Negative result
Learned patch ordering (REOrder) for 3D brain MRI: a rigorous negative result
I implemented a published RL method for 3D MRI from scratch, tested it across about 40 runs and 3 architectures, and showed that its apparent gain was measurement noise.
- Question
- REOrder learns the order in which image patches are fed to a transformer. It was published for 2D images. Does it still help when extended to 3D volumetric MRI for Alzheimer's classification on ADNI?
- What I built
- I implemented REOrder in PyTorch: a Plackett–Luce policy over patch permutations, trained with REINFORCE under an EMA baseline and a Gumbel-temperature curriculum, extended from 2D to 3D volumetric MRI.
- How I tested it
- I benchmarked it against nine fixed orderings across three transformers spanning a 268× capacity range (0.33M–88.4M parameters). That came to ~40 experiments on an 8× V100 HPC cluster, with subject-level splits to prevent leakage.
- Result
-
The learned ordering did not beat a well-chosen fixed ordering, and I traced each architecture's null result to a distinct mechanism:
- Full 3D ViT: it's order-sensitive only through positional embeddings, so it can't benefit from a policy that's resampled every batch.
- 0.33M model: undertrained at the available compute budget.
- Hierarchical model: the apparent gain came from block-locality structure, not from any learned permutation.
- The part I'm proudest of
- One configuration showed an apparent +7.5-point gain. A 2×2 ablation and seed-variance analysis isolated it to measurement noise, not a real improvement.