A5100685776- Stochastic Gradient Optimization Techniques
- Advanced Neural Network Applications
- Multimodal Machine Learning Applications
- 3D Shape Modeling and Analysis
- Image Processing and 3D Reconstruction
- Adversarial Robustness in Machine Learning
- Manufacturing Process and Optimization
- High Temperature Alloys and Creep
- Aluminum Alloys Composites Properties
- Machine Learning and Data Classification
- Advanced Materials Characterization Techniques
- Privacy-Preserving Technologies in Data
- Anomaly Detection Techniques and Applications
- Advanced Image and Video Retrieval Techniques
- Machine Learning and ELM
- Sparse and Compressive Sensing Techniques
- Computational Geometry and Mesh Generation
- Natural Language Processing Techniques
- Advanced Image Processing Techniques
- Machine Learning in Materials Science
- Human Pose and Action Recognition
- Domain Adaptation and Few-Shot Learning
- Computer Graphics and Visualization Techniques
- Gaussian Processes and Bayesian Inference
- Advanced Materials and Mechanics
Dalian University of Technology
2023-2024
University of Chinese Academy of Sciences
2019-2024
Massachusetts Institute of Technology
2023-2024
Institute of Computing Technology
2024
Technical University of Munich
2023
State Key Laboratory of Chemical Engineering
2023
East China University of Science and Technology
2023
University of Southern California
2017-2023
École Polytechnique Fédérale de Lausanne
2023
Harbin Institute of Technology
2023
We propose HOI Transformer to tackle human object interaction (HOI) detection in an end-to-end manner. Current approaches either decouple task into separated stages of and classification or introduce surrogate problem. In contrast, our method, named Transformer, streamlines the pipeline by eliminating need for many hand-designed components. reasons about relations objects humans from global image context directly predicts instances parallel. A quintuple matching loss is introduced force...
This work introduces DiGress, a discrete denoising diffusion model for generating graphs with categorical node and edge attributes. Our utilizes process that progressively edits noise, through the of adding or removing edges changing categories. A graph transformer network is trained to revert this process, simplifying problem distribution learning over into sequence classification tasks. We further improve sample quality by introducing Markovian noise preserves marginal types during...
Generative AI models provide foundational knowledge and sufficient reasoning to aid in individual aspects of a computational design modeling workflow (see Part 1 <https://doi.org/10.1162/99608f92.cc80fe30> ). In this work, we analyze the ability state-of-the-art LLMs (in particular GPT-4) reason about entire end-to-end workflow, from conceptualizing realization, for two target domains: static physical objects (here, furniture), dynamical cyberphysical systems quadcopters). Our investigation...
The progress in generative AI, particularly large language models (LLMs), opens new prospects design and manufacturing. Our research explores the use of these tools throughout entire manufacturing workflow. We assess capabilities LLMs various tasks: converting text prompts into designs, generating spaces variations, transforming designs instructions, evaluating performance, searching for based on performance metrics. identify discuss current strengths limitations LLMs, suggesting areas...
Current multi-modal object detection approaches focus on the vehicle domain and are limited in perception range processing capabilities. Roadside sensor units (RSUs) introduce a new for systems leverage altitude to observe traffic. Cameras LiDARs mounted gantry bridges increase produce full digital twin of In this work, we InfraDet3D, 3D detector roadside infrastructure sensors. We fuse two using early fusion further incorporate detections from monocular cameras robustness detect small...
We introduce a compact, intuitive procedural graph representation for cellular metamaterials, which are small-scale, tileable structures that can be architected to exhibit many useful material properties. Because the structures’ “architectures” vary widely—with elements such as beams, thin shells, and solid bulks—it is difficult explore them using existing representations. Generic approaches like voxel grids versatile, but it cumbersome represent edit individual structures;...
Scene recognition is challenging due to the intra-class diversity and inter-class similarity. Previous works recognize scenes either with global representations or intermediate of objects. In contrast, we investigate more discriminative image object-to-object relations for scene recognition, which are based on triplets <;object, relation, object> obtained detection techniques. Particularly, two types representations, including co-occurring frequency relation (denoted as COOR) sequential...
