Dolphin echolocation: identification of returning echoes using a counterpropagation network

Abstract

The authors report on the result of experiments on the recognition of targets by an echo-locating dolphin and by a counterpropagation neural network. The first experiment describes the success of a counterpropagation network with 20 input bands in classifying four different targets on the basis of the spectral distribution returned in the echo from the objects. Echoes for this experiment were collected in a quiet test pool using a simulated dolphin click as the source. These patterns were classified with 100% accuracy. These data compared well with those obtained from a real dolphin recognizing these same targets in a noisy natural environment (94.5% correct). The same network architecture was then used to classify echoes from three of these targets, collected while the dolphin echo-located in the noisy environment while performing the item recognition task. Under these conditions, the network was 96.7% correct. These results suggest that neural networks of various sorts may be promising computational devices for automated sonar target recognition and for the modeling of cognitive and perceptual processes in dolphins.<>