Plausible Neural Network – A novel and powerful intelligent system for high dimensional data analysis and future analog computing machine

Abstract

There have been many important new developments of science in the last few decades; among them are fractals/chaos, neural networks, and fuzzy systems. These developments are mostly due to the invention of the digital computer. Although we can now explore what is not possible to compute with pen and paper, it also exposes new kinds of problems that still cannot be handled with digital computers. In this seminar we will introduce another paradigm of computation - parallel analog computation. We will discuss why we need analog computation; what is the self-organization principle of a learning/adaptive system; how the information flow of parallel computation can be controlled; and how the physics of computation relates to logic and inference. A decade ago, I have found a way to unify probability theory and fuzzy set theories into a single framework of statistical inference. Recently I have combined neural networks and fuzzy systems, which has shed light onto the mystery of how the brain's higher order cognition network is organized and how logic is computed. Now the brain's higher order cognition can be linked to lower levels of the perception system, and therefore be understood. Using my theory, I have developed a neural network model called Plausible Neural Network (PNN), which can demonstrate the propriety of this approach. A Plausible Neural Network performs associative memory, supervised, semi-supervised, and unsupervised learning, function/relation approximation, and belief judgement in a single network architecture. This network architecture can easily be implemented by analog VLSI circuit design and pave the way for future quantum computation design. The power of PNN stems from its use of belief judgment and fuzzy logic. This provides a new breakthrough, which I believe will revolutionize statistics and data analysis. Traditional statistics is limited by its inflexibility, because it lacks fuzzy concept in dealing with missing and incomplete dada, and it has difficulty combining quantitative and qualitative data. PNN has the ability to incorporate all kind of the data and create high dimensional patterns, where variation can be understood. I have tested many missing and censored data with PNN and found that it can be computed as easily as ordinary data, because fuzzy set coding is so flexible. Humans can observe all kinds of data without any problem, (most images are fuzzy and incomplete data sets). Using PNN to analyze data we can do even better than human judgment, because it mimics the way human incorporate data, but at a much faster rate.