Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where t=0 and t=1. Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gaussian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at t=1 is conditionally removable while it at t=0 is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores.
Structure information is critical for understanding the semantics of text-rich images, such as documents, tables, and charts. Existing Multimodal Large Language Models (MLLMs) for Visual Document Understanding are equipped with text recognition ability but lack general structure understanding abilities for text-rich document images. In this work, we emphasize the importance of structure information in Visual Document Understanding and propose the Unified Structure Learning to boost the performance of MLLMs. Our Unified Structure Learning comprises structure-aware parsing tasks and multi-grained text localization tasks across 5 domains: document, webpage, table, chart, and natural image. To better encode structure information, we design a simple and effective vision-to-text module H-Reducer, which can not only maintain the layout information but also reduce the length of visual features by merging horizontal adjacent patches through convolution, enabling the LLM to understand high-resolution images more efficiently. Furthermore, by constructing structure-aware text sequences and multi-grained pairs of texts and bounding boxes for publicly available text-rich images, we build a comprehensive training set DocStruct4M to support structure learning. Finally, we construct a small but high-quality reasoning tuning dataset DocReason25K to trigger the detailed explanation ability in the document domain. Our model DocOwl 1.5 achieves state-of-the-art performance on 10 visual document understanding benchmarks, improving the SOTA performance of MLLMs with a 7B LLM by more than 10 points in 5/10 benchmarks. Our codes, models, and datasets are publicly available at https://github.com/X-PLUG/mPLUG-DocOwl/tree/main/DocOwl1.5.
In the realm of computational knowledge representation, Knowledge Graph Reasoning (KG-R) stands at the forefront of facilitating sophisticated inferential capabilities across multifarious domains. The quintessence of this research elucidates the employment of reinforcement learning (RL) strategies, notably the REINFORCE algorithm, to navigate the intricacies inherent in multi-hop KG-R. This investigation critically addresses the prevalent challenges introduced by the inherent incompleteness of Knowledge Graphs (KGs), which frequently results in erroneous inferential outcomes, manifesting as both false negatives and misleading positives. By partitioning the Unified Medical Language System (UMLS) benchmark dataset into rich and sparse subsets, we investigate the efficacy of pre-trained BERT embeddings and Prompt Learning methodologies to refine the reward shaping process. This approach not only enhances the precision of multi-hop KG-R but also sets a new precedent for future research in the field, aiming to improve the robustness and accuracy of knowledge inference within complex KG frameworks. Our work contributes a novel perspective to the discourse on KG reasoning, offering a methodological advancement that aligns with the academic rigor and scholarly aspirations of the Natural journal, promising to invigorate further advancements in the realm of computational knowledge representation.
Mathematical capabilities were previously believed to emerge in common language models only at a very large scale or require extensive math-related pre-training. This paper shows that the LLaMA-2 7B model with common pre-training already exhibits strong mathematical abilities, as evidenced by its impressive accuracy of 97.7% and 72.0% on the GSM8K and MATH benchmarks, respectively, when selecting the best response from 256 random generations. The primary issue with the current base model is the difficulty in consistently eliciting its inherent mathematical capabilities. Notably, the accuracy for the first answer drops to 49.5% and 7.9% on the GSM8K and MATH benchmarks, respectively. We find that simply scaling up the SFT data can significantly enhance the reliability of generating correct answers. However, the potential for extensive scaling is constrained by the scarcity of publicly available math questions. To overcome this limitation, we employ synthetic data, which proves to be nearly as effective as real data and shows no clear saturation when scaled up to approximately one million samples. This straightforward approach achieves an accuracy of 82.6% on GSM8K and 40.6% on MATH using LLaMA-2 7B models, surpassing previous models by 14.2% and 20.8%, respectively. We also provide insights into scaling behaviors across different reasoning complexities and error types.
As medical demands grow and machine learning technology advances, AI-based diagnostic and treatment systems are garnering increasing attention. Medication recommendation aims to integrate patients' long-term health records with medical knowledge, recommending accuracy and safe medication combinations for specific conditions. However, most existing researches treat medication recommendation systems merely as variants of traditional recommendation systems, overlooking the heterogeneity between medications and diseases. To address this challenge, we propose DGMed, a framework for medication recommendation. DGMed utilizes causal inference to uncover the connections among medical entities and presents an innovative feature alignment method to tackle heterogeneity issues. Specifically, this study first applies causal inference to analyze the quantified therapeutic effects of medications on specific diseases from historical records, uncovering potential links between medical entities. Subsequently, we integrate molecular-level knowledge, aligning the embeddings of medications and diseases within the molecular space to effectively tackle their heterogeneity. Ultimately, based on relationships at the entity level, we adaptively adjust the recommendation probabilities of medication and recommend medication combinations according to the patient's current health condition. Experimental results on a real-world dataset show that our method surpasses existing state-of-the-art baselines in four evaluation metrics, demonstrating superior performance in both accuracy and safety aspects. Compared to the sub-optimal model, our approach improved accuracy by 4.40%, reduced the risk of side effects by 6.14%, and increased time efficiency by 47.15%.
