Adaptive Response Function Neurons

Abstract

Biological neurons that show a locally tuned response to input may arise from the network topology of interneurons in the system. By considering such a subnetwork, a learning algorithm is developed for the online learning of the centre, width and shape of locally tuned response functions. The response function for each input is trained independently, resulting in a very good fit for the presented data. Two example networks utilising these neurons were considered. The first was a completely supervised network while the second utilised a Kohonen-like training scheme for the hidden layer. The adaptive response function neurons (ARFNs) were able to achieve excellent class separation while maintaining good generalisation with relatively few neurons.