Middle-Level Features for the Explanation of Classification Systems by Sparse Dictionary Methods.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Machine learning (ML) systems often lack transparency, posing challenges for understanding their decision-making processes.
• Existing eXplainable Artificial Intelligence (XAI) methods frequently rely on low-level features, placing a significant cognitive burden on users to interpret explanations.

Purpose of the Study:

• To propose a novel model-agnostic framework to enhance the transparency of ML systems.
• To develop an XAI approach that alleviates the user's cognitive burden by utilizing perceptually salient, middle-level properties instead of low-level features.

Main Methods:

• A model-agnostic framework is introduced, instantiated for ML image classification.
• Sparse dictionary learning techniques are employed to extract perceptually salient, middle-level properties from image classification inputs.
• These middle-level properties serve as building blocks for generating explanations.

Main Results:

• The proposed framework generates parsimonious explanations based on a limited set of middle-level image properties.
• Explanations are contrastive, effectively illustrating why a specific classification was chosen over alternatives.
• The approach avoids reliance on pixel relevance maps or other low-level input features.

Conclusions:

• The developed framework offers a more intuitive and less cognitively demanding method for explaining ML image classifications.
• Its model-agnostic nature makes it adaptable to various ML systems and explanation tasks beyond image classification.