search

dorimedini / robustness_properties

GLOBIGLOBIGLOBI
0 stars 0 forks source link

Is MNIST linear in its bones #60

Open dorimedini opened 5 years ago

dorimedini commented 5 years ago

We think FCN (N-layer FC nets) on MNIST will just try to become 1-layer because MNIST is a linear-solvable problem. Try, say, fitting a 5 layer FC net with all but the first layer frozen.

galshachaf commented 5 years ago

But we know it's linear-solvable. Amir mentioned it on the first lecture.

dorimedini commented 5 years ago

Yes, but that's not exactly the question... from what I understand we also want to know if, given a more powerful model topology, the resulting trained model is in fact just a one-layer linear classifier with added noise before output. Right? @noam