The training of neural classifiers with condensed datasets

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multilayer perceptron neural networks require extensive training datasets.
• Training time can be a significant bottleneck in developing effective neural networks.
• Identifying and removing redundant or overlapping data is crucial for efficiency.

Purpose of the Study:

• To develop and evaluate a data condensing algorithm for accelerating neural network training.
• To assess the impact of data condensation on classification accuracy.
• To demonstrate the effectiveness of the proposed method on diverse datasets.

Main Methods:

• Application of a k-nearest-neighbor-based data condensing algorithm.
• Selective removal of overlapping training data points.
• Retention of training exemplars adjacent to the decision boundary.
• Comparison of network performance trained on condensed versus unedited datasets.

Main Results:

• Significant reduction in neural network training time.
• Achieved an order of magnitude speed-up in typical training.
• Maintained an undegraded misclassification rate compared to traditional training.
• Demonstrated consistent performance across synthetic and real-world datasets.

Conclusions:

• Data condensation using k-nearest-neighbor algorithms is an effective method for accelerating multilayer perceptron training.
• The proposed approach offers substantial computational savings without compromising predictive accuracy.
• This technique presents a valuable optimization for machine learning workflows.