[1805.12177] Why do deep convolutional networks generalize so poorly to small image transformations?

Deep convolutional network architectures are often assumed to guarantee generalization for small image translations and deformations. In this paper we show that modern CNNs (VGG16, ResNet50, and InceptionResNetV2) can drastically change their output when an image is translated in the image plane by a few pixels, and that this failure of generalization also happens with other realistic small image transformations. Furthermore, the deeper the network the more we see these failures to generalize. We show that these failures are related to the fact that the architecture of modern CNNs ignores the classical sampling theorem so that generalization is not guaranteed. We also show that biases in the statistics of commonly used image datasets makes it unlikely that CNNs will learn to be invariant to these transformations. Taken together our results suggest that the performance of CNNs in object recognition falls far short of the generalization capabilities of humans.

1 mentions: @beenwrekt
Date: 2019/06/07 15:48

Referring Tweets

@beenwrekt This work follows up on inspirational work by Azulay and Weiss (https://t.co/6WKVcUdrJs) and Gu et al. (https://t.co/5lPCdU82Bz) which highlighted the variability of image classifiers over videos. (2/4)

Related Entries

Read more CVPR2018のPointCloudのCNN論文とSPLATNet
0 users, 0 mentions 2018/07/07 09:24
Read more GitHub - jiesutd/NCRFpp: NCRF++, an Open-source Neural Sequence Labeling Toolkit. It includes charac...
0 users, 0 mentions 2018/06/16 13:30
Read more CNNの学習に最高の性能を示す最適化手法はどれか - 俺とプログラミング
0 users, 0 mentions 2018/04/22 03:40
Read more 物体検出、セグメンテーションをMask R-CNNで理解してみる (初心者) - QiitaQiita
2 users, 2 mentions 2019/01/18 06:46
Read more CNNで抽出した特徴を使った画像検索 - まんぼう日記
0 users, 0 mentions 2018/04/22 03:41