Do Users Benefit From Interpretable Vision? A User Study, Baseline, And Dataset

Authors: Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiß, Tim Landgraf

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted a user study (N=240) to test how such a baseline explanation technique performs against concept-based and counterfactual explanations. ... Our results show that the baseline outperformed concept-based explanations.
Researcher Affiliation Academia Leon Sixt 1, Martin Schuessler 23, Oana-Iuliana Popescu1, Philipp Weiß3, Tim Landgraf1 Freie Universität Berlin1 Weizenbaum Institut Berlin2 TU Berlin3
Pseudocode Yes Listing 1: Source code example to create a biased sampler.
Open Source Code Yes We open-source our dataset, explanation techniques, model, study design, including instructions and videos to support replicating our results as well as adapting our design to other explanation techniques. ... Model export: https://f002.backblazeb2.com/file/iclr2022/do_users_benefit_from_interpretable_vision_model.tar.gz
Open Datasets Yes We open-source our dataset, explanation techniques, model, study design, including instructions and videos to support replicating our results as well as adapting our design to other explanation techniques. ... Biased dataset: https://f002.backblazeb2.com/file/iclr2022/two4two_obj_color_and_spherical_finer_search_spherical_uniform_0.33_uniform_0.15.tar
Dataset Splits No B: NN predictions explained with 10 sorted rows of 5 images drawn from the validation set (50 original images).
Hardware Specification No We thank the Center for Information Services and High Performance Computing (ZIH) at Dresden University of Technology and the HPC Service of ZEDAT, Freie Universität Berlin, for generous allocations of computation time (Bennett et al., 2020).
Software Dependencies No The paper mentions architectures like Glow and MobileNet V2 but does not provide specific software dependencies with version numbers.
Experiment Setup Yes Our model is based on the Glow architecture (Kingma & Dhariwal, 2018) and contains 7 blocks. A block is a collection of 32 flow steps, followed by a down-sampling layer, and ends with a fade-out layer. A single flow step consists of actnorm, invertible 1x1 convolution and affine coupling layer. ... The model is trained using a supervised loss and an unsupervised objective. ... For the supervised loss Lsup, we use the binary cross entropy. As unsupervised loss Lun, we use the commonly used standard flow loss... We ran the matrix factorization with 10 components and selected the five components that correlated most with the logit score (r is in the range [0.21, 0.34]).