近年來類神經網路的快速發展為人類帶來十分便利的生活，第一代及第二代神經網路在各種不同的領域都大放異彩，例如影像辨識、語音辨識等等。其中第二代神經網路:卷積神經網路(CNN)更是在影像辨識方面取得非常好的成績，但其代價是大量的能量消耗及冗長的運算時間，對於運算核心的要求也相當高，相較於這種高成本的運算方式，第三代神經網路(SNN)因為其更加仿生的特性，能夠以相當低的功耗及運算成本達成訓練神經網路的效果。而傳統的馮紐曼結構式運算將記憶體和運算單元分開，無法應付在訓練神經網路時所需要的大量資料傳輸，作為馮紐曼結構的改進方法，記憶體內運算被提出，他將記憶體和運算單元結合在同一個結構當中，大幅度地改進了運算所需的時間及功耗，很適合運用在類神經網路上。
在人腦當中，訊息的傳遞是透過神經元(neurons)及連結神經元的突觸(synapses)所達成，在研究仿生之神經網路時，許多不同的神經元模型被提出，其中最為仿生的是Hodgkin-Huxley模型，但因其複雜度及運算成本太高而不常被使用，另一種模型也是 最為廣泛使用的是Leaky Integrate-and-Fire (LIF)模型，相較於Hodgkin-Huxley模型它的複雜度不高且也可以有效達到仿生的效果。
運用不同的新型元件，例如鐵電材料電晶體、磁阻式記憶體等等，可以模擬LIF的特性並建構出神經元模型，其中鐵電電晶體作為新型材料的電晶體，具有磁滯特性且基於氧化鉿的鐵電電晶體具有很好的微縮性和CMOS兼容性，因此本篇論文是以運用在第三代神經網路的鐵電電晶體神經元模型之建構及模擬為主題，並用TCAD和PyTorch做為模擬工具呈現最後的模擬和訓練結果。

In recent years, the rapid development of neural networks has brought a convenience life to human beings. The second generation of neural network, convolutional neural networks (CNN), has achieved great results in image recognition but at the cost of large energy consumption and long computation time. Also, the requirements for the computational core to process the CNN are very high. Comparing to the second generation of neural networks, the third generation of neural networks (SNN) is more bionic in nature and can achieve similar effects of training neural networks with lower power and computational cost.
The traditional Von Neumann architecture, which seperates memory units and computational units, can not cope with the large amount of data transmitting required for training neural networks. As an improvement of the Von Neumann architecture, in-memory computing was proposed, which combines memory and computational units in the same structure, dramatically improves the time and power consumption required for computing and training a neural network.　
In the human brain, the transmission of information is accomplished through neurons and synapses that connect neurons. In the studies of bionic neural networks, ,any different neuron models have been proposed, among which the most bionic one is the Hodgkin-Huxley model, but it is not often used because of its high complexity and computational cost. Another model and also the most widely used model is the LIF neuron model, which is less complex than Hodgkin-Huxley model and can effectively achieve the bionic effect.
Using different novel devices, such as ferroelectric transistors, magnetic random access memory, etc., we can simulate the characteristics of LIF and build a model of it. Among the emerging devices, ferroelectric transistors with hysteresis properties and are based on hafnium oxide have good microscopicity and CMOS compatibility. Therefore, this thesis focuses on the construction and simulation of FeFET neuron models for third generation of neural networks, and uses TCAD and Pytorch as simulation tools to present the final simulation and training results.

[1] Hermann Eul, "ICs for mobile multimedia communications," 2006 IEEE International Solid State Circuits Conference - Digest of Technical Papers, San Francisco, CA, USA, 2006, pp. 21-39.
[2] Gong, N., et al., Signal and noise extraction from analog memory elements for neuromorphic computing. Nature Communications, 2018. 9(1): p. 2102.
