Efficient Architecture Search for Continual Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Continual learning in neural networks aims to learn sequential tasks but faces challenges like catastrophic forgetting, adaptation to new tasks, and model complexity control.
• Existing methods struggle to balance knowledge retention from previous tasks with efficient adaptation to new ones.

Purpose of the Study:

• To propose a novel approach, Continual Learning with Efficient Architecture Search (CLEAS), to overcome the limitations of traditional continual learning methods.
• To develop a fine-grained control mechanism for neural architecture search that optimizes for both performance and model simplicity.

Main Methods:

• CLEAS integrates neural architecture search (NAS) with continual learning, utilizing reinforcement learning to find optimal neural architectures for new tasks.
• A neuron-level NAS controller selectively reuses existing neurons (knowledge transfer) and adds new neurons for novel information, without altering the weights of reused neurons.
• This approach ensures perfect memorization of previously learned knowledge while adapting to new tasks.

Main Results:

• CLEAS demonstrated superior performance compared to state-of-the-art methods on various sequential classification tasks.
• The proposed method achieved higher classification accuracy.
• CLEAS resulted in significantly simpler neural architectures, indicating efficient model complexity control.

Conclusions:

• CLEAS offers an effective solution for continual learning by addressing catastrophic forgetting and model complexity.
• The neuron-level NAS controller enables efficient knowledge transfer and adaptation, leading to improved performance and parsimonious models.
• This approach advances the field of continual learning in artificial intelligence, paving the way for more robust and efficient AI systems.