Representation of Imprecision in Deep Neural Networks for Image Classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning uncertainty quantification often overlooks imprecision characterization.
• Some tasks benefit from imprecise results over costly errors.

Purpose of the Study:

• To investigate Representation of Imprecision in Deep-learning (RIDL) techniques.
• To characterize and manage uncertainty-induced imprecision in deep learning models.

Main Methods:

• Reconstructing training image labels using neural networks and a novel label assignment rule.
• Revising multi-labeled images to correct labeling errors while retaining knowledge-driven imprecision.
• Retraining deep network models with the reconstructed dataset and classifying test images.

Main Results:

• RIDL effectively characterizes imprecision in both training and testing datasets.
• The proposed method improves classification accuracy (AC).
• Imprecise test images are assigned to meta-categories, representing ambiguity.

Conclusions:

• RIDL offers a novel approach to managing uncertainty in deep learning.
• The technique enhances model performance and provides meaningful representation of imprecision.
• This method is valuable for applications where understanding ambiguity is crucial.