No abstract

No abstract

Artificial intelligences are indispensable in future societies, and neural networks are representative that mimic biological brains. However, the conventional ones are complicated software on high-spec hardware, the machine size is bulky, and power consumption is huge. Neuromorphic systems are practical solutions composed solely of optimized hardware. Therefore, we are investigating neuromorphic systems with amorphous met-al-oxide-semiconductor thin-film devices as synapse elements and proposing modified Hebbian learning done locally without extra control circuits. As a result, the conductance deterioration can be utilized as synaptic plasticity. It is expected that the neuromorphic systems are three-dimensional-integrated systems, the size can be compact, power can be low, and all functions of biological brains are realized. In this study, we have developed neuromorphic systems with crosspoint-type amorphous Ga-Sn-O thin-film devices as self-plastic synapse elements, and fundamental operations are confirmed. First, crosspoint-type devices are fabricated, and it is found that the electric current gradually decreases along the bias time. Next, a neuromorphic system is actually implemented using a field-programmable-gate-array chip and crosspoint-type devices, and it is confirmed that a function of letter recognition is obtained after learning process. Once the fundamental operations are confirmed, more advanced functions will be obtained by increasing the device and circuit scales.

Artificial intelligences are promising as key technologies in future societies, but the conventional ones are complicated software executed on high-specked hardware. Therefore, we are investigating "brain-type integrated system", namely, neural network built only by hardware, which can be compact, low power, robust, and integrated on everything. In this study, we have succeeded in hardware simplification and are trying to utilize oxide semiconductors for the neuromorphic application because they can be fabricated using low cost fabrication such as sputtering and printing. We made a sandwiched structure of Ti/a-IGZO/Ti and found that the conductance decreases as the time goes by, which is available to modified Hebbian rule, a learning rule we proposed for our network. We modeled the conductance degradation of the a-IGZO thin film, developed a simulator, and obtained the learning results. We would like to develop brain-type integrated system with three-dimensional structure utilizing the oxide semiconductors in future.

No abstract