Developing higher-order networks with empirically selected units.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Polynomial neural networks can suffer from computational complexity due to a high number of monomial terms.
• Developing efficient algorithms for constructing these networks is crucial for practical applications.

Purpose of the Study:

• Introduce a class of simple polynomial neural network classifiers named mask perceptrons.
• Outline algorithms for the practical development of mask perceptrons.
• Evaluate the performance of mask perceptrons on various classification tasks.

Main Methods:

• Developed algorithms involving input attribute ordering based on usefulness.
• Employed heuristic-driven generation and selection of hidden units (monomial terms).
• Tested mask perceptrons on benchmark datasets (mushroom classification, faulty LED-display) and non-standard domains (spoken digit recognition, article category determination).

Main Results:

• Mask perceptrons demonstrated competitive performance against other classifiers on multiple domains.
• Achieved 100% accurate simple logic production rules after training on a small subset (6-20%) of the database.
• Effectively managed the combinatorial explosion of higher-order monomial terms.

Conclusions:

• Mask perceptrons provide an efficient and effective approach to polynomial neural network classification.
• The conversion to logic production rules offers interpretable and highly accurate models.
• This method shows promise for real-world applications requiring efficient and accurate classification.