Fast training algorithms for multilayer neural nets

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Back-propagation is a common algorithm for training neural networks, but can be slow.
• Specifying the number of hidden units in advance is often a limitation.

Purpose of the Study:

• To introduce a new algorithm that is faster than back-propagation.
• To present an algorithm that does not require pre-specification of hidden units.
• To compare its performance against other pattern recognition algorithms.

Main Methods:

• Development and implementation of a novel neural network training algorithm.
• Testing on artificial datasets (e.g., parity problem) and real-world speech recognition tasks.
• Comparison of training times and recognition accuracy with back-propagation and k-d tree-based algorithms.

Main Results:

• The new algorithm demonstrates faster training times compared to back-propagation.
• Achieves recognition accuracy comparable to or better than back-propagation.
• Performance is similar to the nearest-neighbor algorithm but with reduced storage requirements.

Conclusions:

• The developed algorithm provides an efficient alternative to back-propagation for neural network training.
• It offers competitive accuracy with advantages in speed and flexibility regarding hidden unit specification.
• The algorithm shows promise for various pattern recognition applications, including speech recognition.