Classification of Aspect-Dependent Targets by a Biomimetic Neural Network

摘要

A biomimetic neural network was used to model a bottlenose dolphin's ability torecognize aspect-dependent targets. Researchers used echo trains recorded during the dolphin trials to train an Integrator Gateway Network (IGN) to discriminate among the targets using echo spectra. The IGN classifies targets using an average-like sum of the spectra from successive echoes. However, combining echoes may reduce classification accuracy if the spectra vary from echo to echo. The dolphin and the IGN learned to recognize the geometric targets, even though orientation could vary. The process of recognition using cumulated echoes was robust for nonstationary raw input. The results support the notion that ensonified mines with complex shapes and echoes may be reliably classified using neural network architectures that are motivated through understanding of Marine Mammal System echolocation signals and performance.