Comparison of sonar discrimination: dolphin and an artificial neural network

The capability of an echolocating dolphin to discriminate differences in the wall thickness of cylinders (3.81 cm o.d. and 12.7 cm length) was determined by Au and Pawloski [J. Comp. Physiol. A 170, 41-47 (1992)]. The dolphin was required to discriminate a standard target from comparison targets of differing wall thicknesses. Performance varied from 96% to 56% correct depending on the wall thickness of the comparison targets. The 75% correct threshold was determined to be wall thickness differences of -0.23 mm for comparison targets with thinner walls and +0.27 mm for comparison targets with thicker walls than the standard. The dolphin performance was unchanged in the presence of artificial broadband masking noise until the echo-energy-to-noise ratio fell below approximately 15 dB. A counterpropagation artificial neural network was used to examine broadband echo features from the same cylinders. Features of the echoes were determined by passing them through a filter bank of constant-Q filters. Echo features of the standard and each comparison target were analyzed in pairs by a neural network having two output nodes. Twenty echoes per target were used in the training set and 30 additional echoes per target were used in the test set. For the noise free condition, the network performed at a comparable level to the dolphin for Q values between 4 and 5. In the presence of noise, Q values between 7 and 8 were needed before the network could perform at a comparable level to the dolphin for echo-energy-to-noise ratios of 10 and 15 dB.(ABSTRACT TRUNCATED AT 250 WORDS)