As Large Language Models make a breakthrough in natural language processing tasks (NLP), multimodal technique becomes extremely popular. However, it has been shown that multimodal NLP are vulnerable to adversarial attacks, where the outputs of a model can be dramatically changed by a perturbation to the input. While several defense techniques have been proposed both in computer vision and NLP models, the multimodal robustness of models have not been fully explored. In this paper, we study the adversarial robustness provided by modifying loss function of pre-trained multimodal models, by restricting top K softmax outputs. Based on the evaluation and scoring, our experiments show that after a fine-tuning, adversarial robustness of pre-trained models can be significantly improved, against popular attacks. Further research should be studying, such as output diversity, generalization and the robustness-performance trade-off of this kind of loss functions. Our code will be available after this paper is accepted
Bayesian neural networks offer better estimates of model uncertainty compared to frequentist networks. However, inference involving Bayesian models requires multiple instantiations or sampling of the network parameters, requiring significant computational resources. Compared to traditional deep learning networks, spiking neural networks (SNNs) have the potential to reduce computational area and power, thanks to their event-driven and spike-based computational framework. Most works in literature either address frequentist SNN models or non-spiking Bayesian neural networks. In this work, we demonstrate an optimization framework for developing and implementing efficient Bayesian SNNs in hardware by additionally restricting network weights to be binary-valued to further decrease power and area consumption. We demonstrate accuracies comparable to Bayesian binary networks with full-precision Bernoulli parameters, while requiring up to $25\times$ less spikes than equivalent binary SNN implementations. We show the feasibility of the design by mapping it onto Zynq-7000, a lightweight SoC, and achieve a $6.5 \times$ improvement in GOPS/DSP while utilizing up to 30 times less power compared to the state-of-the-art.
Visual sensing of environmental geometry allows robots to use artificial potential fields to avoid sparse obstacles. Yet robots must further traverse cluttered large obstacles for applications like search and rescue through rubble and planetary exploration across Martain rocks. Recent studies discovered that to traverse cluttered large obstacles, multi-legged insects and insect-inspired robots make strenuous transitions across locomotor modes with major changes in body orientation. When viewed on a potential energy landscape resulting from locomotor-obstacle physical interaction, these are barrier-crossing transitions across landscape basins. This potential energy landscape approach may provide a modeling framework for cluttered large obstacle traversal. Here, we take the next step toward this vision by testing whether force sensing allows the reconstruction of the potential energy landscape. We developed a cockroach-inspired, minimalistic robot capable of sensing obstacle contact forces and torques around its body as it propelled forward against a pair of cluttered grass-like beam obstacles. We performed measurements over many traverses with systematically varied body orientations. Despite the forces and torques not being fully conservative, they well-matched the potential energy landscape gradients and the landscape reconstructed from them well-matched ground truth. In addition, inspired by cockroach observations, we found that robot head oscillation during traversal further improved the accuracies of force sensing and landscape reconstruction. We still need to study how to reconstruct landscape during a single traverse, as in applications, robots have little chance to use multiple traverses to sample the environment systematically and how to find landscape saddles for least-effort transitions to traverse.
Ensemble methods exploit the availability of a given number of classifiers or detectors trained in single or multiple source domains and tasks to address machine learning problems such as domain adaptation or multi-source transfer learning. Existing research measures the domain distance between the sources and the target dataset, trains multiple networks on the same data with different samples per class, or combines predictions from models trained under varied hyperparameters and settings. Their solutions enhanced the performance on small or tail categories but hurt the rest. To this end, we propose a modified consensus focus for semi-supervised and long-tailed object detection. We introduce a voting system based on source confidence that spots the contribution of each model in a consensus, lets the user choose the relevance of each class in the target label space so that it relaxes minority bounding boxes suppression, and combines multiple models' results without discarding the poisonous networks. Our tests on synthetic driving datasets retrieved higher confidence and more accurate bounding boxes than the NMS, soft-NMS, and WBF.