An Appraisal of Incremental Learning Methods.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Incremental learning, a subset of machine learning, aims for continuous knowledge acquisition from new data without accessing original datasets.
• A key challenge is catastrophic forgetting, where deep neural networks trained on stationary data lose prior knowledge when encountering new, different data distributions.
• Achieving both plasticity (learning new information) and stability (retaining old information) is crucial for effective incremental learning.

Purpose of the Study:

• To provide a systematic review of state-of-the-art incremental learning methods.
• To categorize and compare existing incremental learning strategies.
• To identify development trends and future research directions in the field.

Main Methods:

• Systematic literature search conducted on Web of Science, IEEEXplore, and DBLP databases up to May 2020.
• Categorization of reviewed papers into three main types: architectural strategy, regularization strategy, and rehearsal/pseudo-rehearsal strategy.
• Comparative analysis and discussion of different incremental learning methods based on the identified categories.

Main Results:

• Identified and categorized various incremental learning techniques, including architectural, regularization, and rehearsal-based approaches.
• Compared and discussed the strengths and weaknesses of different methods in addressing catastrophic forgetting.
• Highlighted the ongoing challenges and the dynamic nature of research in incremental learning.

Conclusions:

• Incremental learning remains a significant and active research area within artificial intelligence.
• Future research should focus on exploring both biological systems and computational models to advance incremental learning capabilities.
• Balancing knowledge acquisition and retention is key to developing robust AI systems capable of continuous learning.