A new neural computation scheme of unsupervised learning with applications to robot biped locomotion

A new neural computational scheme of unsupervised learning is proposed to construct a machine intelligence that is capable of overcoming unpredictable uncertainties and unknowns through proper interactions with environment. Our scheme consists of homogeneous neuron distributions which form layered clusters of computational circuit. Each neuron is very simple and of classical McCulloch-Pitts type equipped with Hebb-type plasticity for their interconnections. The novelty of our neuron lies in its ability to change its threshold according to its firing situation, which makes our scheme stable and configurable. Each cluster of neurons represents the numerical values by the number of firing neurons just like enumerations by fingers. This nonsymbolic nature of computations is shown to be very robust. It is shown that our configuration can act as a type of adaptive control which exhibits brain-like functions in its learning behaviors. Our scheme is shown to be successfully implemented to a biped robot that can walk under unstructured environment.