DIVA: Dataset Derivative of a Learning Task

Authors: Yonatan Dukler, Alessandro Achille, Giovanni Paolini, Avinash Ravichandran, Marzia Polito, Stefano Soatto

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

Reproducibility Variable Result LLM Response
Research Type Experimental To illustrate the flexibility of DIVA, we report experiments on sample auto-curation tasks such as outlier rejection, dataset extension, and automatic aggregation of multi-modal data.
Researcher Affiliation Collaboration Yonatan Dukler1,2 , Alessandro Achille1, Giovanni Paolini1, Avinash Ravichandran1, Marzia Polito1, Stefano Soatto1 1 Amazon Web Services, {aachille, paoling, ravinash, mpolito, soattos}@amazon.com 2 Department of Mathematics, University of California, Los Angeles ydukler@math.ucla.edu
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code No The paper does not provide an explicit statement about the release of source code or a link to a code repository for the described methodology.
Open Datasets Yes For our experiments we use the CUB-200 (Welinder et al., 2010), FGVC-Aircraft, (Maji et al., 2013), Stanford Cars (Krause et al., 2013), Caltech-256 (Griffin et al., 2007), Oxford Flowers 102 (Nilsback & Zisserman, 2008), MIT-67 Indoor (Quattoni & Torralba, 2009), Street View House Number (Netzer et al., 2011), and the Oxford Pets (Parkhi et al., 2012) visual recognition and classification datasets.
Dataset Splits Yes For Ren et al. (2018), we set aside 20% of the training samples as validation for the reweight step, but use all samples for the final training (in parentheses). We can, of course, compute a validation loss using a separate validation set. However, as we will see in Section 3.3, we can also use a leave-one-out cross-validation loss directly on the training set, without any requirement of a separate validation set.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions using specific models and pre-trained networks but does not provide specific software dependencies with version numbers (e.g., programming language, libraries, or frameworks with their respective versions) required for replication.
Experiment Setup Yes In all experiments, we use the network as a fixed feature extractor, and train a linear classifier on top of the network features using the weighted L2 loss eq. (5) and optimize the weights using DIVA. ... We apply only 1-3 gradient optimization steps with a relatively large learning rate ' 0.1. This early stopping both regularizes the solution and decreases the wall-clock time required by the method. We initialize so that i = 1 for all samples.