Convolutional Neural Network With Developmental Memory for Continual Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are pivotal in computer vision but suffer from catastrophic forgetting during sequential task learning.
• Catastrophic forgetting hinders the ability of CNNs to retain previously acquired knowledge when learning new tasks.

Purpose of the Study:

• To introduce a novel approach to mitigate catastrophic forgetting in CNNs for continual learning.
• To enhance sequential learning capabilities of CNNs without compromising performance on previously learned tasks.

Main Methods:

• Developmental Memory (DM): A system generating submemory networks for individual tasks within a CNN.
• Guided Learning (GL): A training method that directs new submemories to specialize in novel tasks.
• Preservation Mechanism: Existing submemories are adapted to retain knowledge of older tasks.

Main Results:

• The proposed CNN with DM demonstrated improved performance on new image classification tasks.
• The system significantly reduced the forgetting of previously learned tasks compared to state-of-the-art methods.
• Experimental results validate the effectiveness of DM and GL in continual learning scenarios.

Conclusions:

• The integration of Developmental Memory and Guided Learning effectively addresses catastrophic forgetting in CNNs.
• This approach facilitates robust continual learning, enhancing both new task acquisition and old task retention.
• The proposed method offers a promising solution for advancing sequential learning in deep neural networks.