Illuminating Salient Contributions in Neuron Activation With Attribution Equilibrium.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep neural networks (DNNs) achieve high performance but lack transparency.
• Interpreting DNN decision-making is crucial for trust and debugging.
• Existing attribution methods offer limited insights into evidence conservation.

Purpose of the Study:

• Introduce Attribution Equilibrium, a novel method for decomposing DNN predictions.
• Provide fine-grained attributions by balancing positive and negative relevance.
• Enhance visualization of evidence supporting network decisions.

Main Methods:

• Define evidence as the gap between positive and negative influences in gradient-derived maps.
• Incorporate antagonistic elements and user-defined positive attribution criteria.
• Consider inactivated neurons in propagation to discern less relevant features like background.

Main Results:

• Attribution Equilibrium decomposes predictions into fine-grained attributions.
• The method balances positive and negative relevance for clearer evidence visualization.
• Evaluations on PASCAL VOC 2007, MS COCO 2014, and ImageNet datasets show superior performance.

Conclusions:

• Attribution Equilibrium outperforms existing methods in identifying key input features.
• The novel approach enhances the discernment of relevant and irrelevant input features.
• This method offers a clearer perspective on the conservation of evidence in DNNs.