NVIDIA, Duke University
Abstract:Although masked image generation models and masked diffusion models are designed with different motivations and objectives, we observe that they can be unified within a single framework. Building upon this insight, we carefully explore the design space of training and sampling, identifying key factors that contribute to both performance and efficiency. Based on the improvements observed during this exploration, we develop our model, referred to as eMIGM. Empirically, eMIGM demonstrates strong performance on ImageNet generation, as measured by Fr\'echet Inception Distance (FID). In particular, on ImageNet 256x256, with similar number of function evaluations (NFEs) and model parameters, eMIGM outperforms the seminal VAR. Moreover, as NFE and model parameters increase, eMIGM achieves performance comparable to the state-of-the-art continuous diffusion models while requiring less than 40% of the NFE. Additionally, on ImageNet 512x512, with only about 60% of the NFE, eMIGM outperforms the state-of-the-art continuous diffusion models.
Abstract:In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled B\v{a}il\'ing in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.
Abstract:While LiDAR and cameras are becoming ubiquitous for unmanned aerial vehicles (UAVs) but can be ineffective in challenging environments, 4D millimeter-wave (MMW) radars that can provide robust 3D ranging and Doppler velocity measurements are less exploited for aerial navigation. In this paper, we develop an efficient and robust error-state Kalman filter (ESKF)-based radar-inertial navigation for UAVs. The key idea of the proposed approach is the point-to-distribution radar scan matching to provide motion constraints with proper uncertainty qualification, which are used to update the navigation states in a tightly coupled manner, along with the Doppler velocity measurements. Moreover, we propose a robust keyframe-based matching scheme against the prior map (if available) to bound the accumulated navigation errors and thus provide a radar-based global localization solution with high accuracy. Extensive real-world experimental validations have demonstrated that the proposed radar-aided inertial navigation outperforms state-of-the-art methods in both accuracy and robustness.
Abstract:Multimodal recommendation systems can learn users' preferences from existing user-item interactions as well as the semantics of multimodal data associated with items. Many existing methods model this through a multimodal user-item graph, approaching multimodal recommendation as a graph learning task. Graph Neural Networks (GNNs) have shown promising performance in this domain. Prior research has capitalized on GNNs' capability to capture neighborhood information within certain receptive fields (typically denoted by the number of hops, $K$) to enrich user and item semantics. We observe that the optimal receptive fields for GNNs can vary across different modalities. In this paper, we propose GNNs with Modality-Independent Receptive Fields, which employ separate GNNs with independent receptive fields for different modalities to enhance performance. Our results indicate that the optimal $K$ for certain modalities on specific datasets can be as low as 1 or 2, which may restrict the GNNs' capacity to capture global information. To address this, we introduce a Sampling-based Global Transformer, which utilizes uniform global sampling to effectively integrate global information for GNNs. We conduct comprehensive experiments that demonstrate the superiority of our approach over existing methods. Our code is publicly available at https://github.com/CrawlScript/MIG-GT.
Abstract:Recently, several studies have combined Gaussian Splatting to obtain scene representations with language embeddings for open-vocabulary 3D scene understanding. While these methods perform well, they essentially require very dense multi-view inputs, limiting their applicability in real-world scenarios. In this work, we propose SparseLGS to address the challenge of 3D scene understanding with pose-free and sparse view input images. Our method leverages a learning-based dense stereo model to handle pose-free and sparse inputs, and a three-step region matching approach to address the multi-view semantic inconsistency problem, which is especially important for sparse inputs. Different from directly learning high-dimensional CLIP features, we extract low-dimensional information and build bijections to avoid excessive learning and storage costs. We introduce a reconstruction loss during semantic training to improve Gaussian positions and shapes. To the best of our knowledge, we are the first to address the 3D semantic field problem with sparse pose-free inputs. Experimental results show that SparseLGS achieves comparable quality when reconstructing semantic fields with fewer inputs (3-4 views) compared to previous SOTA methods with dense input. Besides, when using the same sparse input, SparseLGS leads significantly in quality and heavily improves the computation speed (5$\times$ speedup). Project page: {\tt\small \url{https://ustc3dv.github.io/SparseLGS}}
Abstract:Multi-label data stream usually contains noisy labels in the real-world applications, namely occuring in both relevant and irrelevant labels. However, existing online multi-label classification methods are mostly limited in terms of label quality and fail to deal with the case of noisy labels. On the other hand, the ground-truth label distribution may vary with the time changing, which is hidden in the observed noisy label distribution and difficult to track, posing a major challenge for concept drift adaptation. Motivated by this, we propose an online multi-label classification algorithm under Noisy and Changing Label Distribution (NCLD). The convex objective is designed to simultaneously model the label scoring and the label ranking for high accuracy, whose robustness to NCLD benefits from three novel works: 1) The local feature graph is used to reconstruct the label scores jointly with the observed labels, and an unbiased ranking loss is derived and applied to learn reliable ranking information. 2) By detecting the difference between two adjacent chunks with the unbiased label cardinality, we identify the change in the ground-truth label distribution and reset the ranking or all information learned from the past to match the new distribution. 3) Efficient and accurate updating is achieved based on the updating rule derived from the closed-form optimal model solution. Finally, empirical experimental results validate the effectiveness of our method in classifying instances under NCLD.
