Training reformulated radial basis function neural networks capable of identifying uncertainty in data classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Radial Basis Function Neural Networks (RBFNNs) are widely used for classification tasks.
• Identifying uncertainty in data classification is crucial for robust AI systems.
• Conventional feed-forward neural networks (FFNNs) and standard RBFNNs often struggle with uncertainty quantification.

Purpose of the Study:

• To introduce a novel learning algorithm for training reformulated RBFNNs, specifically cosine RBFNNs.
• To enable these networks to effectively identify and quantify uncertainty in data classification.
• To enhance the performance of quantum neural networks (QNNs) and cosine RBFNNs through improved uncertainty identification.

Main Methods:

• Developed a learning algorithm that updates parameters in cosine RBFNNs to minimize class-conditional variances.
• Trained cosine RBFNNs and quantum neural networks (QNNs) using the proposed algorithm.
• Compared the performance of networks trained with the new algorithm against those trained with original algorithms and conventional FFNNs.

Main Results:

• The proposed learning algorithm successfully trained cosine RBFNNs and QNNs to identify classification uncertainty.
• Cosine RBFNNs trained with the new algorithm demonstrated superior uncertainty identification compared to those trained with the original algorithm.
• Networks trained with the proposed algorithm showed improved classification accuracy, particularly when ambiguous feature vectors were rejected.

Conclusions:

• The novel learning algorithm provides a method for enhancing RBFNNs and QNNs to identify data classification uncertainty.
• This capability is essential for improving the reliability and accuracy of AI classification systems.
• A simple strategy of rejecting ambiguous feature vectors can further boost the performance of these advanced neural networks.