Incremental Deep Neural Network Learning Using Classification Confidence Thresholding

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Modern neural networks often fail to handle unknown classes in classification tasks.
• Testing typically occurs in closed-set scenarios, not reflecting real-world open environments.
• Incremental learning aims to enable models to identify and incorporate new classes autonomously.

Purpose of the Study:

• To propose a novel approach for incremental learning in neural networks.
• To address challenges of resource inefficiency and accuracy degradation in incremental learning.
• To maintain high classification accuracy while limiting forgetting during class expansion.

Main Methods:

• Introduction of the classification confidence threshold (CT) approach.
• Implementation of a lean method to optimize retraining resource utilization.
• Enabling incremental learning with limited samples of new classes.

Main Results:

• The CT approach helps maintain high classification accuracies by limiting forgetting.
• The lean method reduces computational resources required for retraining.
• The proposed method is adaptable to most existing neural network architectures.

Conclusions:

• The classification confidence threshold (CT) approach effectively primes neural networks for incremental learning.
• This method offers a resource-efficient solution for handling unknown classes in dynamic environments.
• The approach allows neural networks to incrementally learn new classes with minimal architectural changes.