Top1 error reported on validation set. We learn transform parameters on validation set also.

Learn on small mnist

(1) Baseline top1: 0.890

(2) \w Learned augmentation: 0.900

Shift Rotation Scale

Learn on small mnist (rotated at random by max. 360)

I left 10k training examples out of 60k in original MNIST.

I did 2 trials:

Shift plot has error.

Learn regular mnist

Two trials: First:

Baseline top1: 0.380

\w Learned augmentation: 0.340

Second:

Baseline top1: 0.420

\w Learned augmentation: 0.380

Shift Rotation Scale

Learn mnist-rot

Standard mnist-rotated dataset

Trial 1:

Baseline top1: 4.280

\w Learned augmentation: 3.524

Trial 2:

Baseline top1: 4.132

\w Learned augmentation: 3.320

Shift Rotation Scale