Probabilistic neural network with homogeneity testing in recognition of discrete patterns set

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Traditional pattern recognition with limited samples faces challenges with standard algorithms like Probabilistic Neural Networks (PNN).
• PNNs, while effective, suffer from memory and speed limitations due to their memory-based approach.

Purpose of the Study:

• To develop an optimized pattern recognition method for small sample datasets.
• To address the limitations of existing PNN architectures in terms of speed and memory usage.
• To propose a novel PNN modification that leverages the discrete nature of patterns.

Main Methods:

• Mapping continuous feature vectors to a discrete range for statistical classification.
• Implementing a modified Probabilistic Neural Network (PNN) with homogeneity testing.
• Utilizing histograms of input and training samples for efficient data processing.
• Comparing performance against the original PNN using character n-grams for text authorship and facial recognition datasets.

Main Results:

• The modified PNN demonstrates improved classification accuracy by 1%-7% compared to the original PNN.
• The proposed method significantly enhances classification speed and reduces memory requirements.
• The modified PNN shows greater robustness to changes in the Gaussian kernel's smoothing parameter.

Conclusions:

• The novel PNN modification offers an optimal solution for pattern recognition tasks with small sample sizes.
• Exploiting the discrete nature of patterns and using histograms provides a more efficient and accurate approach.
• This method is highly effective for real-world applications like text authorship attribution and face recognition.