9. IMAGENET-TRAINED CNNS ARE BIASED
TOWARDS TEXTURE; INCREASING SHAPE
BIAS IMPROVES ACCURACY AND
ROBUSTNESS.
@mo_takusan9922
20181205
Under review at ICLR 2019 (review scores 8,8,7)
27. 参考
? Robert Geirhos. IMAGENET-TRAINED CNNS ARE BIASED
TOWARDS TEXTURE; INCREASING SHAPE BIAS IMPROVES
ACCURACY AND ROBUSTNESS. Under review as a
conference paper at ICLR 2019. 2018
? @karpathy. t-SNE visualization of CNN. Results
codeshttps://cs.stanford.edu/people/karpathy/cnnembed/.
(参照日: 2018/12/02)
? Stanford Vision Lab, Stanford University, Princeton
University.ImageNet. http://www.image-net.org/ .2016
Editor's Notes
#15: for VGG-16, BagNet-9/17/33 reach 0.70 / 0.79 / 0.88