A flat network classifies grayscale handwritten digits, to . Each image is unrolled row by row into a list of numbers, and the first ordinary layer has its own separate weight from every one of those positions to every hidden unit. Trained on the images as they are, it reaches 97% test accuracy.
Four more runs use the same architecture, the same labels and the same amount of training, each starting from scratch. The only change is that every image is rearranged before it is flattened:
| run | training images | test images |
|---|---|---|
| P | pixels scrambled by one fixed scramble, chosen once at random and then used on every image | the same fixed scramble |
| Q | each image scrambled by a fresh random scramble, drawn anew every time the image is used | a fresh random scramble for every image |
| R | one fixed scramble | a different fixed scramble |
| M | every image mirrored left to right | every image mirrored left to right |
A scramble only moves pixels to new positions. Every brightness value survives, just somewhere else.
Which runs can finish with test accuracy close to 97%?
Select all that apply.