Towards global neural network abstractions with locally-exact reconstruction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Network Analysis

Background:

• Neural networks are powerful non-linear functions but their black-box nature hinders explainability and safety certification.
• Existing abstraction techniques offer limited precision, restricting their use to small input regions.

Purpose of the Study:

• To develop a novel abstraction technique for neural networks that provides sound over-approximation bounds globally.
• To ensure exact reconstructions for any local input within the abstracted domain.

Main Methods:

• Propose Global Interval Neural Network Abstractions with Center-Exact Reconstruction (GINNACER).
• Develop a method for sound over-approximation bounds applicable to the entire input domain.
• Incorporate exact reconstruction capabilities for local inputs.

Main Results:

• GINNACER achieves sound over-approximation bounds across the whole input domain.
• The technique guarantees exact reconstructions for specific local inputs.
• Experimental results demonstrate GINNACER is orders of magnitude tighter than existing global methods and competitive with local ones.

Conclusions:

• GINNACER significantly improves the precision of global neural network abstractions.
• This technique enhances the explainability and safety certification of neural networks.
• GINNACER offers a more effective solution for analyzing neural network behavior over their entire input space.