2021.10.07.

[qbeer]

Hi All,

we will collect next time's papers here. :)

The meeting will be online from 16:00, on the same link as always.

Alex

[icsabai]

If there will be enough suggestions from students/colleagues I am not pushing this, but if there is enough time, I will have a detailed presentation of:

Boucaud, A., Heneka, C., Ishida, E.E., Sedaghat, N., de Souza, R.S., Moews, B., Dole, H., Castellano, M., Merlin, E., Roscani, V. and Tramacere, A., 2020. Photometry of high-redshift blended galaxies using deep learning. Monthly Notices of the Royal Astronomical Society, 491(2), pp.2481-2495. arxiv link

This looks like some 'astronomy' paper, but actually very much related to the topic, Oz was talking about today, but now with star spectra instead of cats and dogs. Namely, to find meaningful "physical" interpretation of (unsupervised) autoencoder network's latent representation or as they say “distilling data into knowledge”.

For those, who are not familiar with variational autoencoders, I suggest to read this very informative and didactic TDS post

[qbeer]

For today we will do a quick overview of these methods as well:

You can read the accompanying Facebook AI blog article as well: https://ai.facebook.com/blog/dino-paws-computer-vision-with-self-supervised-transformers-and-10x-more-efficient-training/

DINO PAWS

If time remains we will also check this paper that builds on top of the above articles:

Localizing objects with self-supervised transformers without labels

Will not be discussing this, but is very interesting that ResNets are still extremely powerful and with newer training techniques they can outperform many SOTA models and still be 'in the big league': https://arxiv.org/pdf/2110.00476.pdf

• fun fact: the main author is an American multi-millionaire who developed a model zoo package with fine-tuned SOTA models timm which is extremely useful and does only open-source projects

[qbeer]

We are starting now! :)