News for hackers. Stuff that matters.

A comment on Why do CNNs generalize so poorly to small image transformations?.

Re:Why do CNNs generalize so poorly to small image transform... by John_KZ on Sunday June 10, 2018 @06:05PM (#17280427)
Let me save you some time on why:

>While VGG16 has 5 pooling operations in its 16 layers, Resnet50 has only one pooling operation among its 50 intermediate layers and InceptionResnetV2 has only 5 among its intermediate 134 layers.

Aka being computationally cheap with pooling causes weird sampling issues.

The paper has a terrible pompous title that's just wrong. Modern CNNs generalize wonderfully on small translations. They just managed to break a couple of old CNNs and go on to claim they broke AI research or something.

[ Parent | on HN ]
Re:Why do CNNs generalize so poorly to small image transform... by microtherion on Monday June 11, 2018 @03:39PM (#17287787)
The paper claims the opposite: "[...] jaggedness is greater for the modern, deeper, networks compared to the less modern VGG16 network. While the deeper networks have better test accuracy, they are also less invariant."
[ Parent | on HN ]