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

A precise understanding of why units in an artificial network respond tocertain stimuli would constitute a big step towards explainable artificialintelligence. One widely used approach towards this goal is to visualize unitresponses via activation maximization. These synthetic feature visualizationsare purported to provide humans with precise information about the imagefeatures that cause a unit to be activated - an advantage over otheralternatives like strongly activating natural dataset samples. If humans indeedgain causal insight from visualizations, this should enable them to predict theeffect of an intervention, such as how occluding a certain patch of the image(say, a dog's head) changes a unit's activation. Here, we test this hypothesisby asking humans to decide which of two square occlusions causes a largerchange to a unit's activation. Both a large-scale crowdsourced experiment andmeasurements with experts show that on average the extremely activating featurevisualizations by Olah et al. (2017) indeed help humans on this task ($68 \pm4$% accuracy; baseline performance without any visualizations is $60 \pm 3$%).However, they do not provide any substantial advantage over othervisualizations (such as e.g. dataset samples), which yield similar performance($66\pm3$% to $67 \pm3$% accuracy). Taken together, we propose an objectivepsychophysical task to quantify the benefit of unit-level interpretabilitymethods for humans, and find no evidence that a widely-used featurevisualization method provides humans with better "causal understanding" of unitactivations than simple alternative visualizations.

How Well do Feature Visualizations Support Causal Understanding of CNN Activations?