Deep Learning Q&A
1 minute read
Work in progress
Why don’t we use Mean and Max Pooling?
IDEA: Probably cause the feature vector would be to long. -> GeM But what about using both and doing an autoencoder on it, the network could select and weight each feature on its own. Hybrid Pooling: https://arxiv.org/pdf/1509.06033.pdf -> Concat Mean and Max pooling performes worse then just mean. But still the idea with the autoencoder hangs in the air
Why do we use 2d 3x3 convolutional filter on the first 3 rgb channels?
Why don’t we have color invariant features, or have to learn them?
Has humans we can recognize a wasp in a picture, no matter in what colors it is shown.
| Wasps | Wasps |
|---|---|
 |
 |
| Fir0002/Flagstaffotos | Fir0002/Flagstaffotos |
CNN is not rotation invariant, how to deal with it?
Train it with slight rotation, then get features for 0, 90, 180, 270 degrees