Fuzzy c-means-based architecture reduction of a probabilistic neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Probabilistic Neural Networks (PNNs) exhibit efficiency limitations due to their pattern layer design, which activates all input records.
• This complexity poses challenges for PNNs in big data classification tasks.
• Existing PNN architectures can be computationally intensive for large datasets.

Purpose of the Study:

• To propose a novel algorithm for reducing the structural complexity of PNNs.
• To enhance the efficiency of PNNs, particularly for large-scale classification problems.
• To validate the effectiveness of the proposed reduction algorithm on repository datasets.

Main Methods:

• A new algorithm for PNN structure reduction is presented, utilizing fuzzy c-means clustering.
• Pattern neurons are selected based on cluster centroids derived from fuzzy c-means.
• Input vectors with the highest membership coefficients are chosen to activate pattern neurons.

Main Results:

• The proposed algorithm was applied to repository datasets for classification tasks.
• PNNs were trained using conjugate gradients, reinforcement learning, and the plugin method.
• 10-fold cross-validation demonstrated comparable or improved performance of the reduced PNNs.

Conclusions:

• The developed algorithm effectively reduces the structure of PNNs.
• The reduced PNN architecture maintains classification accuracy while improving efficiency.
• The findings confirm the validity and practical applicability of the introduced algorithm for big data classification.