Not my area of expertise and OOD might be a real problem.
Can you simulate this? Agents are good at hill climbing loops toward specific goals, so that would be the first thing to try.
If not, I would start by building simulation harness (if that's possible)
🚨 New publication alert:
📢 ”Learning a Sampling-Free Variational DNN Plugin from Tiny Training Sets to Refine OOD Segmentation With Uncertainty Estimation.”
🖊️ @jimut_pal, S P Awate.
⬇️
On CLIP ViT-B/32, MonoSoup improves OOD accuracy across 70 ImageNet fine-tuned checkpoints while keeping strong ID performance.
On collapsed checkpoints, it recovers up to ~8% average OOD accuracy.
One checkpoint instead of a full soup 🍲
8/10