[3] Siou-Sian Lin, An Ultra-low-power Fast Transient Low Dropout Regulator for Portable Applications, National Cheng Kung University, 2021
[4] CIBA Collection of Medical Illustrations, Volume 1
[5] Neuron. (2022, July 11). In Wikipedia, the free encyclopedia. Retrieved July17, 2022, from https://en.wikipedia.org/wiki/Neuron
[6] Artifiial neural netwrok. (2022, July 11). In Wikipedia, the free encyclopedia. Retrieved July17, 2022, fromhttps://en.wikipedia.org/wiki/ Artifiial_neural_netwrok
[7] Oliveira, Rodrigo M. S. de et al. A System Based on Artificial Neural Networks for Automatic Classification of Hydro-generator Stator Windings Partial Discharges. Journal of Microwaves, Optoelectronics and Electromagnetic Applications [online]. 2017, v. 16, n. 03 [Accessed 6 September 2022] , pp. 628-645. Available from: <https://doi.org/10.1590/2179-10742017v16i3854>. ISSN 2179-1074. https://doi.org/10.1590/2179-10742017v16i3854.
[8] K. Naik. “Activation Functions in Deep Learning” krichnaik.in https://krishnaik.in/2022/02/14/activation-function (accessed Jul. 30, 2022)
[9] S. Ronaghan. “Deep Learning: Common Architectures” srnghn.medium.com, https://srnghn.medium.com/deep-learning-common-architectures-6071d47cb383(accessed Jul. 30, 2022)
[10] J.C. Martinez. “Introduction to Convolutional Neural Networks CNNs” aigents.co
https://aigents.co/data-science-blog/publication/introduction-to-convolutional-neural-networks-cnns (accessed Jul. 30, 2022)
[11] C. Thomas. “An introduction to Convolutional Neural Networks”, May 2019 towardsdatascience.com https://towardsdatascience.com/an-introduction-to-convolutional-neural-networks-eb0b60b58fd7 (accessed Jul. 30, 2022)
[12] R. Parmar. “Demystifying Convolutional Neural Networks”, Oct 2018, towardsdatascience.com https://towardsdatascience.com/demystifying-convolutional-neural-networks-384785791596 (accessed Jul. 31, 2022)
[13] P. Hoang. “What are some of the limitations or drawbacks of Convolutional Neural Networks?” Quora.com https://www.quora.com/What-are-some-of-the-limitations-or-drawbacks-of-Convolutional-Neural-Networks (accessed Aug. 3, 2022).
[14] F. Ponulak. and A. Kasinski. Introduction to spiking neural networks : Information processing, learning and applications, Acta Neurobiol Exp 2011,71:409-433
[15] Lee, C., Sarwar, S., Panda, P., Srinivasan, G. and Roy, K., 2022. Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures. Advanced Materials ,12 January 2018,https://doi.org/10.1002/adma.201705587
[16] Zaehle, T., Rach, S. and Herrmann, C., 2010. Transcranial Alternating Current Stimulation Enhances Individual Alpha Activity in Human EEG. PLoS ONE, 5(11), p.e13766.
[17] Xie, Y., Kim, W., Kim, Y., Kim, S., Gonsalves, J., BrightSky, M., Lam, C., Zhu, Y. and Cha, J., 2018. Self‐Healing of a Confined Phase Change Memory Device with a Metallic Surfactant Layer. Advanced Materials, 30(9), p.1705587.
[18] Tuma, T., Pantazi, A., Le Gallo, M., Sebastian, A. and Eleftheriou, E., 2016. Stochastic phase-change neurons. Nature Nanotechnology, 11(8), pp.693-699.
[19] Liang, F., Wang, I. and Hou, T., 2021. Progress and Benchmark of Spiking Neuron Devices and Circuits. Advanced Intelligent Systems, 3(8), p.2100007.
[20] M.-H. Wu, M.-C. Hong, C.-C. Chang, P. Sahu, J.-H. Wei, H.-Y. Lee,S.-S. Sheu, T.-H. Hou, in 2019 IEEE Symp. on VLSI Technology(VLSI-T), IEEE, Piscataway, NJ 2019, p. T34.
[21] Celler, G. and Cristoloveanu, S., 2003. Frontiers of silicon-on-insulator. Journal of Applied Physics, 93(9), pp.4955-4978.

