Improving supervised learning by adapting the problem to the learner.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Supervised learning algorithms face challenges in generalizing across all functions.
• Adapting a learning problem to suit a specific algorithm can enhance performance.
• Relabeling training data targets is a key strategy for problem adaptation.

Purpose of the Study:

• To introduce and validate problem adaptation methods for supervised learning.
• To demonstrate that adapting a function can make it easier for a learning algorithm to classify.
• To present Self-Oracle Learning with Confidence-based Target Relabeling (SOL-CTR) as a proof of concept.

Main Methods:

• Developed two variants of Self-Oracle Learning with Confidence-based Target Relabeling (SOL-CTR): SOL-CTR-E and SOL-CTR-P.
• These methods adapt problems by relabeling targets in training data.
• The adapted problems create 'easier' target functions for training artificial neural networks (ANNs).

Main Results:

• SOL-CTR methods consistently produced statistically significant improvements in accuracy (p < 0.05).
• Improvements were observed when compared to standard 0/1 targets.
• The positive impact was most notable on datasets with over 10,000 training examples across 41 datasets.

Conclusions:

• Problem adaptation, specifically through target relabeling, is a viable strategy to enhance supervised learning performance.
• SOL-CTR offers a practical approach to creating more learnable functions for ANNs.
• This method demonstrates significant accuracy gains in large-scale machine learning tasks.