Related Experiment Videos
Optimization-based learning with bounded error for feedforward neural networks
A Alessandri1, M Sanguineti, M Maggiore
1Naval Autom. Inst., Nat. Res. Council of Italy, Genoa.
IEEE Transactions on Neural Networks
|February 5, 2008
Summary
A novel optimization algorithm efficiently trains feedforward neural networks using a sliding-window cost. This method proves effective for large datasets and outperforms traditional backpropagation and extended Kalman filter learning approaches.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Neuroscience
Background:
- Feedforward neural networks (FNNs) are fundamental in machine learning.
- Efficient training algorithms are crucial for handling large datasets.
- Existing methods like backpropagation and extended Kalman filter have limitations.
Purpose of the Study:
- To introduce a new optimization-based learning algorithm for FNNs.
- To address computational efficiency for large-scale data.
- To analyze the algorithm's convergence and robustness.
Main Methods:
- The proposed algorithm determines network weights by minimizing a sliding-window cost function.
- It is designed for batch learning scenarios.
- Convergence and robustness properties are mathematically analyzed.
Main Results:
- Simulation results demonstrate the algorithm's effectiveness.
- The algorithm shows advantages over backpropagation and extended Kalman filter.
- Efficient handling of large datasets is confirmed.
Conclusions:
- The presented optimization algorithm offers an effective and efficient approach for training FNNs.
- It provides a viable alternative to existing learning methods, especially for large datasets.
- The algorithm's robustness and convergence properties are well-established.
