Title Page
ABSTRACT
Contents
Chapter 1. Introduction 14
Chapter 2. Effect of oxygen vacancy on the conduction modulation linearity and improved classification accuracy in a Pr0.7Ca0.3MnO3 memristor[이미지참조] 17
2.1. Introduction 17
2.2. Materials and methods 20
2.3. Results and discussion 22
Chapter 3. Synaptic plasticity through the control of oxygen vacancy concentration to improve classification accuracy of Ta2O5 memristor[이미지참조] 36
3.1. Introduction 36
3.2. Materials and methods 40
3.3. Results and discussion 42
Chapter 4. Detection stress state using brain imaging signals with memristor-based convolutional neural networks 59
4.1. Introduction 59
4.2. Materials and methods 62
4.3. Results and discussion 70
Chapter 5. Estimation of crossbar array yield according to image classification accuracy by memristor-based neural networks 78
5.1. Introduction 78
5.2. Materials and methods 81
5.3. Results and discussion 84
Chapter 6. Conclusions 91
References 93
Fig. 2.1. (a) XRD patterns of the PCMO thin films deposited on the TiN–Si substrate at 300℃ by varying the OP. SEM images of the (b) cross-section and (c)... 24
Fig. 2.2. Variation of conductance with the applications of 100 P-spikes and 100 D-spikes to the PCMO memristor grown under various Ops: (a-1) 300 mTorr, (b-... 27
Fig. 2.3. (a) Schematic diagram of the CNN structure and (b) calculated recognition accuracy of the MNIST patterns as a function of the number of... 29
Fig. 2.4. Schematic diagrams showing (a) the amount of OVs and (b) growth of the filament in the PCMO film grown at 300 mTorr; (c) the amount of OVs and (d)... 31
Fig. 2.5. Variation in the synaptic weight as a function of (a) the retention time after the application of a different number of P-spikes, and (b) the number of... 35
Fig. 3.1. SEM images of (a) the cross section and (b) surface of Ta2O5 film; the inset of Fig. 1(b) shows the AFM image of this Ta2O5 film. (c) I–V curves obtained at...[이미지참조] 44
Fig. 3.2. (a) Voltage signal used to measure the transmission properties of Ta2O5 memristor. Variation of the conductance with the supply of 100 P spikes...[이미지참조] 47
Fig. 3.3. Variation of current as a function of P spike number measured with various magnitudes of P spikes; (ai) - 0.6V, (aii) - 0.7V, and (aiii) - 0.8V for...[이미지참조] 51
Fig. 3.4. Shape, magnitude, and period of (a) the input voltage (P-spike) and (b) the reading voltage that were used to measure the SPT-to-LTP transition curves.... 55
Fig. 3.5. (a) Pre- and post-spikes and (b) net-spike at △t=0.04 ms. Variations of △w as a function △t with the supply of net-spikes to the Ta2O5 memristors heated... 58
Fig. 4.1. Schematic of obtaining brain imaging signals from △HbO and ∆HbR measured by fNIRS. (a) Location of 15 channels at the prefrontal cortex. (b)... 64
Fig. 4.2. Memristor as artificial synapse imitating a biological synapse in the human brain system: (a) biological synapse between a pre-synaptic neuron and a... 67
Fig. 4.3. Bar graphs of the averaged classification accuracy between C-CNN (the conventional CNN) and M-CNN (memristor-based CNN). 72
Fig. 4.4. Confusion matrix proving the reliability of the classification accuracy results between the control group and the stress group. (a) example table... 75
Fig. 4.5. Bar graphs of the averaged classification accuracy to test the reproducibility in various dataset types using (a) CNN and (b) Densenet. In (a),... 77
Fig. 5.1. Synaptic characteristics of HLM, MLM, and LLM measured from fabricated PCMO memristors: (a) I-V curves and (b) normalized conductance... 86
Fig. 5.2. Average classification accuracies of CNN, Squeezenet and FCN classifiers composed of three memristors: HLM, MLM, and LLM. Average... 88
Fig. 5.3. Average classification accuracies and their SDs for 30 repeated calculations with randomly rearranging memristor positions in the crossbar... 90