New paper: Unsupervised Domain Adaptation for Multitask Image Analysis with Extreme Label Shift

Our latest paper, “Unsupervised Domain Adaptation for Multitask Image Analysis with Extreme Label Shift”, has been accepted in Astronomy & Computing.

The problem

Deep learning models trained on Monte-Carlo simulations tend to fail on real CTAO LST-1 data because of differences between the simulated and real feature distributions. One of the main sources of this gap is the level of Night Sky Background (NSB) noise, which is known to hurt CNN performance (Parsons et al. 2022). Domain adaptation methods can help by shifting the learned distributions using real (unlabeled) data at training time.

In this work we implemented and compared three domain adaptation strategies — DANN (adversarial), DeepCORAL (statistics alignment) and DeepJDOT (optimal transport) — within our multitask framework, each combined with two multitask-balancing strategies: Uncertainty Weighting (UW) and GradNorm (GN).

However, gamma-ray data also suffer from extreme label imbalance: fewer than 1 gamma ray per 10,000 protons in real data. This label shift can push standard domain adaptation methods toward the wrong distribution.

Our solution

We propose conditional domain adaptation methods (CDANN, CDeepCORAL, CDeepJDOT) coupled with a Gradient Layer for progressive ramp-up, and show that CDANN + Uncertainty Weighting recovers the performance lost to both the domain shift and the label shift.

The problem

Our solution

Links