Lifelong Classification in Open World With Limited Storage Requirements

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Classical classification methods struggle with incremental datasets and emerging classes in open-world scenarios.
• Existing solutions often require computationally expensive retraining and significant storage for historical data.

Purpose of the Study:

• To enhance lifelong classification performance in open-world environments.
• To develop a method that efficiently handles emerging classes and reduces learning costs.

Main Methods:

• Decomposition of the problem into three sub-tasks: rejecting unknown instances, classifying accepted instances, and minimizing learning costs.
• Utilizing outlier detection for identifying and rejecting unknown instances.
• Employing a variant artificial neural network with reduced weights for classification.

Main Results:

• Demonstrated effectiveness of the proposed approach through experimental results.
• Successful rejection of unknown instances, leading to reduced retraining computation.
• Efficient classification of accepted instances with a cost-effective learning process.

Conclusions:

• The proposed method offers an effective solution for lifelong classification in open-world settings.
• The approach successfully mitigates the challenges of incremental data and emerging classes.
• Significant reduction in computational and storage costs associated with lifelong learning.