[22] Silicon on insulator. (2022, July 11). In Wikipedia, the free encyclopedia. Retrieved July17, 2022, from https://en.wikipedia.org/wiki/Silicon_on_insulator
[23] Dutta, S., Kumar, V., Shukla, A., Mohapatra, N. and Ganguly, U., 2017. Leaky Integrate and Fire Neuron by Charge-Discharge Dynamics in Floating-Body MOSFET. Scientific Reports, 7(1).
[24] Alam, M., Si, M. and Ye, P., 2019. A critical review of recent progress on negative capacitance field-effect transistors. Applied Physics Letters, 114(9), p.090401.
[25] Lone, I., Aslam, J., Radwan, N., Bashal, A., Ajlouni, A. and Akhter, A., 2019. Multiferroic ABO3 Transition Metal Oxides: a Rare Interaction of Ferroelectricity and Magnetism. Nanoscale Research Letters, 14(1).
[26] Z. Wang et al., "Experimental Demonstration of Ferroelectric Spiking Neurons for Unsupervised Clustering," 2018 IEEE International Electron Devices Meeting (IEDM), 2018, pp. 13.3.1-13.3.4, doi: 10.1109/IEDM.2018.8614586.
[27] Yu, S & Chen, PY 2016, 'Emerging Memory Technologies: Recent Trends and Prospects' IEEE Solid-State Circuits Magazine, vol. 8, no. 2, pp. 43-56. https://doi.org/10.1109/MSSC.2016.2546199
[28] SophiaLi, A. (2015). Spin Transfer Torque-RAM Devices as a Future Non-volatile Memory Solution.
[29] Manuel Le Gallo and Abu Sebastian 2020 J. Phys. D: Appl. Phys. 53 213002, An overview of phase-change memory device physics.
[30] "脉冲神经网络2: SNN的仿真介绍_IC_DESIGNER_XZK的博客-CSDN博客", Blog.csdn.net,2022.[Online].Available:https://blog.csdn.net/weixin_41701476/article/details/96155108. [Accessed: 01- Aug- 2022].
[31] I-Ting Huang , A Simulation Framework for STDP-Oriented Unsupervised Learning on Resistive Random Access Memory based Spiking Neural Network, National Cheng Kung University, 2021
[32] W. Gerstner, "2.2 Hodgkin-Huxley Model | Neuronal Dynamics online book", Neuronaldynamics.epfl.ch, 2022. [Online]. Available: https://neuronaldynamics.epfl.ch/online/Ch2.S2.html. [Accessed: 01- Aug- 2022].
[33] Biological neuron model. (2022, July 11). In Wikipedia, the free encyclopedia. Retrieved July17, 2022, from https://en.wikipedia.org/wiki/Biological_neuron_model
[34] Iakymchuk, T., Rosado-Muñoz, A., Guerrero-Martínez, J.F. et al. Simplified spiking neural network architecture and STDP learning algorithm applied to image classification. J Image Video Proc. 2015, 4 (2015). https://doi.org/10.1186/s13640-015-0059-4
[35] M. Madadi Asl, A. Valizadeh and P. Tass, "Propagation delays determine neuronal activity and synaptic connectivity patterns emerging in plastic neuronal networks", Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 28, no. 10, p. 106308, 2018. Available: 10.1063/1.5037309.
[36] Ana Filipa Sousa Silva, "Decoding gait phases from neural activity in rat", 2017
[37] "Google's Tensor Processing Unit explained: this is what the future of computing looks like". TechRadar. Retrieved 2017-01-19.
[38] Y. LeCun. “Introduction to Gradient Descent and Backpropagation Algorithm” https://atcold.github.io/pytorch-Deep-Learning/en/week02/02-1/ (accessed Sep. 2, 2022).
[39] Thorpe, Simon & Gautrais, Jacques. (1998). Rank Order Coding. Computational Neuroscience: Trends in Research. 113-118. 10.1007/978-1-4615-4831-7_19.