hardmaru: SketchTransfer: A Challenging New Task for Exploring Detail-Invariance and the Abstractions Learned by Deep Networks
If we train a neural net to classify CIFAR10 photos but also give it unlabelled QuickDraw doodles, how well can it classify these doodles? https://arxiv.org/abs/1912.11570 https://t.co/YwjFcmTR17
3 replies, 458 likes
Melanie Mitchell: If a neural network learns to recognize objects in photos, can it also recognize human-drawn sketches of these objects? Sort of. Cool paper exploring this question!
3 replies, 143 likes
Pablo Samuel Castro: Very cool idea and paper!
The ability to recognize the most salient features in objects, shared across images, 3D renderings, and sketches, could likely provide a step towards more general and reusable representational primitives.
1 replies, 80 likes
roadrunner01: SketchTransfer: A Challenging New Task for Exploring Detail-Invariance and the Abstractions Learned by Deep Networks
abs: https://arxiv.org/abs/1912.11570 https://t.co/v9m2zqW6XZ
0 replies, 23 likes
NIDHAL SELMI - نضال السالمي: A neural net trained on photos only gets 59% accuracy with doodles (11% is random baseline). Trained on doodles directly gives 87%. Sketches are hard because they are human abstractions. This is a good path towards reasoning since abstractions are noise/detail filters.
0 replies, 11 likes
Although humans can correctly classify even sketches that losses details, the current ML model not. In order to evaluate such generalization performance, the authors propose a task on new illustration SketchTransfer dataset with model trained on CIFAR10. https://t.co/wAEgsi4Ivq
0 replies, 2 likes
Found on Feb 12 2020 at https://arxiv.org/pdf/1912.11570.pdf