Title Page

Contents

Abstract 13

Ⅰ. Introduction 15

1. Necessity of research on sensory-neuromorphic systems 15

2. Necessity of research on real-time recognition 19

References 21

Ⅱ. Real-time data processing for integrating sensory-neuromorphic system 26

1. Introduction 26

2. Parts of sensory-neuromorphic systems 1: Sensory receptor 27

A. Device configuration and triboelectrification mechanism 27

B. Analysis of electrical response to finger tapping 31

C. Summary 34

3. Design procedure for tactile patterns 35

A. Data collection and visualization 35

B. Creating a Dataset for training and inference 39

C. Summary 42

Reference 43

Ⅲ. Implementation of sensory-neuromorphic system 46

1. Introduction 46

2. Tactile neuromorphic system 47

3. Parts of sensory-neuromorphic systems 2: Synaptic device 49

A. Device schematic and ferroelectric mechanism 49

B. Analysis of synaptic characteristics 55

C. Analysis of controllability for dynamic characteristics 59

D. Summary 64

4. Artificial neural networks (ANNs) with tactile-neuromorphic system 65

A. Single-layer perceptron (SLP)-based ANNs 65

B. Multi-layer perceptron (MLP)-based ANNs 69

C. Summary 74

Reference 75

Ⅳ. Conclusion 79

1. Dissertation summary 79

2. Future research direction 82

논문요약 84

Table 3.1. Extracted synaptic parameters (Gmax, Gmin, nonlinearity) with respect to various spike conditions (0.1V to 2V, 5 ms to 50 ms, and 1 Hz and... 62

Table 3.2. Comparison of our ferroelectric synaptic device and other devices with respect to dynamic range (on/off ratio), non-linearity... 73

Figure 1.1. Conventional von Neumann architecture and conceptual parallel computing based on neuromorphic devices. 16

Figure 1.2. Schematic illustration of flexible sensory-neuromorphic system with stretchable resistive sensor and flexible artificial synapse array. 18

Figure 1.3. Acoustic pattern recognition task for training and inference. 20

Figure 2.1. Sensory receptor based on Cu/PDMS/Cu triboelectric sensor. (a) Schematic illustration (left panel) and photograph (right panel) of the... 28

Figure 2.2. Two-capacitor structure of Cu-PDMS-Cu triboelectric sensor. 29

Figure 2.3. (a) Real-time output voltage signals generated by "press" (upper panel) and "release" (bottom panel) stimuli. (b) Schematic illustrations and... 32

Figure 2.4. Real-time output voltage signals with consecutive press/release for two time-intervals, (a) ~0.06 s and (b) ~1 s. 33

Figure 2.5. (a) Procedure employed in the tactile neuromorphic system consisting of a triboelectric sensor, a microcontroller unit, and a simulation... 37

Figure 2.6. Method for capturing external stimuli using triboelectric sensors and microcontroller units (MCU). 38

Figure 2.7. Process of generating a training/test dataset with a randomization method; (a) The original and randomized electrical signal and (b) Visualized... 40

Figure 2.8. Representative input dataset (20 × 20 pixels) for Morse-code alphabet pattern recognition. 41

Figure 3.1. Tactile neuromorphic system for processing external stimuli in real-time. (a) Schematic illustration of the tactile neuromorphic system. This... 48

Figure 3.2. Synaptic plasticity of MoS₂/P(VDF-TrFE) heterostructure-based synaptic device. (a) Schematic illustration of a three-terminal synaptic device... 50

Figure 3.3. (a) Cross-sectional transmission electron microscopy (TEM) images of MoS₂/P(VDF-TrFE) heterostructures and (b) electron energy loss... 51

Figure 3.4. Polarization switching analysis of P(VDF-TrFE) layer using piezoresponse force microscopy (PFM). (a) Schematic illustration of PFM... 53

Figure 3.5. (a) Postsynaptic current characteristics of the synaptic device in response to the excitatory (red) and inhibitory (blue) single pulse spikes (±1... 54

Figure 3.6. (a) Variations in conductance for four consecutive positive/negative voltage pulses (±1 V, 20 ms). (b) Conductance state... 56

Figure 3.7. (a) Representative pulse scheme and LTP/LTD characteristics under 128 successive positive/negative weight update pulses. (b) Operational... 58

Figure 3.8. (a) LTP/LTD characteristic curves according to pulse and (b) The extracted Gmax/Gmin and symmetricity with respect to various pulse spike...[이미지참조] 61

Figure 3.9. (a) Cycle-to-cycle variation in LTP/LTD characteristics for 15 cycles. (b)-(c) Extracted synaptic factors (NLP, NLD, symmetricity, and...[이미지참조] 63

Figure 3.10. Process flow of the training and recognition simulation task for the Morse code alphabet "A" to "Z". Recognition rate as a function of the... 66

Figure 3.11. Recognition rate as a function of the number of epochs for the single-layer perceptron (SLP)-based NN using Morse code alphabets. 68

Figure 3.12. Schematic of multi-layer perceptron neural network (400×100×10). 69

Figure 3.13. MLP simulation with 100 hidden layers for Morse code patterns. 70

Figure 3.14. Recognition rate as a function of the number of epochs for the multilayer perceptron based NN using MNIST handwritten digits. 71

Figure 3.15. MLP simulations for MNIST handwritten digit dataset in terms of (a) pulse frequency (2, 4, and 8 Hz), (b) amplitude (0.5, 1, and 2 V), and (c)... 72