Neuro-Inspired System With Crossbar Array Of Amorphous Metal-Oxide-Semiconductor Thin-Film Devices As Self-Plastic Synapse Units Letter Recognition Of Five Alphabets

Artificial intelligences are essential concept and indispensable in future smart societies, while neural networks are typical representative schemes that imitate human brains and mimic biological functions. However, the conventional neural networks are composed of lengthy software that is executed by high-spec computing hardware, the computer size is enormous, and power dissipation is huge. On the other hand, neuro-inspired systems are practical solutions consisting only of customized hardware, and the hardware size and power dissipation can be saved. Therefore, we have been studying neuro-inspired systems with amorphous metal-oxide-semiconductor (AOS) thin- film devices as synapse units and suggesting revised Hebbian learning that is automatically and locally conducted without additional control circuits. Here, the conductance degradation can be employed as self-plastic weight of synapse units. As a result, it is promising that the neuro-inspired systems become three-dimensional integrated systems, the hardware size can be very compact, the power dissipation can be very low, and all functions of biological brains are obtained. In this study, we have been developing neuro-inspired systems with crossbar array of AOS thin-film devices as self-plastic synapse units. First, the crossbar array is produced, and it is discovered that the electric current continuously decreases along the application time. Next, the neuro-inspired system is really constructed by a field-programmable-gate-array LSI and crossbar array, and it is validated that a function of letter recognition is acquired after learning operation. In particular, we succeed in the letter recognition of five alphabets in this paper, whereas we succeeded in that of three alphabets in the previous paper, which is theoretically discussed, namely, the theoretical maximum performance seems to be achieved. Once the fundamental operations are validated, further progressed functions will be achieved by greatening the device and circuit scales.