The paper presents a new idea of software architecture inspired by the processing mechanism of the human brain. Stimulated by the working of the human brain, we propose an adaptive neural network software architecture that integrates the principles of neuroevolution for adjusting activation functions and an adaptive mechanism for selecting varying numbers of hidden layers to dynamically adjust the structure, functions, and parameters of the neural network. Utilizing genetic algorithms like crossover and mutation strengthens the architecture to optimize its components to adapt in situations like varying data distribution and learning objectives. We conducted an initial experiment on two benchmark image datasets (MNIST and CIFAR-10) and compared the performance for classification, clustering, and reinforcement learning tasks. We found that applying the proposed architecture with a neural network produces 51% better results. We also found that the results are comparable and better for clustering and reinforcement tasks on both datasets. The article concludes that the proposed architecture improves the performance of these machine-learning tasks over classical techniques and can offer a framework for developing robust and adaptable neural network systems.
Paper
Full text
Brain-Inspired Software Architecture: An Adaptive Neural Network Systems
Semantic Scholar · Computer Science · 2024
Abstract
The paper presents a new idea of software architecture inspired by the processing mechanism of the human brain. Stimulated by the working of the human brain, we propose an adaptive neural network software architecture that integrates the principles of neuroevolution for adjusting activation functions and an adaptive mechanism for selecting varying numbers of hidden layers to dynamically adjust the structure, functions, and parameters of the neural network. Utilizing genetic algorithms like crossover and mutation strengthens the architecture to optimize its components to adapt in situations like varying data distribution and learning objectives. We conducted an initial experiment on two benchmark image datasets (MNIST and CIFAR-10) and compared the performance for classification, clustering, and reinforcement learning tasks. We found that applying the proposed architecture with a neural network produces 51% better results. We also found that the results are comparable and better for clustering and reinforcement tasks on both datasets. The article concludes that the proposed architecture improves the performance of these machine-learning tasks over classical techniques and can offer a framework for developing robust and adaptable neural network systems.