Zhiheng Li

Pronunciation (🔊): Zhiheng: /ʤihɛnŋ/, Li: /li/

I am an Applied Scientist at Amazon AGI, working on multi-modal foundation models. I received my PhD degree in computer science at University of Rochester in 2023 (PhD advisor: Prof. Chenliang Xu) and my bachelor's degree at Wuhan University in 2018.

My research interests are computer vision and responsible AI (e.g., fairness, robustness, and interpretability). I also conducted research in multi-modal generative AI, such as text-to-image synthesis and talking face video generation.

Email: zhiheng.li -at- ieee.org

CV  /  Google Scholar  /  OpenReview  /  LinkedIn  /  Github

We propose FairRAG to mitigate demographic biases of diffusion models (e.g., Stable Diffusion) based on Retrieval Augmented Generation (RAG).

We proposed a new method to discover multiple biased subgroups based on Partial Least Squares (PLS), which enables dimension reduction guided by useful supervisions from the image classifier.

We introduce two new datasets (UrbanCars and ImageNet-W) to study multi-shortcut learning. Especially, ImageNet-W is created based on the newly found watermark shortcut in ImageNet affecting a broad range of vision models, including ResNet, RegNet, ViT, MoCov3, MAE, SEER, SWAG, model soups, and CLIP. Our work surfaces an overlooked challenge in shortcut learning: multi-shortcut mitigation resembles a Whac-A-Mole game, i.e., mitigating one shortcut amplifies others.

We introduce Debiasing Alternate Networks (DebiAN) to discover and mitigate unknown biases of an image classifier. DebiAN trains two networks in an alternate fashion. The discoverer network identifies unknown biases in the classifier. The classifier mitigates biases found by the discoverer.

Unlike existing approaches (e.g., GroupDRO) focusing on reweighting or rebalancing training data, we show that simply using embeddings from a large pretrained vision model extractor (e.g., SWAG) and training a linear classifier on top of it without training group information achieve state-of-the-art results in combating group shift on Waterbirds.

We introduce a new framework, StyleT2I, to improve the compositionality of text-to-image synthesis, e.g., faithfully synthesizing a face image in “a man wearing lipstick” that is underrepresented in the training data (e.g., due to societal stereotypes).

We study a new problem: discover the unknown bias (i.e., the bias that is out of human's conjecture) of an image classifier. We tackle it by optimizing a hyperplane in generative model's latent space. The semantic meaning of the bias can be interpreted from the variation in the synthesized traversal images based on the optimized latent hyperplane.

We propose DGM, which incorporates the traditional perceptual grouping process into modern CNN architecture for better contextual modeling, interpretability, and a lower computational overhead.

A novel method to generate and evolve the pseudo-annotations for weakly-supervised video actor-action segmentation task.

We train a GNN to predict coarse-grained (CG) molecule, i.e., graph partitioning on the given molecule. We also collect HAM dataset, which provides CG mapping annotations of molecules.

Given an audio speech and a lip image of an arbitrary target identity, synthesize lip movements of the target identity saying the speech.

The template is based on Jon Barron's website.

The Whack-A-Mole icon created by Flat Icons - Flaticon