A spatio-temporal artificial neural network for object recognition using bioacoustic signals

This research designs and implements a biologically based spatio-temporal artificial neural network (ANN) that processes spatio-temporal bioacoustic signals to perform object recognition regardless of orientation or distance to target. The bioacoustic signals used in this research are identical to those used by the echolocating bat Myotis lucifugus, the little brown bat. These signals are chosen because echolocating bats provide an excellent example of a biological system that can perform target identification and ranging.; The ANN architecture consists of cascaded two dimensional modified sequential competitive avalanche field (MSCAF) layers. The output of the final MSCAF layer is input into an associative memory. The network is first tested on artificially generated spatio-temporal signals. The network is shown to be able to classify the binary spatio-temporal patterns correctly when corrupted by noise, occlusions or time shifting.; Next, the network receives information from a cortical response map, which models the neural activity in the auditory cortex of the Myotis (Palakal et al.). The cortical response map generates a spatio-temporal pattern (STP) based upon the echo signature of the target the network is attempting to identify. This STP is input to the ANN. By using the output of one MSCAF layer as the input to another MSCAF layer, the sampling rate is lowered with each layer. The output of the last MSCAF layer represents an entire spatio-temporal pattern and this output is input to an associative memory network. The associative memory network maps the output of the last MSACF layer to the object that it represents.