Can linear problems survive training with all but the first layer frozen?

dorimedini / robustness_properties

GLOBIGLOBIGLOBI

0 stars 0 forks source link

Can linear problems survive training with all but the first layer frozen? #109

Closed dorimedini closed 5 years ago

dorimedini commented 5 years ago

Compare accuracy results between:

Vanilla trained FCN model
FCN model, with all layers frozen except the first

Hypothesis: MNIST should take a small accuracy penalty, CIFAR should take a larger penalty. In general, linear problems resist freezing, non-linear problems... less so.