Densenets supermask

[bhack]

Have you never tried to find supermask over densenets?

[mitchellnw]

This seems like more of a question for

• https://github.com/uber-research/deconstructing-lottery-tickets
• https://github.com/allenai/hidden-networks though I don't believe anyone has tried this! Very cool idea.

[bhack]

I was interested in your specific context :wink: and the comments and FAQ section in https://mitchellnw.github.io/blog/2020/supsup/ was poiting to this repo :smile_cat:

[bhack]

P.s. I got this vague idea reading the conclusions of https://arxiv.org/abs/2006.12156.

If he is wondering about skip connections why not about dense connections?

[mitchellnw]

Oops! Sorry about that :)

We tried skip-connections with resnets here which worked well.

I believe dense-connections have not been explored with supermasks and it seems like a really interesting direction!

[bhack]

Yes I know but I meant in the mentioned work the conclusion was more related to their strong claim that subnetworks "only needs a logarithmic factor (in all variables but depth) number of neurons per weight of the target subnetwork".

So the open question was more about the impact of convolutional and batch norm layers, skip-connections, (densenet like connections?) and LSTMs on the number of required sampled neurons to maintain a good accuracy.

[bhack]

I also meant that this claim could has an interesting impact in your continual learning specific setup. If you can free-up "more resources" it is useful when you need to expand on new task.

[mitchellnw]

Thanks, that could definitely help!

[bhack]

If you are interested in this see also Optimal Lottery Tickets via SubsetSum: Logarithmic Over-Parameterization is Sufficient

[mitchellnw]

Thank you, we have seen this but haven't taken a close look! Hopefully we can soon it seems awesome

[bhack]

Other then densenets another interesting direction are Transformers. Some early exploring efforts were made in:

https://arxiv.org/abs/2005.00561 https://arxiv.org/abs/2005.03454