Ideas for next projects

[bnord]

@voetberg @samueldmcdermott

1. focus on applications that expose interpretability strengths -- e.g., SZ galaxy clusters
2. perform more experiments on benchmark data sets

[bnord]

@samueldmcdermott do you have our big list of priority next steps -- the one we put together at the retreat?

[samueldmcdermott]

I don't have a photo of the whiteboard sheet -- is that what you were asking about? Is there somewhere else it might be? The only list I maintained otherwise is by using the Notion board with the tag future. Also, in the slack I found this list posted by @bnord which seems to reflect what I recall of our conversations at FNAL:

1. VAEs: yes and medium priority
2. FPGA: yes and low priority
3. Graphs: nada
4. UQ: maybe and low low priority
5. Symmetries/Equivariance: yes, and medium/high priority

[samueldmcdermott]

it's from April 13 btw and here's a link to it
https://deepskies.slack.com/archives/C0424RWFRJR/p1681408496181649

[voetberg]

Adding a screenshot of that convo, just in case

[voetberg]

I had already spoken to Sam about it as well, but I do have a draft ldrd in progress to explore a different direction entirely; something Shuhbendu and I had originally thought when Sam brought the project to us originally. Explaination from the ldrd sheet:

Exploit the decompositional nature of a multi-resolution analysis using wavelets to construct a convolutional filter constrained to the 3-directional nature of a Daubechies-N wavelet decomposition. 
Train the individual operations instead of using a canonical decomposition transform. 
Exploit the built-in symmetries to decrease computational requirements on known symmetrical problems (good for HEP? Def good for astro. Probably)

[samueldmcdermott]

from an email conversation, Brian wrote:
There are six students at various stages of skill who could potentially take another step with this project. Avatars of skill:

1. No deep learning experience, but some physics experience
2. Some moderate GNN experience
3. significant deep learning no physics experience

[samueldmcdermott]

GNNs – do a graph NN on the wavelet decomposition (rather than parallel MLPs), possibly suitable for a student with GNN experience (or maybe we bring in a different GNN expert)

FPGAs – suitable for some specific buddies (Ben Hawks at FNAL, Dylan Rankin at Penn)

VAEs – seems suitable for a student

CNNs – Maggie has an idea here (to be clarified) potentially related to interpretability

Symmetries (concrete) – train a network on a symmetrical data set, split data to break the symmetry, does it perform as well after this split?

Interpretability – potentially related to CNNs

Symmetries (abstract) – Shubhendu?

Write fast code for the decomposition (jax, torch, something else?)

[bnord]

Some more ideas

1. conditional VAEs, conditioned on the type of Wavelet or some wavelet parameters
2. VAEs could better de-noise

[bnord]

Interpretability:
https://arxiv.org/abs/2211.00024

[bnord]

https://journals.aps.org/prd/abstract/10.1103/PhysRevD.105.103533

[bnord]

REGRESSION with wavpool?

[bnord]

Wavpool as an embedding network -- e.g., for SBI