Arrow Research search

Author name cluster

Yifeng Yu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

3 papers
2 author rows

Possible papers

3

NeurIPS Conference 2025 Conference Paper

Advancing Wasserstein Convergence Analysis of Score-Based Models: Insights from Discretization and Second-Order Acceleration

  • Yifeng Yu
  • Lu Yu

Score-based diffusion models have emerged as powerful tools in generative modeling, yet their theoretical foundations remain underexplored. In this work, we focus on the Wasserstein convergence analysis of score-based diffusion models. Specifically, we investigate the impact of various discretization schemes, including Euler discretization, exponential integrators, and midpoint randomization methods. Our analysis provides the first quantitative comparison of these discrete approximations, emphasizing their influence on convergence behavior. Furthermore, we explore scenarios where Hessian information is available and propose an accelerated sampler based on the local linearization method. We establish the first Wasserstein convergence analysis for such a Hessian-based method, showing that it achieves an improved convergence rate of order $\widetilde{\mathcal{O}}\left(\frac{\sqrt{d}}{\varepsilon}\right)$, which significantly outperforms the standard rate $\widetilde{\mathcal{O}}\left(\frac{d}{\varepsilon^2}\right)$ of vanilla diffusion models. Numerical experiments on synthetic data and the MNIST dataset validate our theoretical insights.

PDF Details

ICLR Conference 2025 Conference Paper

Global Well-posedness and Convergence Analysis of Score-based Generative Models via Sharp Lipschitz Estimates

  • Connor Mooney
  • Zhongjian Wang
  • Jack Xin
  • Yifeng Yu

We establish global well-posedness and convergence of the score-based generative models (SGM) under minimal general assumptions of initial data for score estimation. For the smooth case, we start from a Lipschitz bound of the score function with optimal time length. The optimality is validated by an example whose Lipschitz constant of scores is bounded at initial but blows up in finite time. This necessitates the separation of time scales in conventional bounds for non-log-concave distributions. In contrast, our follow up analysis only relies on a local Lipschitz condition and is valid globally in time. This leads to the convergence of numerical scheme without time separation. For the non-smooth case, we show that the optimal Lipschitz bound is $O(1/t)$ in the point-wise sense for distributions supported on a compact, smooth and low-dimensional manifold with boundary.

Details

TAAS Journal 2025 Journal Article

TT-DSC: Enhancing YOLO for Marine Ecosystem through Efficient Tensor Train-based Depthwise Separable Deep Neural Network

  • Yunduan Lou
  • Pu Sun
  • Yifeng Yu
  • Shangping Ren
  • Yu Bai

The current era of Artificial Intelligence (AI) has witnessed significant and continuous advancements based on the powerful learning capabilities of Deep Neural Networks (DNNs), particularly those featuring convolutional (CONV) layers. In the field of marine ecosystem conservation, these advancements have revolutionized our ability to monitor and protect ocean environments. DNNs, especially those utilizing YOLO (You Only Look Once) architecture, have been instrumental in tasks such as real-time marine species identification, tracking of marine mammal migrations, detection of coral bleaching events, and monitoring of illegal fishing activities. These AI-powered tools provide unprecedented insights into marine ecosystems, enabling more timely and effective conservation actions. As we aim further to enhance the computational and storage efficiency of these networks, Tensor Train (TT) decomposition has emerged as a notable compression technique due to its high compression ratio and ability to maintain strong performance. However, the CONV layer in TT format still incurs substantial computational costs, stemming from convolution calculations and the additional multiplication operations intrinsic to TT usage. Consequently, reducing these computational costs is critical to improving the effectiveness of DNNs. To advance the computational efficiency of DNNs, this paper introduces a novel separable TT decomposition that offers an efficient TT-format CONV layer using depthwise separable convolution. Remarkably, this method not only reduces computation costs significantly but also maintains a similar capacity for parameter compression and accuracy compared to the conventional TT-format model. Furthermore, our method facilitates distributed learning based on the factorization of CONV layers. By scheduling the smaller-factored weight tensors, we significantly mitigate the GPU memory requirements of the larger model, thereby enhancing the availability and speed of training. The primary contributions of this paper are twofold: (1) a simultaneous reduction in computational cost and parameter count in TT-based CONV layers, achieved by minimizing TT redundancy and optimizing convolution, leading to up to 7–10× improvements per layer, an overall one-third reduction in parameters, and 15% reduction in FLOPs at the model level. 2) We demonstrate how our approach enables effective distributed learning and resource allocation. By merging TT decomposition and depthwise separable convolution, we present TTDSC, a TT-based depthwise separable convolution approach. This study opens new avenues to improve the efficiency of CONV. layers compression and has significant implications for large-scale deep learning applications.

Details DOI
v2026.09.13