Research

Three themes, one method.

A clinician-defensible model is not one technique — it’s a stack: calibration, interpretability, pediatric-specific data, and external validation. Each theme contributes one layer.

LRFormer: Lipschitz-bounded calibrated uncertaintyA diagram showing a mean prediction curve μ(x) with a ±σ confidence envelope that is narrow over in-distribution inputs and widens for out-of-distribution inputs, enclosed by a straight Lipschitz cone. Illustrative; not a measured result.LRFORMER · UAI 2023LIPSCHITZ-BOUNDED · SINGLE FORWARD PASSσx →IN-DISTRIBUTIONOUT-OF-DISTRIBUTIONLIPSCHITZ BOUNDμ(x)± σ(x)ILLUSTRATIVE · CONFIDENCE WIDENS WITH DISTRIBUTIONAL SHIFT
Theme 01 · Foundations

Responsible AI for medicine.

We design transformer architectures that are calibrated, robust, and Lipschitz-bounded — so a model's confidence has clinical meaning. LRFormer (UAI 2023) provides a theoretical guarantee for uncertainty estimation in single-forward-pass vision models, addressing the overconfidence that has kept transformers off the wards.

ARCH · LRFORMER
Wenqian Ye et al.
ViTASD: patch-wise pediatric face attentionA diagram showing a generic face silhouette divided into a 4x4 patch grid; a few patches are highlighted to indicate attention concentration; an attention plus Gaussian-process head produces a calibrated P(ASD) ± σ density on the right. Illustrative; not a real prediction.VITASD · ICASSP 2023VIT BACKBONE · CALIBRATED PEDIATRIC ASD HEADFACE INPUT · PATCH 16×16ATTN+ GPP(ASD)μ ± σILLUSTRATIVE · ATTENTION-WEIGHTED PATCHES → CALIBRATED HEAD
Theme 02 · Applied

AI for Autism Spectrum Disorder.

ViTASD (ICASSP 2023) introduced a robust ViT baseline for pediatric ASD facial diagnosis — distilling knowledge from large facial-expression datasets and adding a Gaussian-process decoder for clinical robustness.

BENCHMARK · VITASD
Xu Cao et al.
AggPose: multi-scale feature aggregation for infant poseA diagram showing four stacked feature maps at decreasing resolutions on the left, lines converging into an aggregation node in the center, and a generic 15-keypoint infant skeleton on the right. The pose graph mirrors the homepage hero artifact. Illustrative; not a measured result.AGGPOSE · IJCAI 2022MULTI-SCALE · 15-LANDMARK INFANT POSE1×½¼⅛MULTI-SCALE FEATURESAGGPOSE · 15 LANDMARKSILLUSTRATIVE · MULTI-SCALE FEATURES → UNIFIED POSE OUTPUT
Theme 03 · Open

Open pediatric AI.

We build and release open tools where pediatric medicine has been overlooked. AggPose (IJCAI 2022) is a large-scale infant pose-estimation dataset and a high-resolution transformer model intended to support pediatric research on early neurodevelopmental assessment. Datasets, weights, and training recipes are released under permissive licenses.

DATASET · AGGPOSE
Xu Cao et al.
— Selected Works

Recent publications & preprints.

2024 · WACV

A Survey on Multimodal Large Language Models for Autonomous Driving

Cui C, Ma Y, Cao X, Ye W, …, Cao J, Wang Z, Zheng C.

(opens in new tab)
2024 · WACV

MACP: Efficient Model Adaptation for Cooperative Perception

Ma Y, Lu J, Cui C, Zhao S, Cao X, Ye W, Wang Z.

(opens in new tab)
2023 · UAI

Mitigating Transformer Overconfidence via Lipschitz Regularization

Ye W, Ma Y, Cao X, Tang K.

(opens in new tab)
2023 · ICASSP

ViTASD: Robust Vision Transformer Baselines for Autism Spectrum Disorder Facial Diagnosis

Cao X, Ye W, Sizikova E, Bai X, Coffee M, Zeng H, Cao J.

(opens in new tab)
2022 · IJCAI

AggPose: Deep aggregation vision transformer for infant pose estimation

Cao X, Li X, Ma L, Huang Y, Feng X, Chen Z, Zeng H, Cao J.

(opens in new tab)
Preprint

PIE: Simulating Disease Progression via Progressive Image Editing

Liang K, Cao X, Liao KD, Gao T, Ye W, Chen Z, Cao J, Nama T, Sun J.

(opens in new tab)
All publications