Identifying Model Weakness with Adversarial Examiner

Authors: Michelle Shu, Chenxi Liu, Weichao Qiu, Alan Yuille11998-12006

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on Shape Net object classification. We show that our adversarial examiner can successfully put more emphasis on the weakness of the model, preventing performance estimates from being overly optimistic.
Researcher Affiliation Academia Michelle Shu, Chenxi Liu,( ) Weichao Qiu, Alan Yuille Johns Hopkins University {mshu1, cxliu}@jhu.edu, {qiuwch, alan.l.yuille}@gmail.com
Pseudocode Yes Algorithm 1: Adversarial Examiner Procedure
Open Source Code No The paper does not provide a link or explicit statement about the availability of open-source code for the methodology described in this paper.
Open Datasets Yes We conduct experiments on visual recognition of objects in the Shape Net dataset (Chang et al. 2015), which contains 55 classes and 51190 instances.
Dataset Splits No The paper mentions 'For each class, we choose one 3D object in the validation set that has the highest post-softmax probability on the true class.' but does not provide specific split percentages, sample counts, or clear predefined split information for reproducibility.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments were provided in the paper.
Software Dependencies No The paper mentions 'Blender software for rendering', 'Res Net34', 'Alex Net', 'Adam optimizer', and 'Bayesian Optimization package1' with a link, but does not provide specific version numbers for these or other key software components, nor does it specify the ML framework used.
Experiment Setup Yes The Res Net34 model is trained with learning rate of 0.005, and Alex Net model with 0.001, both with Adam optimizer (Kingma and Ba 2014) for 40 epochs. ... We set the learning rate to 0.001 and batch size to 32, and use Adam optimizer (Kingma and Ba 2014) to update model parameters.