Artificial neural networks imitating human brain has been greatly improved performance of artificial intelligence used for image-voice recognition, data analysis and other variety of fields. Recently, AlphaGo, based on deep learning algorithm created by Google DeepMind, defeated a human professional player in Go game. Despite this remarkable performance, there is a limit to the artificial intelligence that works in the conventional Von Neumann structure. For example, AlphaGo requires a high integration area of neurons and synapses and the operation power of about 1 MW, that is about four orders of magnitude higher than that of a human brain consuming about 20 W of power. In order to overcome these drawbacks, recent researches have been carried out to implement the functions of neurons and synapses, which are core components of neural networks, with electrical devices such as memristor, resistance random-access-memory (ReRAM), Mott transistor, partially-depleted silicon-on-insulator nMOSFET (PD SOI n-MOSFET), and spin-torque-transfer magnetic-random-access-memory (STT MRAM). In particular, ReRAM has intensively researched as a synapse because of a behavior of memristor, a simple device structure (i.e., top electrode/ReRAM material/bottom electrode), a fast write/erase speed of several tens of n sec., and a possibility of tera-bit integration. Many researches on ReRAM based synapses have remarkably studied to achieve a good linearity and symmetry of potentiation and depression. In this works, we proposed amorphous carbon oxide (α-COx) based synapse, showing synapse characteristic such as potentiation/depression as shown in Fig. 1. (a) and (b). For the more, it has been reported that a conventional neuron has been operated by a circuit using several C-MOSFETs, capacitor, and resistance, generating a high operation power and a heavy burden of neuron integration area. In our study, we proposed a novel neuron using a strained fully-depleted strained silicon-on-insulator (FD ε-SOI) n-MOSFET having leaky integration (LI) function as shown in Fig. 2. (a) and (b). Moreover, several neural networks have been researched; i.e., SNN, deep-neural-network (DNN), convolutional-neural-network (CNN), and Recurrent-neural-network (RNN). In particular, SNN has been intensively attracted since it demonstrated a lower power consumption and a possibility of real-time proceed event although its showed a lower accuracy of the pattern recognition. In our study, we tested the accuracy of the SNN pattern recognition for our proposed α-CO_x synapse and FD ε-SOI n-MOSFET neuron as shown in Fig. 3. (a) and (b).

* This research was supported by Nano·Material Technology Development Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Science, ICT and Future Planning(grant number. 2016910249)

* This material is based upon work supported by the Ministry of Trade, Industry & Energy(MOTIE, Korea) under Industrial Technology Innovation Program (1006855).

* The EDA tool was supported by the IC Design Education Center(IDEC), Korea.