Unshuffling Data for Improved Generalization in Visual Question Answering

Simple item page
dc.contributor.authorTeney, D.
dc.contributor.authorAbbasnejad, E.
dc.contributor.authorvan den Hengel, A.
dc.contributor.conferenceIEEE/CVF International Conference on Computer Vision (ICCV) (11 Oct 2021 - 17 Oct 2021 : Virtual Online)
dc.date.issued2021
dc.description.abstractGeneralization beyond the training distribution is a core challenge in machine learning. The common practice of mixing and shuffling examples when training neural networks may not be optimal in this regard. We show that partitioning the data into well-chosen, non-i.i.d. subsets treated as multiple training environments can guide the learning of models with better out-of-distribution generalization. We describe a training procedure to capture the patterns that are stable across environments while discarding spurious ones. The method makes a step beyond correlation-based learning: the choice of the partitioning allows injecting information about the task that cannot be otherwise recovered from the joint distribution of the training data. We demonstrate multiple use cases with the task of visual question answering, which is notorious for dataset biases. We obtain significant improvements on VQA-CP, using environments built from prior knowledge, existing meta data, or unsupervised clustering. We also get improvements on GQA using annotations of “equivalent questions”, and on multidataset training (VQA v2 / Visual Genome) by treating them as distinct environments.
dc.description.statementofresponsibilityDamien Teney, Ehsan Abbasnejad, Anton van den Hengel
dc.identifier.citationProceedings / IEEE International Conference on Computer Vision. IEEE International Conference on Computer Vision, 2021, pp.1397-1407
dc.identifier.doi10.1109/ICCV48922.2021.00145
dc.identifier.isbn9781665428125
dc.identifier.issn1550-5499
dc.identifier.orcidTeney, D. [0000-0003-2130-6650]
dc.identifier.orcidvan den Hengel, A. [0000-0003-3027-8364]
dc.identifier.urihttps://hdl.handle.net/2440/136312
dc.language.isoen
dc.publisherIEEE
dc.publisher.placeLos Alamitos, CA
dc.relation.ispartofseriesIEEE International Conference on Computer Vision
dc.rights©2021 IEEE
dc.source.urihttps://ieeexplore.ieee.org/xpl/conhome/9709627/proceeding
dc.titleUnshuffling Data for Improved Generalization in Visual Question Answering
dc.typeConference paper
pubs.publication-statusPublished

Files

Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
hdl_136312.pdf
Size:
2.1 MB
Format:
Adobe Portable Document Format
Description:
Submitted version
Download

Collections

Australian Institute for Machine Learning publications