None
EN
Visualizing the Impact of Feature Attribution Baselines
['Sturmfels', 'University Of Washington', 'Lundberg', 'Microsoft Research', 'Lee']
Distill
Although the original paper discusses the need for a baseline and even proposes several different baselines for image data - including the constant black image and an image of random noise - there is little existing research about the impact of this baseline. There are a myriad of methods to interpret machine learning models, including methods to visualize and understand how the network represents inputs internally , feature attribution methods that assign an importance score to each feature for a specific input , and saliency methods that aim to highlight which regions of an image the model was looking at when making a decision .