Ensemble probabilistic quantization encoding for information preservation of numerical variables in convolutional...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• One-hot encoding is standard for converting numeric to categorical data but loses quantitative information.
• This data loss negatively impacts the performance of Convolutional Neural Networks (CNNs).

Purpose of the Study:

• Introduce and evaluate ensemble probabilistic quantization encoding for improved CNN performance.
• Compare this new method against one-hot encoding, label smoothing, and mean square error.

Main Methods:

• Developed ensemble probabilistic quantization encoding, treating classes as collaborating quanta.
• Utilized cross-entropy loss for robustness and ensemble collaboration for enriched outcomes.
• Compared encoding techniques on the same dataset and CNN model structure.

Main Results:

• Quantitative information loss significantly degrades CNN functionality.
• Ensemble probabilistic quantization demonstrated less dependency on the number of classes.
• The new method maintained effectiveness even with fewer classes compared to alternatives.

Conclusions:

• Efficient transmission of quantitative information is vital for optimal CNN performance.
• Ensemble probabilistic quantization effectively conveys diverse quantitative information with fewer classes.
• This method outperforms one-hot encoding and label smoothing, particularly in low-class scenarios.