Exploring the capacity of neural networks to recognize objects from dolphin echoes across multiple orientations

Dolphins naturally recognize objects from multiple angles using echolocation. With training, humans can also learn to accurately classify objects based on their echoic features. In this study, we used neural networks to identify acoustic cues tat enable objects to be recognized from varying aspects. In Simulation 1, a self-organizing map was able to differentiate a subset of objects using only amplitude and frequency cues, but it classified some echoes from different objects as being from the same object. In Simulation 2, a multilayer perceptron was trained through error correction to identify objects based on echoes from a single aspect, and then tested on its ability to recognize those objects using echoes from different orientations. Overall, perceptrons performed similarly to trained undergraduates. Analysis of network connection weights revealed that both the amplitude and frequency of echoes, as well as the temporal dynamics of these features over the course of an echo train, enabled perceptrons to accurately identify objects when presented with novel orientations. These findings suggest that learning may strongly impact an organism's ability to echoically recognize an object from any viewpoint.