IEEE ISBI 2025Oral
A brain-network transformer pre-trained with contrastive learning over randomly altered connectivity graphs, reaching 82.6 AUROC and 74% accuracy on ABIDE autism detection.
SPIE Medical Imaging 2025
Picks harder negatives by image similarity and pairs them with a biased contrastive loss, lifting red blood cell classification accuracy to 0.840 over SimCLR, MoCo, and DINO.
AAAI 2025
Brings the prior of a pretrained inpainting model to large-area watermark removal by gated fusion of watermark-cleaned features, and trains on coarse masks so it holds up when the mask is imprecise.
AAAI 2024
Separates watermark exclusion from background restoration, then fuses content recovered from beneath the watermark with content imagined from the surrounding context (HWVOC PSNR 39.52).