Wing structure and neural encoding jointly determine sensing strategies in insect flight

Abstract

Animals rely on sensory feedback to generate accurate, reliable movements. In many flying insects, strain-sensitive neurons on the wings provide rapid feedback that is critical for stable flight control. While the impacts of wing structure on aerodynamic performance have been widely studied, the impacts of wing structure on sensing are largely unexplored. In this paper, we show how the structural properties of the wing and encoding by mechanosensory neurons interact to jointly determine optimal sensing strategies and performance. Specifically, we examine how neural sensors can be placed effectively on a flapping wing to detect body rotation about different axes, using a computational wing model with varying flexural stiffness. A small set of mechanosensors, conveying strain information at key locations with a single action potential per wingbeat, enable accurate detection of body rotation. Optimal sensor locations are concentrated at either the wing base or the wing tip, and they transition sharply as a function of both wing stiffness and neural threshold. Moreover, the sensing strategy and performance is robust to both external disturbances and sensor loss. Typically, only five sensors are needed to achieve near-peak accuracy, with a single sensor often providing accuracy well above chance. Our results show that small-amplitude, dynamic signals can be extracted efficiently with spatially and temporally sparse sensors in the context of flight. The demonstrated interaction of wing structure and neural encoding properties points to the importance of understanding each in the context of their joint evolution.

Nonuniform structural properties of wings confer sensing advantages

“…For results in the main text, the threshold β is held constant at 1 × 10 -4 . We also tested the effects of varying the threshold from 1 × 10 -10 to 1 because the neural threshold substantially impacted results in previous work [15]. In the present study, altering the neural threshold does not alter conclusions.…”

“…The local normal strain in the direction of the wing span is then encoded by a dense grid of neural-inspired sensors, from which an optimal subset will be chosen to assess sensing performance, as in previous work [15]. At each node across the 25 x 50 grid of elements (corresponding to every 1 mm), strain is converted to a series of temporally sparse all-or-none sensing events using a linear-nonlinear model, a common model of neural responses [26].…”

“…For straightforward comparison across conditions, we consistently use the top 10 sensors (i.e., sensors with the largest weights in s ) to assess classification accuracy. In previous work, we find that 10 sensors are typically enough to achieve near-peak accuracy without including a large number of extraneous sensors [15]. This is also a reasonable number of sensors for engineered systems to ensure reduced noise and latency along with higher classification accuracies [33].…”

“…For both of these cases, we determine the time to first spike within each wingbeat with 0.1 ms precision and use only this spike timing information to classify the data. For wingbeats where no spike is elicited, we designate the spike time as zero, as in previous work [15]. Data are standardized in this optimization step, but the original (non-standardized) data are used to evaluate accuracy.…”

“…Previous work on the sensory functions of wings has been limited by relying on highly idealized wing models [8, 14, 15]. These models are unable to capture the rich morphological characteristics of real insect wings or the diverse array of structural properties one might consider in designing an engineered system, such as complex shape and nonuniform material properties.…”