Learning procedure in a neural control model for the urinary bladder

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Systems Biology

Background:

• Urinary bladder dynamics are complex and challenging to model accurately.
• Controlling bladder function often requires sophisticated computational approaches.
• Existing control methods may rely on gradient-based learning, which is not always feasible.

Purpose of the Study:

• To develop a continuous neural network control system for a dynamical urinary bladder model.
• To train the neural network to accurately track prescribed volume fluctuations.
• To implement a learning procedure that avoids gradient descent due to unknown error gradients.

Main Methods:

• A continuous neural network was coupled with a dynamical urinary bladder model.
• The neural network was trained by adjusting weights and time constants.
• A novel learning procedure minimized the error functional without utilizing gradient descent.
• This approach addresses the challenge of unknown gradients in the neural network's output neurons.

Main Results:

• The neural network successfully controlled the bladder model to track target volume fluctuations.
• The learning procedure effectively minimized the error functional despite the absence of gradient information.
• The study demonstrates a viable alternative to gradient-based learning for complex dynamical systems.

Conclusions:

• A continuous neural network coupled to a dynamical bladder model can be trained effectively without gradient descent.
• This approach offers a promising method for controlling complex biological systems where gradients are not readily available.
• The findings have implications for developing advanced biofeedback and control systems for bladder management.