Is it that simple? The use of linear models in cognitive neuroscience

Anna Ivanova, Martin Schrimpf, Leyla Isik, Stefano Anzellotti, Noga Zaslavsky, Evelina FedorenkoOpenReview 2020 · OpenReview

TL;DR

This paper questions the pervasive reliance on linear models in cognitive neuroscience, arguing that their simplicity might be hindering deeper insights into brain function. The authors propose a comprehensive approach to systematically evaluate the validity, benefits, and limitations of both linear and nonlinear models, aiming to guide future research towards more accurate and biologically plausible interpretations of brain activity.

What Problem Does This Paper Address?

Cognitive neuroscience heavily relies on linear models to map brain activity to stimuli, behaviors, or other brain states. This reliance stems from the perceived neural plausibility and ease of use of linear models. However, the authors highlight a growing tension: neural computations are inherently nonlinear, and even indirect measures of brain activity (like fMRI signals) might not accurately reflect linear processes. While prediction-oriented studies are increasingly adopting nonlinear models, the "explanation-oriented" core of cognitive neuroscience remains largely linear. This creates a gap, potentially causing the field to overlook crucial aspects of brain function by unnecessarily restricting the modeling toolkit. The paper aims to bridge this gap by rigorously assessing when linear models suffice and when nonlinear approaches are essential.

Methods

The proposed approach is two-pronged:

  1. Theory Branch:
    • Synthesize existing literature on the "linear readout assumption" in neuroscience.
  2. Introduce novel information-theoretic criteria for evaluating the performance and suitability of linear versus nonlinear models in neuroimaging research.
  3. Establish the theoretical validity of the linear readout assumptions in the context of neuroimaging data.
  4. Empirical Branch:
    • Evaluate practical limitations of both model types, such as data requirements and the complexity of features they can handle.
  5. Develop theory-driven metrics for systematic comparison.
  6. Compare the predictive and explanatory power of linear and nonlinear models across three key domains:
  7. Mapping stimulus features to neural activity.
  8. Mapping neural activity to behavior.
  9. Mapping neural activity between brain regions.

Datasets

The paper does not explicitly list specific datasets used for training or evaluation. Instead, it proposes to evaluate models on "existing datasets" within the empirical branch of their approach.

Key Insights and Ideas

  • The "Linear Readout Assumption" is a double-edged sword: While it simplifies analysis and is often considered neurally plausible, it might be an oversimplification that masks true neural complexity.
  • A growing divide: There's a divergence between prediction-focused studies (using nonlinear models) and explanation-focused studies (sticking to linear models), which this work aims to reconcile.
  • Beyond individual neurons: Even if individual neurons exhibit linear behavior, the aggregated signals from larger populations or indirect measures like fMRI might not.
  • The risk of stagnation: An over-reliance on linear models could be hindering the field's progress, preventing the discovery of important insights into brain function.
  • A call for systematic evaluation: The paper advocates for a rigorous, principled framework to compare linear and nonlinear models, moving beyond ad-hoc applications.

Key Results

This paper is a proposal for future research, not a report of completed experiments. Therefore, there are no quantitative results to summarize at this stage. The "Concrete outcomes" section outlines the expected deliverables:

  1. An information-theoretic framework for evaluating linear vs. nonlinear models in neuroimaging.
  2. Guidelines for using linear vs. nonlinear models based on theoretical and empirical findings.
  3. An online platform for researchers to compare models systematically.

Limitations and Future Work

The paper acknowledges the potential for nonlinear models to be complex and difficult to interpret, cautioning neuroscientists to be careful in their application. The primary future work is the execution of the proposed two-pronged approach (theory and empirical branches) to systematically evaluate linear and nonlinear models. The authors believe this work will either:

  • Demonstrate the continued utility of linear models, saving researchers time and resources.
  • Show the benefits of nonlinear models in specific cases, potentially catalyzing new lines of inquiry in cognitive computational neuroscience.

Summary updated

Abstract

Title: Is it that simple? The use of linear models in cognitive neuroscience Scientific question: Do linear models provide an accurate, interpretable, and biologically plausible description of brain activity?
BibTeX
@misc{ivanova2020is,
  author = {Anna Ivanova and Martin Schrimpf and Leyla Isik and Stefano Anzellotti and Noga Zaslavsky and Evelina Fedorenko},
  title = {{Is it that simple? The use of linear models in cognitive neuroscience}},
  year = {2020},
  url = {https://openreview.net/forum?id=-o0dOwashib},
  howpublished = {OpenReview},
}
Figure