well-understood XY model of statistical mechanics, particularly when neurons exhibit regular firing.

Key Results

  • The authors demonstrate that the temporal auto-correlation functions of individual cells in simulated arrays of different sizes (1x1, 3x3, 5x5, 7x7) show distinct patterns that depend on the array's dimensions.
  • They show that the correlation length predicted by their effective Hamiltonian model for the spin system closely matches the correlation lengths observed in simulations of spiking neurons, with no free parameters.
  • The model predicts that the temporal correlations of individual neurons can encode object dimensions and topology in a scale-independent manner, meaning it works for both small and large objects.

Limitations and Future Work

  • Idealized Neurons: The model uses simplified representations of neurons (planar spins, Fitzhugh-Nagumo model) and assumes regular firing, which may not fully capture the complexity of biological neurons.
  • Experimental Validation: While the simulations are promising, direct experimental validation in biological systems is the next crucial step. The authors suggest direct analogs of their numerical experiments in the visual cortex.
  • More Complex Stimuli: The current work focuses on simple spatial extents. Future work could explore how more complex shapes, textures, and dynamic stimuli are encoded.
  • Biological Plausibility of Criticality: While the paper argues for the plausibility of critical operation, further research is needed to confirm if biological neural networks naturally operate at or near critical points.

Summary updated

BibTeX
@inproceedings{1990analog,
  title = {{Analog Computation at a Critical Point: A Novel Function for Neuronal Oscillations?}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure