Incremental learning methods with retrieving of interfered patterns.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Neural networks often forget previously learned patterns when trained incrementally on new data.
• This forgetting occurs because network parameters adjust for new information, impacting existing memories.
• Learning new patterns with all previous data requires significant computational power, posing a practical challenge.

Purpose of the Study:

• To develop novel incremental learning methods that mitigate catastrophic forgetting in neural networks.
• To propose a system that efficiently incorporates new patterns while preserving existing knowledge.
• To enhance the generalization capabilities of neural network models in incremental learning scenarios.

Main Methods:

• The study proposes incremental learning methods with retrieving interfered patterns (ILRI).
• ILRI utilizes a modified Resource Allocating Network (RAN), a type of Generalized Radial Basis Function (GRBF) network.
• Two ILRI approaches are presented: one using a database of past patterns and another regenerating patterns without a database.

Main Results:

• Both ILRI methods demonstrated comparable performance in preserving learned patterns.
• The proposed ILRI systems exhibited superior generalization ability compared to standard neural networks and k-nearest neighbors.
• The database-free ILRI approach effectively approximated pattern regeneration, reducing computational overhead.

Conclusions:

• ILRI offers an effective solution to catastrophic forgetting in incremental neural network learning.
• The methods enhance model robustness and generalization without demanding excessive computational resources.
• ILRI provides a practical approach for systems requiring continuous learning and memory retention.