Continual learning for recurrent neural networks: An empirical evaluation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Deploying machine learning solutions requires robustness to data distribution drifts.
• Continual Learning (CL) with recurrent neural networks (RNNs) is essential for non-stationary data, common in NLP and robotics.
• Existing CL research for sequential data is fragmented, using varied protocols and datasets.

Purpose of the Study:

• To organize and categorize existing literature on CL for sequential data processing.
• To introduce two novel benchmarks for CL with sequential data, simulating real-world conditions.
• To empirically evaluate CL strategies with RNNs in a class-incremental setting.

Main Methods:

• Literature review and categorization of CL contributions for sequential data.
• Development of two new CL benchmarks using existing datasets.
• Empirical evaluation of various CL strategies with RNNs on the proposed benchmarks.

Main Results:

• The study provides a structured overview of CL for sequential data.
• New benchmarks were introduced to facilitate standardized evaluation.
• Empirical results underscore the significant impact of sequence length on model performance.
• The importance of clearly defining the CL scenario was emphasized.

Conclusions:

• Standardized benchmarks and clear scenario definitions are vital for advancing CL research in sequential data processing.
• RNNs combined with appropriate CL strategies show promise for applications with evolving data distributions.
• Further research is needed to develop more robust and generalizable CL methods for sequential data.