03-06-2019: Meeting w/ Camille, Carole

gkiar / analytical-stability

Notes, presentations, plans, misc code, and links relating to the Analytical Stability project with Camille Maumet and Elisa Fromont in INRIA.

GNU General Public License v3.0

0 stars 1 forks source link

03-06-2019: Meeting w/ Camille, Carole #6

Open gkiar opened 5 years ago

gkiar commented 5 years ago

topics:

a brief overview of our project (a tiny bit outdated, but summarized here)

describing the model architecture we've adopted (brief paper summary here)

overview of the current performance (notebook here, in particular the first set of figures)

asking for recommendations on how to improve our performance

gkiar commented 5 years ago

Notes from Antoine (pre meeting):

[ ] check performance against test and validation sets
[ ] possibly add skip connections to increase spatial context
[ ] decrease learning rate to take more advantage of the high number of samples/epochs

cmaumet commented 5 years ago

Questions/things to look at (from me):

Can we avoid training in the background (i.e. we know those should be zeros, no need to train the autoencoder in those regions). We could use use the MNI brainmask for that.

cmaumet commented 5 years ago

From Carole: There is a paper referring to square like artefacts in autoencoders. Could be related to max-pooling. Something in the architecture. Zara had the same king of artefacts but when using the latent representation not the reconstructed image.

The problem may come from the upsampling part.

cmaumet commented 5 years ago

Meeting w/ Carole Latizien, @efromont, @cmaumet.

cmaumet commented 5 years ago

gkiar commented 5 years ago

https://distill.pub/2016/deconv-checkerboard/

gkiar commented 5 years ago

https://github.com/Microsoft/CNTK/issues/1741