Toward Training Recurrent Neural Networks for Lifelong Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Parametric lifelong learning systems face significant challenges with catastrophic forgetting and capacity saturation.
• Recurrent neural networks (RNNs) are crucial for sequential supervised learning tasks.

Purpose of the Study:

• To investigate catastrophic forgetting and capacity saturation in RNNs within a lifelong learning context.
• To develop and evaluate a unified approach combining memory and capacity expansion for RNNs.

Main Methods:

• Proposed a curriculum-based benchmark for evaluating lifelong learning models with increasing task difficulty.
• Integrated gradient episodic memory (GEM) for mitigating catastrophic forgetting with Net2Net for capacity expansion.
• Assessed the feasibility of these methods for recurrent neural networks.

Main Results:

• The unified model, combining GEM and Net2Net, showed improved suitability for lifelong learning compared to individual methods.
• Evaluation on the proposed benchmark demonstrated the effectiveness of the integrated approach in addressing core lifelong learning challenges.
• The unified model proved more effective in recurrent network settings.

Conclusions:

• The unified approach offers a promising direction for developing more robust lifelong learning systems.
• Addressing both forgetting and capacity limitations is essential for advancing sequential learning in RNNs.
• The proposed benchmark provides a valuable tool for future lifelong learning research.