Yidong Ouyang

Hey, I am Yidong Ouyang, a Ph.D. student at UCLA where I am fortunate to be advised by Guang Cheng. Prior to this, I earned my MPhil from CUHK Shenzhen, where I was advised by Liyan Xie. My research interests lie in Generative Models, especially diffusion models.

Email  /  CV  /  Github  /  Google Scholar

We prove that the optimal diffusion model for the target domain integrates pre-trained diffusion models on the source domain with additional guidance from a domain classifier. We further extend TGDP to a conditional version for modeling the joint distribution of data and its corresponding labels, together with two additional regularization terms to enhance the model performance.

We first observe that the widely adopted "impute-then-generate" pipeline may lead to a biased learning objective. Then we propose to mask the regression loss of Denoising Score Matching in the training phase. We prove the proposed method is consistent in learning the score of data distributions, and the proposed training objective serves as an upper bound for the negative likelihood in certain cases.

We first analyze the optimality condition of synthetic distribution for achieving non-trivial robust accuracy. We show that enhancing the distinguishability among the generated data is critical for improving adversarial robustness. Thus, we propose the Contrastive-Guided Diffusion Process (Contrastive-DP), which adopts the contrastive loss to guide the diffusion model in data generation.

We theoretically prove the self-attention model can achieve zero approximation error. Moreover, our proposed model can avoid the classical trade-off between approximation error and sample error in the mean squared error analysis.

We presents the first review for recent advances in domain generalization, including the theory of generalization, the taxonomy of domain generalization methods, and the potential research topics.

We propose a mutual information maximum module to explicitly drop superfluous information related to the domain labels. We thoroughly reviewed our methods from both theoretical and empirical perspective and demonstrated the connections and advantages to domain adversarial training and triplet loss.

We investigated the regularization technique from a Fourier perspective and pinpointed an extreme small but valid spectral range for different layers. our regularization technique reduces the generalization gap on computer vision benchmarks and improves the robustness of models, especially against low frequency attack.

We propose RotateConv kernels as an interpolation-based method that transforms traditional square kernels to line segments. Our approach can massively reduce the number of parameters and calculations while maintaining acceptable performance.

Website credits go to Jon Barron!