The advancement of Large Language Models (LLMs), including GPT-4, provides exciting new opportunities for generative design. We investigate the application this tool through sequential steps computational design and manufacturing workflow. In particular, we examine how LLMs can aid in tasks including: converting a text-based prompt into quantitative specification, transforming instructions, producing space variations within that space, computing performance given design, optimizing designs...
Botanical simulation plays an important role in many fields including visual effects, games and virtual reality. Previous plant research has focused on computing physically based motion, under the assumption that material properties are known. It is too tedious impractical to manually set spatially-varying of complex trees. In this paper, we give a method mass density, stiffness damping individual tree components (branches leaves) using small number intuitive parameters. Our rooted...
Understanding the loss surface of a neural network is fundamentally important to understanding deep learning. This paper presents how piecewise linear activation functions substantially shape surfaces networks. We first prove that {\it many networks have infinite spurious local minima} which are defined as minima with higher empirical risks than global minima. Our result demonstrates activations possess substantial differences well-studied holds for any arbitrary depth and (excluding...
The regeneration process of fluid catalytic cracking (FCC) units produces amounts greenhouse gases (GHGs), including carbon dioxide (CO2), nitrous oxide (N2O), and methane (CH4). However, the flue gas GHG emission characteristics FCC are not well understood in China. In our work, three typical taken as stack tests for CO2, N2O, CH4 emissions gas. on-site monitoring results show that regenerated form unit has greatest impact on CH4, CO2 is primary GHG, with total >99%. Meanwhile, spent...
We present a method for modeling solid objects undergoing large spatially varying and/or anisotropic strains, and use it to reconstruct human anatomy from medical images. Our novel shape deformation uses plastic strains the finite element successfully model shapes specified by sparse point constraints on boundary of object. extensively compare our standard second-order methods, variational surface-based demonstrate that avoids spikiness, wiggliness, other artifacts previous methods. how...
We provide a simple convergence proof for AdaGrad optimizing non-convex objectives under only affine noise variance and bounded smoothness assumptions. The is essentially based on novel auxiliary function $\xi$ that helps eliminate the complexity of handling correlation between numerator denominator AdaGrad's update. Leveraging proofs, we are able to obtain tighter results than existing \citep{faw2022power} extend analysis several new important cases. Specifically, over-parameterized regime,...
Deep Neural Networks (DNNs) have obtained impressive performance across tasks, however they still remain as black boxes, e.g., hard to theoretically analyze. At the same time, Polynomial (PNs) emerged an alternative method with a promising and improved interpretability but yet reach of powerful DNN baselines. In this work, we aim close gap. We introduce class PNs, which are able ResNet range six benchmarks. demonstrate that strong regularization is critical conduct extensive study exact...
This paper studies the relationship between generalization and privacy preservation in iterative learning algorithms by two sequential steps. We first establish an alignment for any algorithm. prove that $(\varepsilon, \delta)$-differential implies on-average bound multi-database which further leads to a high-probability also PAC-learnable guarantee differentially private algorithms. then investigate how nature shared most influence generalization. Three composition theorems are proposed...
Due to the abstraction of scenes, comprehensive scene understanding requires semantic modeling in both global and local aspects. Scene recognition is usually researched from a point view, while dense captioning typically studied for regions. Previous works separately research on captioning. In contrast, we propose joint learning framework that benefits mutual coupling models. Generally, these two tasks are coupled through steps, 1) fusing supervision by considering contexts between labels...
Adaptive Moment Estimation (Adam) optimizer is widely used in deep learning tasks because of its fast convergence properties. However, the Adam still not well understood. In particular, existing analysis cannot clearly demonstrate advantage over SGD. We attribute this theoretical embarrassment to $L$-smooth condition (i.e., assuming gradient globally Lipschitz continuous with constant $L$) adopted by literature, which has been pointed out often fail practical neural networks. To tackle...