A novel architecture design for multi-layer neural networks by using genetic algorithms: neural logic networks circuits

Created by W.Langdon from gp-bibliography.bib Revision:1.8344

@PhdThesis{Unal:thesis,

• author = "Hamit Taner Unal",
• title = "A novel architecture design for multi-layer neural networks by using genetic algorithms: neural logic networks circuits",
• school = "Selcuk Universitesi Fen Bilimleri Enstitusu",
• year = "2023",
• month = "9 " # mar,
• keywords = "genetic algorithms, genetic programming, ANN, Artificial Intelligence, Deep Learning, Machine Learning, Artificial Neural Networks, Optimization, Yapay Zeka, Derin Ogrenme, Makine Ogrenmesi, Yapay Sinir Aglari, Optimizasyon, Genetik Algoritma",
• URL = "https://hdl.handle.net/20.500.12395/50746",
• URL = "https://acikerisim.selcuk.edu.tr/bitstreams/395b70b2-a99f-454c-b960-61f130fc9c7c/download",
• size = "170 pages",
• abstract = "We introduce Neural Logic Circuits (NLC), a novel artificial intelligence paradigm inspired by all-or-none character of biological neurons and neuroplasticity of human brain. Developed as an alternative to multilayer artificial neural networks and deep learning models, the proposed paradigm is constructed by artificial neurons arranged in three-dimensional space and synaptic connections between those neurons. The artificial neurons are represented by logic Gates and Genetic Algorithms are used for the design and continuous optimization of neural connections in the network architecture. Unlike classical ANNs, our proposed structure has no layers and weights. All operations are performed in binary and there exists no complex operations such as activation functions or backpropagation. With the help of training data, the optimization algorithm builds the network from minimal scale and augments gradually. The proposed NLC paradigm is evaluated through a serious of experiments by using advanced accuracy metrics and the results are compared with modern and competitive machine learning algorithms. The experimental results proved the efficiency of our proposed paradigm by outperforming all other machine learning models on all test instances. Our NLC model achieved 23.83percent better performance on F1-Score and 11.47percent better performance on Balanced Accuracy compared to average of best results by other evaluated models. This is a clear indicator of our models superior ability to classify imbalanced datasets. The resulting NLC networks turned out to be very sparse structures, indicating strong plausibility of biological facts. The sparsity of those networks also proved that NLC can be used as a competitive feature selector. The transparent structure of the circuits obtained with NLC highlights the parameters that affect the result and makes important contributions to explainable artificial intelligence.",
• notes = "In turkish",

}

Genetic Programming entries for Hamit Taner Unal

Citations