Analysis and Design of Cultured Neuronal Networks for Adaptive and Reconfigurable Control

用于自适应和可重构控制的培养神经网络的分析和设计silvia Ferrari， Craig Henriquez和Antonius M.J. VanDongenDuke大学提出的活动的目标是开发一种训练培养神经网络的方法，以解决最优控制和预测中的挑战性问题。工程系统，如航空航天和机器人系统，已经从反馈控制器和估计器中受益，这些控制器和估计器是人为设计的，用于处理系统动力学已知的正常操作条件。然而，这些设计还不能处理不可预见的损坏和故障，涉及高度非线性和未建模的动力学，这是先验未知的。生物系统能够在各种约束条件下以最佳状态运行，并在新的和具有挑战性的情况出现时实时学习和适应。对于任何形式的非线性动力学和性能函数，神经动力学规划可以实时解决随机最优控制和估计问题。但是，目前人工神经网络和基于梯度的学习的形式与生物大脑的机制相去甚远。通过整合理论和实验，该项目将开发神经动力学规划算法，这些算法在生理上是合理的，并且可以在培养的光敏海马和皮层神经元上进行测试。在这个实验装置中，培养的神经元可以通过光模式进行训练，并且可以在网络中定义病变，评估记忆保留如何恢复，以及控制孤立神经元网络之间的连接。通过克服在体外或动物体内实验中限制这种系统级研究的复杂性，我们有望揭示神经元网络存储和检索感官信息的能力，并了解奖励途径对这一过程的影响。这个项目更广泛的影响是在芯片上逆向工程多巴胺和皮层神经元培养，并帮助揭示哺乳动物大脑中感觉运动学习的潜在机制

Analysis and Design of Cultured Neuronal Networks for Adaptive and Reconfigurable ControlSilvia Ferrari, Craig Henriquez, and Antonius M.J. VanDongenDuke UniversityThe goal of the proposed activity is to develop a methodology for training cultured neuronal networks to solve challenging problems in optimal control and prediction. Engineering systems, such as aerospace and robotic systems, already benefit from feedback controllers and estimators that are man-designed to handle normal operating conditions for which system dynamics are known a priori. However, these designs are not yet capable of handling unforeseen damages and failures involving highly nonlinear and unmodeled dynamics that are unknown a priori. Biological systems are capable of operating optimally, subject to a variety of constraints, and to learn and adapt in real time when new and challenging situations arise. Neurodynamic programming can solve stochastic optimal control and estimation problems in real time, for any form of nonlinear dynamics and performance functions. But, the current formalisms for artificial neural networks and gradient-based learning are far removed from the mechanisms found in biological brains. By integrating theory and experiments, this project will develop neurodynamic programming algorithms that are physiologically plausible and testable on light-sensitive hippocampal and cortical neurons growing in culture. In this experimental setup, the cultured neurons can be trained by light patterns, and it is possible to make defined lesions in the network and evaluate how memory retention is restored, as well as control the connectivity between isolated networks of neurons. By overcoming the complexities that are known to limit such system-level studies in in vitro or in vivo experiments on animals, it is expected that we will uncover the abilities of neuronal networks to store and retrieve sensory information, as well as understand the effect that reward pathways have on this process. The broader impact of this project is to reverse-engineer dopamine and cortical neuronal cultures on a chip, and to help uncover the mechanisms underlying sensorimotor learning in the mammalian brain