Tracing the Flow of Perceptual Features in an Algorithmic Brain Network

Robin A. A. Ince, Nicola van Rijsbergen, Gregor Thut, Guillaume A. Rousselet, Joachim Groß, Stefano Panzeri, Philippe G. SchynsScientific 2015 · DOI

Abstract

The model of the brain as an information processing machine is a profound hypothesis in which neuroscience, psychology and theory of computation are now deeply rooted. Modern neuroscience aims to model the brain as a network of densely interconnected functional nodes. However, to model the dynamic information processing mechanisms of perception and cognition, it is imperative to understand brain networks at an algorithmic level—i.e. as the information flow that network nodes code and communicate. Here, using innovative methods (Directed Feature Information), we reconstructed examples of possible algorithmic brain networks that code and communicate the specific features underlying two distinct perceptions of the same ambiguous picture. In each observer, we identified a network architecture comprising one occipito-temporal hub where the features underlying both perceptual decisions dynamically converge. Our focus on detailed information flow represents an important step towards a new brain algorithmics to model the mechanisms of perception and cognition.

Introduction

Contemporary neuroscience frequently models the brain as a complex network of interconnected functional nodes. While tools such as Transfer Entropy and Granger Causality have successfully mapped the "functional connectivity" or the "plumbing" of these networks—identifying which regions influence others—they often fail to address the "content gap." Specifically, traditional measures do not reveal the nature of the information being exchanged. To transition from a structural map to an algorithmic understanding of the brain, researchers must identify the specific perceptual features (e.g., color, shape, or texture) flowing through these neural circuits. This study introduces a novel metric to bridge this gap and map the precise flow of visual information during perception.

Methodology: Directed Feature Information (DFI)

To isolate the communication of specific stimulus features, the researchers developed Directed Feature Information (DFI). This metric refines standard Directed Information (DI) by subtracting the communication that occurs independently of a specific stimulus feature ($F$). The resulting formula, $DFI = DI - DI|F$, effectively isolates the "data" relevant to the task from background neural "handshaking."

Mutual Information (MI) over time, quantifying the degree to which a specific brain node codes a particular visual feature.

The experimental design utilized Salvador Dali’s Slave Market with Disappearing Bust of Voltaire, an ambiguous image that can be perceived as either two nuns or the bust of Voltaire. Using a "Bubbles" sampling technique, the researchers presented random fragments of the image across thousands of trials to five human observers. Brain activity was recorded via Magnetoencephalography (MEG), and Independent Component Analysis (ICA) was employed to isolate distinct functional nodes. Mutual Information (MI) was then used to quantify the coding of specific visual features within these nodes.

Results: Spatiotemporal Dynamics of Feature Integration

The application of DFI revealed that algorithmic communication is significantly sparser than general functional connectivity. Only 5% to 20% of significant neural connections were found to carry specific information relevant to the perceptual task.

Directed Information (DI) plot illustrating the temporal flow and delay of information between neural nodes.

The study identified a consistent temporal sequence in information processing:

  1. Initial Coding: Feature coding was first detected between 110–200ms post-stimulus, primarily in contra-lateral early visual areas.
  2. Information Flow: This information flowed toward specialized hubs between 140–300ms.
  3. Integration Hub: In all observers, features from both visual fields converged on a common hub in the right occipito-temporal region, a key component of the ventral "what" pathway.
  4. Decision Dispatch: Integrated information was finally transmitted to inferior parietal regions around 300ms, likely for final decision-making.

Crucially, the DFI networks for the "Nuns" perception differed from those for the "Voltaire" perception, despite the physical stimulus remaining identical, highlighting the algorithmic specificity of the brain's response to ambiguous input.

Discussion and Future Directions

This research represents a significant shift from identifying where brain activity occurs to understanding what information is being processed and communicated. While DFI currently provides a lower-bound estimate of information flow and assumes stimulus-dependent processing changes, it offers a powerful framework for reverse-engineering brain algorithms. Future work may expand this approach to higher-level generative features and other sensory modalities, moving closer to a comprehensive "algorithmics" of the human mind.

Summary updated

Abstract

The model of the brain as an information processing machine is a profound hypothesis in which neuroscience, psychology and theory of computation are now deeply rooted. Modern neuroscience aims to model the brain as a network of densely interconnected functional nodes. However, to model the dynamic information processing mechanisms of perception and cognition, it is imperative to understand brain networks at an algorithmic level--i.e. as the information flow that network nodes code and communicate. Here, using innovative methods (Directed Feature Information), we reconstructed examples of possible algorithmic brain networks that code and communicate the specific features underlying two distinct perceptions of the same ambiguous picture. In each observer, we identified a network architecture comprising one occipito-temporal hub where the features underlying both perceptual decisions dynamically converge. Our focus on detailed information flow represents an important step towards a new brain algorithmics to model the mechanisms of perception and cognition.
BibTeX
@inproceedings{ince2015tracing,
  author = {Robin A. A. Ince and Nicola van Rijsbergen and Gregor Thut and Guillaume A. Rousselet and Joachim Groß and Stefano Panzeri and Philippe G. Schyns},
  title = {{Tracing the Flow of Perceptual Features in an Algorithmic Brain Network}},
  booktitle = {Scientific},
  year = {2015},
  doi = {10.1038/srep17681},
}
Figure