Incremental Concept Learning via Online Generative Memory Recall.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks struggle with catastrophic forgetting when learning new data incrementally.
• This forgetting occurs due to the absence of past data and alterations in neural weights during new concept learning.

Purpose of the Study:

• To develop an incremental concept learning framework that addresses catastrophic forgetting.
• To enable lifelong learning systems capable of retaining previously acquired knowledge.

Main Methods:

• Proposed a novel incremental concept learning framework with ICLNet and RecallNet components.
• ICLNet utilizes a trainable feature extractor and dynamic concept memory matrix.
• Introduced a concept-contrastive loss and a balanced online memory recall strategy.

Main Results:

• The proposed approach effectively mitigates catastrophic forgetting in deep neural networks.
• Experiments on MNIST, Fashion-MNIST, and SVHN datasets demonstrated superior performance compared to pseudorehearsal methods.
• The framework successfully consolidates old concept memory and recalls pseudo-samples.

Conclusions:

• The developed incremental concept learning framework significantly reduces catastrophic forgetting.
• The combination of ICLNet and RecallNet with proposed strategies offers an effective solution for lifelong learning systems.
• The approach shows strong potential for real-world applications requiring continuous learning.