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? arxiv.org/abs/1912.11570
If we train a neural net to classify CIFAR10 photos but also give it unlabelled QuickDraw doodles, how well can it classify these doodles? arxiv.org/abs/1912.11570
@MILAMontreal @GoogleAI Answer: Better than we thought!
Recent paper by Alex Lamb, @sherjilozair, Vikas Verma + me looks motivated by abstractions learned by humans and machines.
Alex trained SOTA domain transfer methods on labelled CIFAR10 data + unlabelled QuickDraw doodles & reported his findings:
Recent paper by Alex Lamb, @sherjilozair, Vikas Verma + me looks motivated by abstractions learned by humans and machines.
Alex trained SOTA domain transfer methods on labelled CIFAR10 data + unlabelled QuickDraw doodles & reported his findings:
@MILAMontreal @GoogleAI @sherjilozair Alex found SOTA transfer methods (labelled CIFAR10 + unlabelled doodles) achieves ~60% accuracy on doodles. Supervised learning on doodles gets ~90%, leaving ~30% gap for improvement.
Surprisingly, training a model only on CIFAR10 still does quite well on ships, planes & trucks!
Surprisingly, training a model only on CIFAR10 still does quite well on ships, planes & trucks!
2. 
2. 
2. 
