%0 Generic %1 zhai2019visual %A Zhai, Xiaohua %A Puigcerver, Joan %A Kolesnikov, Alexander %A Ruyssen, Pierre %A Riquelme, Carlos %A Lucic, Mario %A Djolonga, Josip %A Pinto, Andre Susano %A Neumann, Maxim %A Dosovitskiy, Alexey %A Beyer, Lucas %A Bachem, Olivier %A Tschannen, Michael %A Michalski, Marcin %A Bousquet, Olivier %A Gelly, Sylvain %A Houlsby, Neil %D 2019 %K from:pmorales representation %T The Visual Task Adaptation Benchmark %U http://arxiv.org/abs/1910.04867 %X Representation learning promises to unlock deep learning for the long tail of vision tasks without expansive labelled datasets. Yet, the absence of a unified yardstick to evaluate general visual representations hinders progress. Many sub-fields promise representations, but each has different evaluation protocols that are either too constrained (linear classification), limited in scope (ImageNet, CIFAR, Pascal-VOC), or only loosely related to representation quality (generation). We present the Visual Task Adaptation Benchmark (VTAB): a diverse, realistic, and challenging benchmark to evaluate representations. VTAB embodies one principle: good representations adapt to unseen tasks with few examples. We run a large VTAB study of popular algorithms, answering questions like: How effective are ImageNet representation on non-standard datasets? Are generative models competitive? Is self-supervision useful if one already has labels?

@misc{zhai2019visual, abstract = {Representation learning promises to unlock deep learning for the long tail of vision tasks without expansive labelled datasets. Yet, the absence of a unified yardstick to evaluate general visual representations hinders progress. Many sub-fields promise representations, but each has different evaluation protocols that are either too constrained (linear classification), limited in scope (ImageNet, CIFAR, Pascal-VOC), or only loosely related to representation quality (generation). We present the Visual Task Adaptation Benchmark (VTAB): a diverse, realistic, and challenging benchmark to evaluate representations. VTAB embodies one principle: good representations adapt to unseen tasks with few examples. We run a large VTAB study of popular algorithms, answering questions like: How effective are ImageNet representation on non-standard datasets? Are generative models competitive? Is self-supervision useful if one already has labels?}, added-at = {2019-10-19T15:33:23.000+0200}, author = {Zhai, Xiaohua and Puigcerver, Joan and Kolesnikov, Alexander and Ruyssen, Pierre and Riquelme, Carlos and Lucic, Mario and Djolonga, Josip and Pinto, Andre Susano and Neumann, Maxim and Dosovitskiy, Alexey and Beyer, Lucas and Bachem, Olivier and Tschannen, Michael and Michalski, Marcin and Bousquet, Olivier and Gelly, Sylvain and Houlsby, Neil}, biburl = {https://www.bibsonomy.org/bibtex/224a78fb1976d5a297c577b85c8b45a8b/groupxpapers}, description = {The Visual Task Adaptation Benchmark}, interhash = {da80e9102e3fb130114999fee73f604b}, intrahash = {24a78fb1976d5a297c577b85c8b45a8b}, keywords = {from:pmorales representation}, note = {cite arxiv:1910.04867}, timestamp = {2019-10-19T15:33:23.000+0200}, title = {The Visual Task Adaptation Benchmark}, url = {http://arxiv.org/abs/1910.04867}, year = 2019 }