Self-supervised ARTMAP.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Traditional machine learning models use supervised, unsupervised, or semi-supervised learning with fixed features.
• Human and machine learning can benefit from self-directed learning with novel features beyond initial training.
• Existing models struggle to integrate new information without compromising previously learned knowledge.

Purpose of the Study:

• To introduce a new self-supervised learning paradigm and a neural network, self-supervised ARTMAP.
• To enable learning from novel features in unlabeled data without degrading existing knowledge from labeled data.
• To enhance machine learning models' ability to adapt and learn in dynamic environments.

Main Methods:

• Developed a novel neural network architecture: self-supervised ARTMAP.
• Integrated three knowledge sources: labeled patterns (teacher), unlabeled patterns (environment), and self-labeled patterns (internal activation).
• Implemented a category selection function and distributed network activation for scalable, confident learning on novel features.

Main Results:

• Self-supervised ARTMAP successfully learns novel features from unlabeled data without forgetting previously acquired knowledge.
• The model demonstrates improved test accuracy on both low-dimensional and high-dimensional benchmark datasets.
• Distributed learning on unlabeled patterns refines ambiguous classification boundaries.

Conclusions:

• Self-supervised ARTMAP offers a robust framework for self-supervised learning, enhancing adaptability and knowledge integration.
• The model's ability to learn from diverse data sources and novel features represents a significant advancement in machine learning.
• This approach holds promise for more sophisticated and human-like learning systems.