Modern synergetic neural network for imbalanced small data classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep learning struggles with imbalanced small datasets due to overfitting.
• Recurrent neural networks offer robustness but Synergetic Neural Networks (SNNs) face association errors.
• Current SNN research often uses genetic algorithms, limiting parameter optimization.

Purpose of the Study:

• To introduce the modern synergetic neural network (MSNN) model.
• To address association errors and enhance SNN application capabilities.
• To improve classification performance on imbalanced datasets.

Main Methods:

• Correcting state initialization to resolve association errors.
• Utilizing error backpropagation and gradient bypass for attention parameter optimization.
• Enabling joint training with other network layers through self-learning attention parameters.

Main Results:

• MSNN liberates parameter optimization space by correcting state initialization.
• Self-learning attention parameters adapt to imbalanced sample sizes, boosting classification.
• MSNN achieved the best average rank across 75 UCI Machine Learning Datasets classification tasks.

Conclusions:

• MSNN effectively overcomes SNN limitations, particularly association errors.
• The model demonstrates superior adaptability and classification performance on imbalanced data.
• MSNN significantly outperforms existing neural and non-neural machine learning methods.