rotating MNIST

[murphyk]

• Use (sin(a), cos(a)) as output targets to handle periodic boundary conditions
• For eval, measure performance on some range (eg input images from -45 to 45) and also the current task
• Compare SGD+RB, offline SGD, Lofi with q in {0, 0.1} x alpha in {0, 0.1}, gamma=1, R=fixed, eta_0=0.01 (say)

The paper below uses a CNN with target ouputs $[max(0, α), max(0, −α)]$. and uses L1 loss. They can predict orientation of natural images to within 3 degrees if the input is rotated within -30..30 degrees.

P. Fischer, A. Dosovitskiy, and T. Brox, “Image Orientation Estimation with Convolutional Networks,” in German Conference on Pattern Recognition, 2015, pp. 368–378, doi: 10.1007/978-3-319-24947-6_30 [Online]. Available: https://lmb.informatik.uni-freiburg.de/Publications/2015/FDB15/image_orientation.pdf

Other links

• https://github.com/ZJCV/RotNet/tree/master
• https://d4nst.github.io/2017/01/12/image-orientation

[murphyk]

Also try classification as a task (with the same input data)