Techniques for biased data distributions and variable classification with neural networks applied to otoneurological...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neuroscience

Background:

• Neural networks are powerful for classification but understanding their decisions, especially misclassifications, remains challenging.
• Investigating the relationship between input variables and output classes is crucial for scientific discovery and model improvement.
• Existing methods often struggle with biased output class distributions and quantifying variable influence.

Purpose of the Study:

• To develop novel techniques for analyzing neural network classification efficacy and interpretability.
• To address challenges posed by biased output class distributions in machine learning models.
• To quantify the information content of input variables and their contribution to classification outcomes.

Main Methods:

• A novel network structure and learning strategy designed for imbalanced datasets.
• A method to measure the classification information embedded within individual variables and variable groups.
• Techniques to visualize and interpret the properties learned by a neural network based on its internal structure.

Main Results:

• Successfully disentangled reasons for misclassification in neural networks.
• Quantified the influence of input variables on classification outcomes.
• Demonstrated the utility of the developed techniques on otoneurological data related to vertiginous diseases.

Conclusions:

• The novel techniques provide deeper insights into neural network decision-making processes.
• These methods enhance the interpretability of machine learning models in complex scientific domains.
• The approach is particularly valuable for analyzing otoneurological data and understanding vertiginous diseases.