Related Experiment Videos
A new constructive algorithm for architectural and functional adaptation of artificial neural networks
Md Monirul Islam1, Md Abdus Sattar, Md Faijul Amin
1Department of Computer Science and Engineering, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh. monirul@synapse.his.fukui-u.ac.jp
Summary
A new constructive algorithm (NCA) automatically designs artificial neural network (ANN) architectures. NCA enhances generalization by adapting both structure and function, outperforming existing methods in benchmark tests.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Neural Networks
Background:
- Artificial neural network (ANN) generalization heavily relies on architecture.
- Constructive algorithms offer automated ANN architecture determination.
- Existing methods have limitations in adapting ANN structure and function.
Purpose of the Study:
- Introduce a novel constructive algorithm (NCA) for automated ANN architecture design.
- Emphasize architectural and functional adaptation in the architecture determination process.
- Analyze the convergence and computational aspects of NCA.
Main Methods:
- NCA determines the number of hidden layers and neurons per layer.
- Functional adaptation is achieved by training hidden neurons on distinct data subsets, inspired by boosting.
- Analytical study of NCA's convergence and computational complexity.
Main Results:
- NCA's computational complexity is O(W xP(t) xtau), comparable to backpropagation.
- Evaluated on eight classification and two approximation problems.
- NCA produced ANNs with fewer hidden neurons and superior generalization.
Conclusions:
- NCA effectively determines ANN architectures with improved generalization ability.
- The algorithm balances architectural and functional adaptation for better learning.
- NCA presents a viable alternative to existing constructive and nonconstructive methods.
