Online semi-supervised growing neural gas.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional semi-supervised classification methods, including those based on Growing Neural Gas (GNG), often necessitate explicit storage of training data.
• Labeling in existing GNG-based approaches is typically performed post-training, requiring data to be held in memory.

Purpose of the Study:

• To introduce a novel online semi-supervised classification approach, Online Semi-Supervised Growing Neural Gas (OSSGNG), that processes labeled and unlabeled data uniformly and in an online fashion.
• To develop a method that avoids the explicit storage of training examples by utilizing online labeling and prediction functions.

Main Methods:

• Developed OSSGNG, an extension of Growing Neural Gas (GNG), that integrates online labeling strategies.
• Employed online labeling and prediction functions to process both labeled and unlabeled data concurrently without explicit data storage.
• Evaluated the performance of OSSGNG against existing semi-supervised GNG extensions and other state-of-the-art methods on benchmark datasets.

Main Results:

• On-the-fly labeling strategies were found to not significantly degrade the performance of GNG-based classifiers.
• OSSGNG achieves performance comparable to previous semi-supervised GNG extensions that utilize offline labeling.
• OSSGNG demonstrates competitive performance against other leading semi-supervised learning approaches on standard datasets.

Conclusions:

• OSSGNG offers an effective and efficient online semi-supervised classification method, eliminating the need for explicit training data storage.
• The proposed online labeling approach maintains classification accuracy while improving computational efficiency and memory usage.
• OSSGNG represents a significant advancement in online semi-supervised learning, offering a viable alternative to traditional offline methods.