Hello Naitik,
Thank you for your interest in the project.
I would encourage you to focus on the direction you are most interested
in pursuing, as long as it is aligned with the guidelines. All
interesting and feasible plans and ideas would be considered and could
be combined with our development paths.
Best Regards,
Alex Apostolakis
On 3/16/2026 12:59 PM, Naitik Agrawal wrote:
Hi Alexis, Christos, Simon
I’ve been digging into the TIRAuxCloud codebase and paper following
your last broadcast. I'm focusing my GSoC proposal on the XAI and
uncertainty quantification extensions you requested.
I wanted to ask a quick technical question regarding the "unimodal
dominance" issue you highlighted in Section IV. Since the TIR band is
so closely correlated with the cloud mask, it provides a simpler,
high-signal pathway. It makes complete sense that the network neglects
the more complex multimodal relationships found in the DEM and
meteorological variables.
My proposed architectural solution is auxiliary-attribution
regularization: using DeepSHAP (computed periodically on a validation
subset) to monitor per-channel feature attribution during training. If
auxiliary channels (DEM, meteorological variables) show near-zero
attribution while TIR dominates, a regularization term would penalize
this imbalance effectively forcing the model to extract signal from
the atmospheric context. Combined with input-level modality dropout on
TIR, this creates both a push (dropout forces reliance on auxiliaries)
and a pull (attribution regularization rewards it)
Before I finalize my implementation plan, I wanted to ask: Have you
already experimented with any specific architectural interventions
like the parallel encoder branches operating at lower spatial
frequencies mentioned in your discussion to force the network to
utilize the auxiliary data?
I'd love to know if any of these directions have already been
explored internally, so I can build on your findings rather than
spending time on the same.
Best regards,
Naitik Agrwal
----
Λαμβάνετε αυτό το μήνυμα απο την λίστα: Λίστα αλληλογραφίας και συζητήσεων που απευθύνεται σε φοιτητές developers \& mentors έργων του Google Summer of Code - A discussion list for student developers and mentors of Google Summer of Code projects.,
https://lists.ellak.gr/gsoc-developers/listinfo.html
Μπορείτε να απεγγραφείτε από τη λίστα στέλνοντας κενό μήνυμα ηλ. ταχυδρομείου στη διεύθυνση<gsoc-developers+unsubscribe [ at ] ellak [ dot ] gr>.
----
Λαμβάνετε αυτό το μήνυμα απο την λίστα: Λίστα αλληλογραφίας και συζητήσεων που απευθύνεται σε φοιτητές developers \& mentors έργων του Google Summer of Code - A discussion list for student developers and mentors of Google Summer of Code projects.,
https://lists.ellak.gr/gsoc-developers/listinfo.html
Μπορείτε να απεγγραφείτε από τη λίστα στέλνοντας κενό μήνυμα ηλ. ταχυδρομείου στη διεύθυνση <gsoc-developers+unsubscribe [ at ] ellak [ dot ] gr>.