Towards Understanding Flight Characteristics by Reconstructing Motor Program Muscular Spikes

Abstract

Bio-inspired engineering has seen increased interest in recent years. Specifically, neural motor control research has attempted to reconnect the neural and motor programs. However, muscle spikes have complex, nonlinear, and non-unique conditional dependencies which make modeling muscle spikes extremely difficult. Further, biological variance in recordings adds difficulty to this problem.This work models the relationship between muscle spikes and flight characteristics for Manduca sexta using probabilistic modeling techniques. Leveraging advancements in neurobiology, this work examines loss functions and sampling techniques specific to spike trains. A Conditional Variational Autoencoder (CVAE) trained with these techniques is able to model this relationship. Using forces and torques as the underpinnings for flight characteristics, this model maintains intra-species variation while maintaining accuracy on spike timing. After verifying the model results, the latent space is examined to draw novel insights into the flight mechanisms of M. sexta.