The Computational Properties of a Simplified Cortical Column Model

Nicholas Cain, Ramakrishnan Iyer, Christof Koch, Stefan MihalasPLOS Computational Biology 2016 · DOI

TL;DR

This paper uses a sophisticated computational model of a cortical column to investigate its fundamental computational capabilities. They discover that the model can perform linear operations and, intriguingly, a subtraction operation, which they link to theories of predictive coding and error signaling in the brain.

What Problem Does This Paper Address?

The mammalian neocortex is a highly structured and complex part of the brain responsible for higher cognitive functions. A long-standing question in neuroscience is how the specific structural organization of cortical circuits, particularly the "canonical microcircuit" within a cortical column, gives rise to these complex computations. While many models exist, this paper aims to understand the inherent computational properties of a well-established, biophysically-plausible model of a cortical column, rather than starting with a computational theory and building a model to fit it. They want to see what computations naturally emerge from this detailed model.

Methods

The authors employ a population density modeling approach using a displacement integro-partial differential equation (DiPDE) formalism. This method allows them to efficiently simulate the collective behavior of large populations of neurons, approximating the mean firing rate dynamics without the computational cost of simulating individual neurons. They adapt a previously parameterized four-layer cortical column model (Potjans and Diesmann, 2014) and apply various input patterns (step and sinusoidal) to different layers and target cell types (excitatory, inhibitory, or balanced). They analyze the resulting changes in population firing rates to characterize the model's input-output transformations, focusing on linearity, homogeneity, and additivity.

Datasets

Not discussed in the paper. The study uses a pre-existing computational model parameterized from existing literature.

Key Insights and Ideas

  • Subtraction as a Core Computation: The most striking finding is that when the model receives balanced, simultaneous inputs to layers 2/3 and 4, the output in layer 5 exhibits a subtractive relationship between these two inputs. This suggests that even a relatively simple cortical circuit can perform fundamental operations like subtraction, which is crucial for comparing different streams of information.
  • Linearity Under Balanced Input: When inputs are "balanced" (targeting both excitatory and inhibitory neurons equally), the model's response is remarkably linear. This means that the output scales proportionally with the input strength and that responses from different input layers can be added together linearly. This linearity is surprising given the recurrent nature of the network.
  • Nonlinearity with Cell-Type Specific Targeting: The linear behavior breaks down when inputs are preferentially targeted to specific cell types, particularly inhibitory neurons. This highlights the critical role of cell-type-specific connectivity in shaping cortical computations, leading to more complex, nonlinear transformations.
  • Hierarchical Processing Implications: The observed subtraction in layer 5, with layer 4 potentially representing bottom-up sensory input and layer 2/3 representing top-down predictions or internal models, aligns with theories of hierarchical predictive coding. This suggests that cortical columns might implement a mechanism for comparing sensory evidence with expectations and signaling the difference (error).
  • Low-Pass Filtering in Deep Layers: When driven by sinusoidal inputs, deeper layers (like L5) exhibit low-pass filtering characteristics, attenuating higher frequencies. This is consistent with predictions from hierarchical predictive coding theories, where information is smoothed as it passes through successive processing stages.

Key Results

  • The model demonstrates a linear regime for input-output transformations when inputs are balanced and target both excitatory and inhibitory neurons.
  • A subtractive computation is observed in layer 5 when inputs are simultaneously applied to layers 2/3 and 4 in a balanced manner, where the output reflects the difference between the two inputs.
  • Nonlinear responses emerge when inputs preferentially target specific cell types, especially inhibitory neurons.
  • Sensitivity analysis shows that the linear response under balanced input is robust to small perturbations in model parameters, suggesting it's not an artifact of fine-tuning.
  • Sinusoidal inputs reveal low-pass filtering in deeper layers and band-pass filtering in superficial layers, consistent with hierarchical processing theories.

Limitations and Future Work

The authors acknowledge that their findings are specific to the Potjans and Diesmann model and its parameterization. They suggest that future work should involve:

  • More Comprehensive Models: Expanding the model to include more cortical layers and cell types, parameterized with more detailed experimental data.
  • Experimental Validation: Combining computational models with in vivo and in silico perturbation experiments (e.g., optogenetics) to validate model predictions and refine our understanding of specific cell populations' roles.
  • Exploring Other Canonical Computations: Investigating if other population-level computational operations can be identified in more complex models.
  • Testing Predictive Coding Hypotheses: Further exploring how the observed subtractive and filtering properties can be directly mapped onto specific aspects of predictive coding theories, particularly concerning error signaling and reward prediction.

Summary updated

BibTeX
@article{cain2016computational,
  author = {Nicholas Cain and Ramakrishnan Iyer and Christof Koch and Stefan Mihalas},
  title = {{The Computational Properties of a Simplified Cortical Column Model}},
  journal = {PLOS Computational Biology},
  year = {2016},
  doi = {10.1371/journal.pcbi.1005045},
}
Figure