Spike-based classification of tactile features with an insect-inspired antenna neuromechanical model

Description and abstract.
A hallmark of insect intelligence is in its efficiency and economy of behaviors in complex environments. For instance, many insects rely on touch information from antennae to guide rapid decision-making. To study how touch information is represented at the neural level for efficient feature classification, we developed a neuro-mechanical computational model of the cockroach antenna P. americana that simulates antenna mechanics and neural encoding of mechanosensory information. The mechanics of the antenna were modeled within a physics-based simulation environment (MuJoCo), which revealed spatiotemporal strain patterns (tactile tensors) arising from antennal deflections during contact with features. These tactile tensors were then used as input to a phenomenological model of campaniform and hair sensilla, where individual mechanosensory units encoded local mechanical inputs into spike rate responses. We validated the predicted firing activity against extracellular recordings from the antennal nerve and determined a strong correlation between simulated and experimental neural activity. To determine how spikes could be used for efficient feature classification, we trained spiking neural networks (SNNs) on the model-generated spike trains to classify tactile stimuli. The SNN classifiers achieved robust predictions, demonstrating the potential of sparse, spike-based codes for efficient tactile feature discrimination. Together, our results highlight how the coupling of mechanics and neural computation supports energy-efficient sensing in insects and provide a pathway for developing efficient tactile sensors.