Myopic control of neural dynamics

David Hocker, Il Memming ParkPLOS Computational Biology 2019 · OpenReview

TL;DR

This paper introduces "myopic control," a novel approach to stimulating neural systems that aims to make them behave like a different target dynamical system. Unlike traditional methods that force a system along a predefined path, myopic control focuses on steering the system's immediate future dynamics, allowing for more natural and flexible behavior, which is crucial for both understanding brain function and treating neurological disorders.

What Problem Does This Paper Address?

Current methods for stimulating neural systems, whether for research or therapeutic purposes, often struggle to account for the inherent complexity and variability of neural dynamics. Traditional control approaches, like trajectory tracking, impose rigid paths that can counteract natural neural fluctuations, hindering the system's ability to perform computations or adapt flexibly. This is particularly problematic when trying to restore healthy function to a diseased brain or to precisely probe how neural systems process information. The authors identify a gap in control strategies that can effectively guide a neural system to adopt the dynamics of another system without stifling its natural, moment-to-moment operations.

Methods

The core of the paper is the development of "myopic control," a control strategy formulated within a nonlinear state-space feedback framework. The key innovation is a cost function that minimizes the difference between the current system's dynamics and a target dynamical system's dynamics over a single time step. This "myopic" approach avoids the need for long-term trajectory planning, making it computationally less demanding and more adaptable to the unpredictable nature of neural systems.

The authors derive both linear and nonlinear versions of the myopic controller. For linear systems, the controller depends only on the mean of the state distribution. For nonlinear systems, they employ a moment expansion approximation, deriving first-order and second-order controllers that depend on the mean and covariance of the state distribution, respectively. They also propose design principles for creating target dynamical systems, often involving multiplicative filters and additive barrier functions to shape the desired dynamics.

Datasets

Not discussed in the paper. The paper focuses on simulating dynamical systems and applying the proposed control method.

Key Insights and Ideas

  • "Myopic" control is key: The core idea is that controlling neural systems requires a focus on the immediate future dynamics rather than a rigid, pre-planned trajectory. This allows the system to maintain its inherent flexibility.
  • Mimicking dynamics, not just states: The goal is to make one dynamical system behave like another, not just reach specific states. This is a more powerful concept for understanding and manipulating complex biological systems.
  • Design principles for target dynamics: The paper offers a framework for constructing desired neural dynamics, suggesting methods like filtering and barrier functions to shape system behavior. This shifts the focus from just controlling a system to actively designing its functional properties.
  • First-order control is often sufficient: Surprisingly, the simpler first-order myopic controller often performs comparably to the more complex second-order controller, suggesting that for many practical applications, a less computationally intensive approach is viable.

Key Results

The paper demonstrates the effectiveness of myopic control through two main examples:

  1. Winner-Take-All to Robust Neural Integrator: They successfully transformed a decision-making system (Winner-Take-All) into a robust neural integrator. In a simulated "change-of-mind" stimulus task, myopic control achieved 92% accuracy, significantly outperforming trajectory control (0% accuracy) and closely matching the target RNI dynamics (91% accuracy).
  2. Diseased Motor System to Healthy System: They controlled a simulated "diseased" motor system exhibiting an unwanted beta-oscillation spiral attractor, guiding it towards healthy dynamics with stable attractors. Performance metrics (log-cost) showed a significant improvement after control was applied, effectively removing the disease-like oscillatory behavior.

Limitations and Future Work

  • Known Dynamics Assumption: The current formulation assumes the controlled and target dynamics are known. Future work should integrate adaptive estimation of these dynamics, potentially using deep neural networks.
  • State Estimation: While the controller is designed to be separable from state estimation, accurate estimation of neural states and dynamics is crucial for real-time implementation.
  • Parameter Tuning: Designing target dynamical systems, especially for complex behaviors like the robust neural integrator, can require significant parameter tuning.
  • High-Dimensional Systems: The current examples are two-dimensional. Extending the filtering and analysis tools to higher-dimensional systems is a necessary next step.
  • Experimental Implementation: The authors highlight the need for quality neural measurements and fast processing for real-time implementation, suggesting that myopic control is promising for its relatively low computational demands and demonstrated performance with realistic time lags.

Summary updated

Abstract

Author summary Stimulating a neural system and observing its effect through simultaneous observation offers the promise to better understand how neural systems perform computations, as well as for the treatment of neurological disorders. A powerful perspective for understanding a neural system’s behavior undergoing stimulation is to conceptualize them as dynamical systems, which considers the global effect that stimulation has on the brain, rather than only assessing what impact it has on the recorded signal from the brain. With this more comprehensive perspective comes a central challenge of determining what requirements need to be satisfied to harness neural observations and then stimulate to make one dynamical system function as another one entirely. This could lead to applications such as neural stimulators that make a diseased brain behave like its healthy counterpart, or to make a neural system previously capable of only hasty decision making to wait and accumulate more evidence for a more informed decision. In this work we explore the implications of this new perspective on neural stimulation and derive a simple prescription for using neural observations to inform stimulation protocol that makes one neural system behave like another one.
BibTeX
@article{hocker2019myopic,
  author = {David Hocker and Il Memming Park},
  title = {{Myopic control of neural dynamics}},
  journal = {PLOS Computational Biology},
  year = {2019},
  url = {https://openreview.net/forum?id=d5632nwNssM},
}

Loading related content...

Figure