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

Authors: Roland S. Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas Wallis, Wieland Brendel

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here, we test this hypothesis by asking humans to decide which of two square occlusions causes a larger change to a unit s activation. Both a large-scale crowdsourced experiment and measurements with experts show that on average the extremely activating feature visualizations by Olah et al. [40] indeed help humans on this task (68 4 % accuracy; baseline performance without any visualizations is 60 3 %). ... We run an extensive psychophysical experiment with more than 12, 000 trials distributed over 323 crowdsourced participants on Amazon Mechanical Turk (MTurk) and two experts
Researcher Affiliation Academia 1 Tübingen AI Center, University of Tübingen, Germany. 2 Institute of Psychology and Centre for Cognitive Science, Technical University of Darmstadt, Germany.
Pseudocode No No pseudocode or algorithm blocks are explicitly presented or labeled in the paper.
Open Source Code Yes Code and data are available at github.com/brendel-group/causal-understanding-via-visualizations.
Open Datasets Yes To generate stimuli, we follow Olah et al. [40] and use an Inception V1 network [53] trained on Image Net [12, 49]. ... Natural references: The reference images are the most strongly activating3 dataset samples from Image Net [12, 49].
Dataset Splits No The paper mentions using ImageNet and a 'random subset of the training set (∼50%)' for activation computations, but does not specify the train/validation/test dataset splits used for training the InceptionV1 network or for partitioning the data in their human psychophysical experiment.
Hardware Specification No The paper does not explicitly describe the hardware used for running its experiments, such as specific GPU or CPU models.
Software Dependencies No The paper mentions using an 'Inception V1 network' and refers to other methods, but does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks.
Experiment Setup Yes To generate stimuli, we follow Olah et al. [40] and use an Inception V1 network [53] trained on Image Net [12, 49]. ... We test units sampled from 9 layers and 2 Inception module branches (namely 3x3 and POOL). ... To generate query images, we place a square patch of 90 × 90 pixels of the average RGB color of the occluded pixels... We test the five different reference image types as separate experimental conditions. In each condition, we collect data from a total of 50 different MTurk participants...