Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: Restricting the Flow: Information Bottlenecks for Attribution
(Submitted on 2 Jan 2020 (v1), last revised 25 May 2020 (this version, v4))
Abstract: Attribution methods provide insights into the decision-making of machine learning models like artificial neural networks. For a given input sample, they assign a relevance score to each individual input variable, such as the pixels of an image. In this work we adapt the information bottleneck concept for attribution. By adding noise to intermediate feature maps we restrict the flow of information and can quantify (in bits) how much information image regions provide. We compare our method against ten baselines using three different metrics on VGG-16 and ResNet-50, and find that our methods outperform all baselines in five out of six settings. The method's information-theoretic foundation provides an absolute frame of reference for attribution values (bits) and a guarantee that regions scored close to zero are not necessary for the network's decision. For reviews: this https URL For code: this https URL
Submission history
From: Leon Sixt [view email][v1] Thu, 2 Jan 2020 11:24:35 GMT (7629kb,D)
[v2] Sat, 15 Feb 2020 18:37:23 GMT (6385kb,D)
[v3] Thu, 7 May 2020 17:31:52 GMT (6386kb,D)
[v4] Mon, 25 May 2020 14:21:37 GMT (6386kb,D)
Link back to: arXiv, form interface, contact.