Joshua Bassey, Prairie View A&M University

Complex Valued Neural Networks with Multi-Valued Neurons

Abstract:

In a vast majority of deep learning applications today like computer vision, speech processing, wireless communications and signal processing, the need to handle data in form of complex numbers is a recurrent theme. Traditional methods either treats each part of the complex number independently or converts the complex number into its polar form and operate only on the phase. Both these processes have their disadvantages; the first approach ignores the relationship between the real and imaginary part of a complex number. The latter approach leaves out some information since it makes use of only the phase of the complex number. In this work, we present preliminary results on a fully complex-valued neural networks with multi-valued neurons, that is able to process complex number inputs. The network with multi-valued neurons also makes use of complex weights and do not need traditional backpropagation as traditional neural networks do. Other advantages to this method are also evaluated.

Presentation Author(s):
Joshua Bassey*

Judging Forms Official judges only