3DB: A Framework for Debugging Computer Vision Models

Authors: Guillaume Leclerc, Hadi Salman, Andrew Ilyas, Sai Vemprala, Logan Engstrom, Vibhav Vineet, Kai Xiao, Pengchuan Zhang, Shibani Santurkar, Greg Yang, Ashish Kapoor, Aleksander Madry

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate, through a wide range of use cases, that 3DB allows users to discover vulnerabilities in computer vision systems and gain insights into how models make decisions. In all our experiments, we analyze a Res Net-18 [30] trained on the Image Net [53] classification task.
Researcher Affiliation Collaboration Guillaume Leclerc LECLERC@MIT.EDU Hadi Salman HADY@MIT.EDU Andrew Ilyas AILYAS@MIT.EDU Sai Vemprala SAIHV@MICROSOFT.COM Microsoft Research Logan Engstrom ENGSTROM@MIT.EDU Vibhav Vineet VIVINEET@MICROSOFT.COM Microsoft Research Kai Xiao KAIX@MIT.EDU Pengchuan Zhang PENZHAN@MICROSOFT.COM Microsoft Research Shibani Santurkar SHIBANI@MIT.EDU Greg Yang GE.YANG@MICROSOFT.COM Microsoft Research Ashish Kapoor AKAPOOR@MICROSOFT.COM Microsoft Research Aleksander M adry MADRY@MIT.EDU
Pseudocode No The paper describes the 3DB workflow and its components but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes We are releasing 3DB as a library1 alongside a set of examples2, guides3, and documentation4. 1https://github.com/3db/3db
Open Datasets Yes In all our experiments, we analyze a Res Net-18 [30] trained on the Image Net [53] classification task. [53] Olga Russakovsky et al. Image Net Large Scale Visual Recognition Challenge . In: International Journal of Computer Vision (IJCV). 2015.
Dataset Splits Yes In all our experiments, we analyze a Res Net-18 [30] trained on the Image Net [53] classification task (its validation set accuracy is 69.8%).
Hardware Specification No The paper states 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix B.' However, Appendix B is not provided in the given text, thus specific hardware details are not available.
Software Dependencies No The paper mentions 'Py Torch classification module' and 'Blender' but does not specify any version numbers for these or other software dependencies.
Experiment Setup No The paper specifies the model (ResNet-18) and the dataset (ImageNet) used, and describes how 3DB generates scenes with various transformations (e.g., 'random poses, orientations, and scales'), but it does not provide specific training hyperparameters such as learning rate, batch size, number of epochs, or optimizer details for the model being analyzed.