Artificial intelligences are almighty ? No, although some advantages are useful for various applications, there still remain many problems, such as, bulky size, high power consumption, less robustness, etc., because the current neural networks are long and complicated software executed on non-optimized Neumann-type hardware. Neuromorphic systems are biomimetic systems from hardware level and have the same advantages as living brains, especially, compact size, low power consumption, and operation robustness. On the other hand, thin-film semiconductor electronic devices can be fabricated on large areas, and three-dimensional layered structure can be acquired.
Neuromorphic systems using thin-film devices will be studied in this doctoral dissertation. First, we will investigate a neuromorphic system, where we will simplify a neuron element to three simple circuits and synapse element to one variable resistor or capacitor, and propose tug-of-war method and modified Hebbian learning, whose advantage is that the synaptic connection strength is automatically controlled using the local electrical conditions. Next, we will examine low-temperature poly-Si (LTPS) device, amorphous In-Ga-Zn-O (a-IGZO) device, and amorphous Ga-Sn-O (a-GTO) device, where, it will be confirmed that the electrical conductance gradually decreases when electric current flows, which is available as a synaptic connection strength. Finally, we will investigate Hopfield neural networks using crosspoint-type devices and cellular neural networks using separated architecture, surfaced architecture, layered architecture, and planar-type devices and confirm the correct operations of simple logic learning and letter reproduction. It is believed that these results will be theoretical bases to realize ultra-large scale integration for neuromorphic systems. Neuromorphic systems using thin-film devices have great potentials that the size can be compact, the power can be low, and the operation can be robust. Energy crisis can be avoided, and artificial intelligence on everything (AIoE) may be realized. Although we will not succeed in integration of an astronomical number of processing elements with three-dimensional layered structure, the research results will suggest that it is possible in the future.