Maximum likelihood training of probabilistic neural networks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Pattern Recognition

Background:

• Probabilistic Neural Networks (PNNs) are effective for classification tasks.
• Existing training methods may lack efficiency or robustness for complex datasets.

Purpose of the Study:

• To introduce a novel maximum likelihood training method for PNNs.
• To enhance PNN performance in nonlinear discrimination tasks.

Main Methods:

• Developed a maximum likelihood training algorithm utilizing a Gaussian kernel (Parzen window).
• Generalized Fisher's linear discrimination method for nonlinear classification.
• Incorporated class pooling for improved generalization with small training sets.

Main Results:

• The method economizes the Parzen window estimator while maintaining feedforward neural network architecture.
• Achieved smooth, statistically robust, piece-wise flat discriminant boundaries.
• Demonstrated significantly faster computation compared to backpropagation.
• Exhibited superior numerical stability.

Conclusions:

• The proposed maximum likelihood training method provides an efficient and robust approach for PNNs.
• This technique generalizes nonlinear discrimination and improves classification performance, especially with limited data.