Spiking Neural Network Based on ReRAM for Character Recognition With Unsupervised Learning

Hasan, Rahmat Munir (2024) Spiking Neural Network Based on ReRAM for Character Recognition With Unsupervised Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 07111942000009-Undergraduate_Thesis.pdf] Text
07111942000009-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 April 2026.

Download (3MB) | Request a copy


The development of artificial intelligence is increasing along with the development of electronic devices. ReRAM is expected to be able to overcome the problem of the Von Nueman architecture being a link between the processor and RAM as well as the problem of the CPU having the weakness of a long learning process and the GPU consuming high power consumption. This study reviews the I-V characteristics of ReRAM and the implementation of spiking neural network (SNN) using ReRAM in CMOS technology. The SNN is designed for character recognition tasks with datasets using binary image data. Circuit simulation was performed using images of digits as input, crossbar stage, preprocessing stage, encoding stage, current amplifier, and time to first spike stage. The research results show that ReRAM meets I-V characteristics in the form of a hysteresis loop. Simulation results of the LIF neurons, ReRAM crossbar, current mirror stage, ISI encoding block shows correct operation. Circuit simulation shows 100% accuracy for recognizing numbers based on the first to spike results under a limited input pattern combination. High-level software simulation was performed to determine weights of the SNN using both supervised and unsupervised learning algorithms. For both learning algorithms, the resulting SNN shows an accuracy of 100%.
Perkembangan artificial intelegence semakin meningkat seiring dengan perkembangan perangkat elektronik. ReRAM diharapkan mampu mengatasi masalah arsitektur Von Nueman menjadi penghubung antara prosesor dan RAM serta masalah CPU memiliki kelemahan proses learning dengan waktu lama dan GPU menghabiskan konsumsi daya tinggi. Penelitian ini meninjau I-V karakteristik ReRAM dan implementasi spiking neural network menggunakan ReRAM di teknologi CMOS. SNN di desain untuk pekerjaan pengenalan karakter menggunakan dengan dataset data gambar biner. Simulasi rangkaian dilakukan menggunakan gambar digit sebagai input, crossbar stage, preprocessing stage, encoding stage, current amplifier, dan time to first spike. Hasil penelitian menunjukkan ReRAM memenuhi I-V karakteristik berupa hysteris loop. Hasil simulasi dari LIF neuron, ReRAM crossbar, current mirror stage, ISI encoding block menunjukkan operasi yang benar. Simulasi Rangkaian menunjukkan hasil akurasi 100% untuk mengenali angka berbasis first to spike dalam kombinasi pola masukan terbatas. High level Simulasi software dilakukan untuk menentukan bobot SNN menggunakan algoritma Supervised dan unsupervised learning. Untuk kedua aloritma pembelajaraan SNN yang dihasilkan menunjukkan akurasi sebesar 100%

Item Type: Thesis (Other)
Uncontrolled Keywords: ReRAM, Circuit, Simulation, Rangkaian, Simulasi, SNN
Subjects: T Technology > T Technology (General) > T57.62 Simulation
T Technology > T Technology (General) > T58.8 Productivity. Efficiency
T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User: Rahmat Munir Hasan
Date Deposited: 02 Feb 2024 08:28
Last Modified: 02 Feb 2024 08:28
URI: http://repository.its.ac.id/id/eprint/105987

Actions (login required)

View Item View Item