Looking inside self-organizing map ensembles with resampling and negative correlation learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Ensemble learning methods improve model robustness and accuracy.
• Self-organizing maps (SOMs) are unsupervised neural networks effective for dimensionality reduction and visualization.
• Negative Correlation Learning (NCL) is a theoretical framework for training ensemble models by penalizing correlated errors.

Purpose of the Study:

• To investigate the efficacy of training self-organizing map (SOM) ensembles using negative correlation learning (NCL).
• To explore the relationship between diversity and accuracy within SOM ensembles.
• To determine if non-independent learning benefits SOM ensemble training.

Main Methods:

• Training ensembles of self-organizing maps (SOMs) with a small number of neurons.
• Applying negative correlation learning (NCL) to penalize correlated errors among individual SOMs.
• Analyzing the diversity (explicit/implicit, inter/intra) within SOMs and its correlation with prediction performance.

Main Results:

• SOMs with few neurons are suitable as weak ensemble components.
• The proposed NCL approach yields efficiently trained SOM ensembles that outperform reference learners.
• Quantified diversity measures show high correlation with prediction performance.
• Insights into the interplay between diversity and sublocal accuracy within SOMs were gained.

Conclusions:

• Negative correlation learning (NCL) effectively enhances self-organizing map (SOM) ensemble performance.
• Diversity, arising from various factors, is crucial for successful SOM ensemble learning.
• The transparency of SOMs allows for detailed analysis of ensemble dynamics and prediction accuracy.