Parameter Optimization and Learning in a Spiking Neural Network for UAV Obstacle Avoidance Targeting Neuromorphic...

Area of Science:

• Computational neuroscience
• Robotics
• Biomimetic systems

Background:

• The Lobula giant movement detector (LGMD) neuron in locusts is crucial for detecting looming objects and initiating escape behaviors.
• Understanding LGMD function can inform the development of efficient robotic obstacle avoidance systems.
• Neuromorphic engineering offers a pathway to implement biologically inspired neural networks on hardware.

Purpose of the Study:

• To develop and optimize a neuromorphic spiking neural network model of the locust LGMD.
• To investigate parameter optimization techniques for robust performance in noisy, variable hardware.
• To validate the optimized model using real-world data from a dynamic vision sensor.

Main Methods:

• Implementation of a spiking neural network model of the LGMD with adaptation and plasticity.
• Application of differential evolution (DE), Bayesian optimization (BO), and self-adaptive DE (SADE) for parameter tuning.
• Evaluation of optimizer performance through comprehensive comparisons.
• Testing the model with data from a dynamic vision sensor (DVS) on an unmanned aerial vehicle (UAV).

Main Results:

• Successful parameter optimization of the neuromorphic LGMD model using DE and BO techniques.
• SADE demonstrated effectiveness in overcoming challenges associated with DE parameter selection.
• The optimized model showed robust and reliable responses, validated by real-world UAV flight data.
• Demonstrated the feasibility of mapping the model to neuromorphic hardware.

Conclusions:

• Parameter optimization is essential for robust neuromorphic LGMD models in real-world applications.
• DE, BO, and SADE are effective methods for tuning complex spiking neural network parameters.
• The developed neuromorphic LGMD model shows promise for advanced robotic vision and navigation systems.