Abstract:We propose TG-LMM (Text-Guided Large Multi-Modal Model), a novel approach that leverages textual descriptions of organs to enhance segmentation accuracy in medical images. Existing medical image segmentation methods face several challenges: current medical automatic segmentation models do not effectively utilize prior knowledge, such as descriptions of organ locations; previous text-visual models focus on identifying the target rather than improving the segmentation accuracy; prior models attempt to use prior knowledge to enhance accuracy but do not incorporate pre-trained models. To address these issues, TG-LMM integrates prior knowledge, specifically expert descriptions of the spatial locations of organs, into the segmentation process. Our model utilizes pre-trained image and text encoders to reduce the number of training parameters and accelerate the training process. Additionally, we designed a comprehensive image-text information fusion structure to ensure thorough integration of the two modalities of data. We evaluated TG-LMM on three authoritative medical image datasets, encompassing the segmentation of various parts of the human body. Our method demonstrated superior performance compared to existing approaches, such as MedSAM, SAM and nnUnet.
Abstract:With the rapid development of the short video industry, traditional e-commerce has encountered a new paradigm, video-driven e-commerce, which leverages attractive videos for product showcases and provides both video and item services for users. Benefitting from the dynamic and visualized introduction of items,video-driven e-commerce has shown huge potential in stimulating consumer confidence and promoting sales. In this paper, we focus on the video retrieval task, facing the following challenges: (1) Howto handle the heterogeneities among users, items, and videos? (2)How to mine the complementarity between items and videos for better user understanding? In this paper, we first leverage the dual graph to model the co-existing of user-video and user-item interactions in video-driven e-commerce and innovatively reduce user preference understanding to a graph matching problem. To solve it, we further propose a novel bi-level Graph Matching Network(GMN), which mainly consists of node- and preference-level graph matching. Given a user, node-level graph matching aims to match videos and items, while preference-level graph matching aims to match multiple user preferences extracted from both videos and items. Then the proposed GMN can generate and improve user embedding by aggregating matched nodes or preferences from the dual graph in a bi-level manner. Comprehensive experiments show the superiority of the proposed GMN with significant improvements over state-of-the-art approaches (e.g., AUC+1.9% and CTR+7.15%). We have developed it on a well-known video-driven e-commerce platform, serving hundreds of millions of users every day
Abstract:Motor imagery electroencephalogram (EEG)-based brain-computer interfaces (BCIs) offer significant advantages for individuals with restricted limb mobility. However, challenges such as low signal-to-noise ratio and limited spatial resolution impede accurate feature extraction from EEG signals, thereby affecting the classification accuracy of different actions. To address these challenges, this study proposes an end-to-end dual-branch network (EEG-DBNet) that decodes the temporal and spectral sequences of EEG signals in parallel through two distinct network branches. Each branch comprises a local convolutional block and a global convolutional block. The local convolutional block transforms the source signal from the temporal-spatial domain to the temporal-spectral domain. By varying the number of filters and convolution kernel sizes, the local convolutional blocks in different branches adjust the length of their respective dimension sequences. Different types of pooling layers are then employed to emphasize the features of various dimension sequences, setting the stage for subsequent global feature extraction. The global convolution block splits and reconstructs the feature of the signal sequence processed by the local convolution block in the same branch and further extracts features through the dilated causal convolutional neural networks. Finally, the outputs from the two branches are concatenated, and signal classification is completed via a fully connected layer. Our proposed method achieves classification accuracies of 85.84% and 91.60% on the BCI Competition 4-2a and BCI Competition 4-2b datasets, respectively, surpassing existing state-of-the-art models. The source code is available at https://github.com/xicheng105/EEG-DBNet.
Abstract:The delineation of tumor target and organs-at-risk is critical in the radiotherapy treatment planning. Automatic segmentation can be used to reduce the physician workload and improve the consistency. However, the quality assurance of the automatic segmentation is still an unmet need in clinical practice. The patient data used in our study was a standardized dataset from AAPM Thoracic Auto-Segmentation Challenge. The OARs included were left and right lungs, heart, esophagus, and spinal cord. Two groups of OARs were generated, the benchmark dataset manually contoured by experienced physicians and the test dataset automatically created using a software AccuContour. A resnet-152 network was performed as feature extractor, and one-class support vector classifier was used to determine the high or low quality. We evaluate the model performance with balanced accuracy, F-score, sensitivity, specificity and the area under the receiving operator characteristic curve. We randomly generated contour errors to assess the generalization of our method, explored the detection limit, and evaluated the correlations between detection limit and various metrics such as volume, Dice similarity coefficient, Hausdorff distance, and mean surface distance. The proposed one-class classifier outperformed in metrics such as balanced accuracy, AUC, and others. The proposed method showed significant improvement over binary classifiers in handling various types of errors. Our proposed model, which introduces residual network and attention mechanism in the one-class classification framework, was able to detect the various types of OAR contour errors with high accuracy. The proposed method can significantly reduce the burden of physician review for contour delineation.