Simulation of brain learning process through a novel fuzzy hardware approach

The problem of understanding and simulating human learning and inferencing procedures is faced from a hardware point of view. A fuzzy modeling method called active learning (ALM) has been developed by the authors. ALM is an adaptive recursive algorithm which tries to express a multi-input multi-output system as a fuzzy combination of some single-input single-output systems. It uses a fuzzy curve fitting technique for behavior extraction or finding the input-output transformation of each of the single-input single-output systems. The mentioned curve fitting technique is called ink drop spread (IDS) which also serves as processing engine for the fuzzy rule of composition. In this paper first we try to develop a digital hardware for executing the IDS. Then, we show that the natural difference between the fuzzy calculating procedure of IDS and the precise nature of digital hardware appears as a serious problem for achieving very high processing speed comparable to that of the brain. As the next step in the hardware development procedure an analog merged digital hardware approach is introduced. It is shown that when preserving the necessary calculating resolution for ALM, the analog merged digital approach reduces the number of circuit elements and increases the processing speed by at least two orders of magnitude. Then it is discussed how a non-precise version of the technology for binary-multi valued-analog-merged operations called Neuron-MOSFET (neuMOS) may suit the non-precise calculating nature of IDS, which appears very promising for achieving the speed of brain learning process. Finally based on the advantages in circuit integration scale and hardware speed obtained through utilizing the binary-multi valued-analog-merged operations for the hardware implementation of IDS a primary concept for a new generation of artificial neural networks structure is developed.