In this dissertation, we investigated techniques to develop high-performance and mass productive neuromorphic system using resistive switching random access memory (ReRAM) devices for application to memristors as artificial synapses. It is important to stably implement memristor crossbar array mimicking synapses of brain in neuromorphic systems. Our interest was to extend neuromorphic applications of memristors by adopting the-state-of-the-art deep neural network, novel synaptic plasticity, and advanced conduction modulation linearity with artificial synapse devices.
First, we introduced effect of the conduction modulation linearity (CML) on the convolutional neural network (CNN) of Pr0.7Ca0.3MnO3 (PCMO) memristor. An amorphous PCMO thin film was grown on a TiN/SiO₂/Si substrate at 300 °C under varying oxygen pressure to adjust the number of oxygen vacancies (OVs) in the PCMO film. PCMO films deposited at low oxygen pressure of 100 mTorr contain a large number of OVs and exhibit high CML, suggesting that CML can be obtained by controlling the oxygen pressure during the growth process. Moreover, we found that an improvement in classification accuracy in CNN simulations using PCMO memristors with high CML.
Second, we developed synaptic plasticity through the OV concentration for the improvement of learning accuracy in a Ta2O5 memristor. Ta2O5 memristor annealed under N₂ at 10 Torr has abundant OVs and enhance the CML. Essential biological synaptic properties are implemented in Ta2O5 memristor, which confirms that the Ta2O5 memristor based CNN exhibits a high classification accuracy of 93% due to high CML. Furthermore, this simple method used to optimize CML can also be applied to other ReRAM memristors whose switching behavior is determined by the growth of OV filaments.
Third, we introduced detection stress using brain imaging signals with memristor-based convolutional neural networks. The functional near-infrared spectroscopy (fNIRS) is widely employed by CNN to recognize a user's stress state. The fNIRS signal is supplied to the CNN model in the form of image. In order to classify stress states, we suggested a memristor-based CNN (M-CNN) that updates its weights according to the conductances of the memristors. We calculated the classification accuracies between the control and stress groups using the M-CNN, and then compared with the conventional CNN (C-CNN). To ensure a fair comparison, we compared their accuracies with the memristor-based Densenet (M-Densenet) and the conventional Densenet (C-Densenet). We analyzed the reproducibility of M-CNN by calculating the classification accuracy of 93.33% exceeded that of C-CNN (87.92%), and proving reliability from a confusion matrix. Moreover, M-Densenet (92.38%) has higher accuracy than C-Densenet (90.00%), but shows lower accuracy than M-CNN.
At last, we performed a theoretical study to estimate the yield of memristor crossbar-arrayed neural networks using classification accuracies affected by the conductance linearity of memristors. The neural networks are composed of one or two memristors-high linear memristors (HLMs), medium linear memristors (MLMs), or low linear memristors (LLMs)-depending on the conductance linearities measured from the fabricated PCMO memristors. We calculate the classification accuracy of images using three types of neural networks-convolutional neural network (CNN), fully connected neural network (FCN), and Squeezenet-with two randomly arranged memristors from 0 to 100%. Performance of CNN, FCN, Squeezenet in the same database are presented. We achieve reliable results by averaging the calculated accuracies by randomly arranging the cell positions 30 times between two memristors. The neural networks with 100% HLM, 100% MLM, or all combinations of HLM and MLM produce the accuracy above 80%. However, for combinations of MLM and LLM, ratios of 10%, 20%, and 90% for LLM can produce the accuracy above 80% for FCN, Squeezenet, and CNN, respectively. Thus, we suggest that LLMs should be contained less than 10% to achieve over 80% accuracy for manufacturing memristor-based crossbar-arrayed neural networks.
Consequently, in this dissertation, it can be concluded that the memristor implemented is potential artificial synapse which is exhibiting various synaptic plasticity for manufacturing high density array circuits and highly accurate neuromorphic computing applications.