Self-Net: Lifelong Learning via Continual Self-Modeling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Neural Networks

Background:

• Continual learning (CL) in AI faces challenges with storing past knowledge.
• Existing methods require new networks per task, store old data, or limit learning.
• Efficiently adapting AI models to new tasks over time is crucial.

Purpose of the Study:

• To propose a novel framework, Self-Net, for effective continual learning.
• To address limitations of current CL approaches regarding storage and adaptability.
• To enable AI systems to learn new tasks sequentially without forgetting previous ones.

Main Methods:

• Developed Self-Net, a framework utilizing autoencoders.
• Autoencoders learn low-dimensional representations of task-specific network weights.
• These representations are used to reconstruct original weights for task recall.

Main Results:

• Achieved over 10X storage compression for continual learning.
• Demonstrated high-fidelity reconstruction of network weights.
• Outperformed state-of-the-art methods on multiple benchmark datasets (MNIST, CIFAR, Atari, CORe50).

Conclusions:

• Self-Net offers an efficient solution for continual learning in deep neural networks.
• The autoencoder-based approach enables learning new tasks with minimal retraining and no prior data storage.
• This method significantly advances the field of sequential learning in AI.