Layer-Wise Relevance Propagation

Layer-wise relevance propagation is a method for understanding deep neural networks that uses a particular design path to observe how the individual layers of the program work.

These types of techniques help engineers to learn more about how neural networks do what they do, and they are crucial in combating the problem of "black box operation" in artificial intelligence, where technologies become so powerful and complex that it's hard for humans to understand how they produce results.

Specifically, experts contrast layer-wise relevance propagation with a deepLIFT model which uses backpropagation to examine activation differences between artificial neurons in various layers of the deep network. Some describe layer-wise relevance propagation as a deepLIFT method that sets all reference activations of artificial neurons to the same baseline for analysis.

Techniques like layer-wise relevance propagation, deepLIFT and LIME can be attached to Shapley regression and sampling techniques and other processes that work to provide additional insight into machine learning algorithms.