Recognizing successive dolphin echoes with an integrator gateway network

A novel network architecture was developed to classify multiple successive echoes from targets ensonified by a dolphin echolocating in a naturalistic environment. The inputs to the network were spectral vectors of the echo plus one unit representing the start of each scan. This network combined information from successive echoes from the same target and reset between scans of different targets. The network was trained on a small subset (4%) of the total set of available echoes (1,335). Depending on the measure used to assess it, the network correctly classified between 90% and 93% of all echo trains. In contrast a standard backpropagation network with the same number of units and variable connections performed with only about 63% accuracy in classifying echo trains. The integration model seems to provide a better account of the dolphin´s performance than a decision model that does not combine information from multiple echoes