Desiderata of evidence for representation in neuroscience (2403.14046v1)

Abstract: This paper develops a systematic framework for the evidence neuroscientists use to establish whether a neural response represents a feature. Researchers try to establish that the neural response is (1) sensitive and (2) specific to the feature, (3) invariant to other features, and (4) functional, which means that it is used downstream in the brain. We formalize these desiderata in information-theoretic terms. This formalism allows us to precisely state the desiderata while unifying the different analysis methods used in neuroscience under one framework. We discuss how common methods such as correlational analyses, decoding and encoding models, representational similarity analysis, and tests of statistical dependence are used to evaluate the desiderata. In doing so, we provide a common terminology to researchers that helps to clarify disagreements, to compare and integrate results across studies and research groups, and to identify when evidence might be missing and when evidence for some representational conclusion is strong. We illustrate the framework with several canonical examples, including the representation of orientation, numerosity, faces, and spatial location. We end by discussing how the framework can be extended to cover models of the neural code, multi-stage models, and other domains.

• The paper introduces an information-theoretic framework that defines sensitivity, specificity, invariance, and functionality as core criteria for establishing neural representations.
• It demonstrates how common analysis methods like decoding and encoding models align with these criteria using canonical examples such as visual stimulus orientation and spatial location.
• The study fosters interdisciplinary dialogue by standardizing terminology, enabling clearer comparison of evidential claims across neuroscience and AI research.

An Information-Theoretic Framework for Neural Representation Evidence

The paper by Pohl et al. presents a comprehensive framework designed to systematize the evidence supporting representational claims within neuroscience. It formalizes the criteria necessary for determining whether a neural response serves as a representation of a particular environmental feature. These criteria, referred to as desiderata, include sensitivity, specificity, invariance, and functionality. By leveraging information theory, the authors provide a unified framework that enables researchers to articulate and compare evidential claims across various studies.

Core Desiderata

The framework identifies four core desiderata that are necessary to establish a neural representation:

1. Sensitivity: A neural response is sensitive to a feature if it carries substantial information about that feature. This is quantified as the mutual information between the neural response and the feature, normalized by the entropy of the feature.
2. Specificity: A neural response is specific to a feature if the variability of the response is largely explained by changes in that feature. This desideratum is represented by mutating information between the neural response and the feature, normalized by the entropy of the neural response.
3. Invariance: Invariance is achieved when a neural response remains constant across variations of other features that are not of interest. It is defined as the degree to which a neural response to a feature remains unchanged irrespective of other confounding features.
4. Functionality: For a neural representation to be considered functional, it must causally impact downstream processes, such as influencing a behavioral outcome. This involves demonstrating that the information carried by the neural response actively contributes to subsequent cognitive processes.

Application and Empirical Illustrations

The paper outlines how common analysis methods in neuroscience, such as encoding and decoding models, representational similarity analysis, and tests of statistical dependence, relate to these desiderata. For instance, decoding models are associated with sensitivity and are often used to gauge how well a neural response can predict an external feature. Conversely, encoding models assess specificity by examining the extent to which features explain variations in neural activity.

Several canonical examples illustrate the implementation of these desiderata:

• Orientation of Visual Stimuli: Hubel and Wiesel's work on V1 neurons demonstrated sensitivity and specificity to stimulus orientation.
• Numerosity: Neurons in parietal and prefrontal cortices exhibit sensitivity to numerical quantities, showcasing specificity and task-dependent sensitivity.
• Face Perception: The fusiform face area (FFA) exemplifies the sensitivity and specificity of neural representations for facial recognition.
• Spatial Location: The function of place cells in the hippocampus displays high specificity and invariance with respect to environmental cues.

Extensions and Implications

The authors explore extensions of the framework to other domains, such as mental simulations, planning, and the representation of uncertainty, which suggests that the same principles can be adapted to understand cognitive processes not directly tied to real-world states. The framework also aligns with philosophical considerations of representation and positions itself as a tool to facilitate interdisciplinary dialogue.

Beyond the theoretical contributions, the paper calls for a systematic use of common terminology to improve communication between research groups, allowing for a more integrated understanding of evidence across neuroscience.

Conclusion

Overall, Pohl et al.'s paper offers a precise and robust framework defined in information-theoretic terms for evidential claims about neural representation. While its core contribution is theoretical, the framework has practical implications for broader applications in neuroscience and AI. By fostering a clearer articulation of representational evidence, this work serves as a foundational reference, encouraging more rigorous and consistent approaches to studying representations in neural systems.