From physiological principles to computational models of the cortex

Jérémy Fix, Raphaël Monat, Frédéric AlexandreJournal of Physiology-Paris 2007 · DOI

TL;DR

This paper proposes a multi-level computational framework for understanding the cortex, bridging physiological principles with computational models. By integrating data from microscopic neuronal activity to macroscopic behavior, the authors aim to build more biologically plausible and functional artificial systems, particularly for tasks like visual search.

What Problem Does This Paper Address?

Understanding the human brain is a monumental task, with vast amounts of data emerging from various levels of analysis: single-cell recordings (microscopic), brain imaging (mesoscopic), and behavioral studies (macroscopic). The challenge lies in integrating these disparate levels of information into a coherent understanding of cognition and, crucially, in using this understanding to build more sophisticated computational models. Existing models often focus on one level, leading to a disconnect between biological realism and functional performance. This paper seeks to bridge this gap by proposing a unified approach that leverages insights from all three levels to create computational models that are both biologically inspired and capable of performing complex tasks.

Methods

The authors propose a hierarchical approach to computational modeling, working across three levels:

  • Microscopic Level: This level focuses on the elementary units of computation, often inspired by neuronal models. They discuss the trade-offs between detailed spiking neuron models and more abstract mean-firing rate models, opting for the latter for computational tractability when dealing with large networks. This level also considers how individual neurons are modulated by attention, drawing on concepts like contrast gain.
  • Mesoscopic Level: This level deals with the organization of the cortex into regions and the information flow between them. The authors discuss the ventral and dorsal pathways in visual processing and how these are represented computationally. They emphasize the importance of both feedforward and feedback connections in modeling these interactions.
  • Macroscopic Level: This level focuses on the overall task or behavior being modeled. Computational models at this level are used to simulate complex behaviors like visual search and to generate predictions that can be tested against psychological and behavioral data.

The core methodology involves building computational models that integrate these levels. They propose using assemblies of units organized into "maps," which communicate through local, distributed, and asynchronous computations. This architecture is designed to mimic the brain's distributed processing and emergent behavior. They illustrate this with a model of visual attention and search, incorporating feature-based and spatial attention mechanisms, working memory, and anticipation of eye movements.

Datasets

Not discussed in the paper. The paper focuses on synthesizing existing knowledge from neuroscience and psychology rather than training models on specific datasets.

Key Insights and Ideas

  • Multi-level Integration is Crucial: A truly biologically inspired computational model must integrate information from the microscopic (neuronal) to the macroscopic (behavioral) levels.
  • Computational Tractability vs. Biological Detail: While spiking neuron models are more biologically realistic at the microscopic level, mean-firing rate models are necessary for simulating large-scale cortical networks. This highlights a fundamental trade-off in computational neuroscience.
  • Feedback is Key for Attention: Attentional mechanisms, particularly feature-based and spatial attention, are heavily influenced by feedback projections, suggesting that top-down influences play a critical role in shaping sensory processing.
  • Distributed and Asynchronous Computation: The brain's computational power arises from the interaction of many simple, distributed units operating asynchronously, leading to emergent behaviors. This principle is central to their modeling approach.
  • Modeling as a Tool for Hypothesis Testing: Computational models are not just for replicating brain function but also for testing hypotheses about neural mechanisms and cognitive processes, providing a rigorous way to assess theoretical ideas.

Key Results

The paper doesn't present novel experimental results but rather demonstrates how existing knowledge can be integrated into computational models. The key "results" are the successful conceptualization and illustration of a multi-level modeling framework.

  • Visual Search Model: They describe a computational model that can perform visual search tasks, distinguishing between feature search (fast, "pop-out") and conjunction search (slower, requiring more systematic scanning).
  • Simulation of Behavioral Phenomena: The model, when treated as a "black box," can reproduce observed behavioral patterns, such as the difference in reaction times between feature and conjunction search as a function of the number of distractors (illustrated in Figure 5).
  • Scanpath Generation: The model can generate realistic scanpaths (sequences of eye movements) during visual search tasks, demonstrating its ability to integrate visual perception with motor control (illustrated in Figure 6).

Limitations and Future Work

  • Simplification of Neuronal Models: The use of mean-firing rate models, while computationally advantageous, means that phenomena occurring at the level of individual spikes are not captured. The authors acknowledge this as a potential limitation but argue for its functional relevance at higher levels.
  • Exact Mechanisms of Integration: While the paper outlines the general framework, the precise details of how information is integrated across different maps and pathways are complex and not fully elaborated, with the authors referring readers to other works for complete descriptions.
  • Learning Rules: The authors identify the development of learning rules to describe the mutual influence of information flows as a crucial area for future work at the mesoscopic level.
  • Embodied Cognition: They suggest that applying these models to autonomous agents like robots, enabling embodied cognition, is a promising direction for future research.
  • Refining Descriptions: The paper emphasizes the opportunity to refine the level of description in computational models and compare them with more precise observations, suggesting an iterative process of model development and validation.

Summary updated

BibTeX
@article{fix2007from,
  author = {Jérémy Fix and Raphaël Monat and Frédéric Alexandre},
  title = {{From physiological principles to computational models of the cortex}},
  journal = {Journal of Physiology-Paris},
  year = {2007},
  doi = {10.1016/j.jphysparis.2007.10.009},
}
Figure