English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Recent years have witnessed an explosion in the development of novelprediction-based attribution methods, which have slowly been supplanting oldergradient-based methods to explain the decisions of deep neural networks.However, it is still not clear why prediction-based methods outperformgradient-based ones. Here, we start with an empirical observation: these twoapproaches yield attribution maps with very different power spectra, withgradient-based methods revealing more high-frequency content thanprediction-based methods. This observation raises multiple questions: What isthe source of this high-frequency information, and does it truly reflectdecisions made by the system? Lastly, why would the absence of high-frequencyinformation in prediction-based methods yield better explainability scoresalong multiple metrics? We analyze the gradient of three representative visualclassification models and observe that it contains noisy information emanatingfrom high-frequencies. Furthermore, our analysis reveals that the operationsused in Convolutional Neural Networks (CNNs) for downsampling appear to be asignificant source of this high-frequency content -- suggesting aliasing as apossible underlying basis. We then apply an optimal low-pass filter forattribution maps and demonstrate that it improves gradient-based attributionmethods. We show that (i) removing high-frequency noise yields significantimprovements in the explainability scores obtained with gradient-based methodsacross multiple models -- leading to (ii) a novel ranking of state-of-the-artmethods with gradient-based methods at the top. We believe that our resultswill spur renewed interest in simpler and computationally more efficientgradient-based methods for explainability.

Gradient strikes back: How filtering out high frequencies improves explanations