Yu Zhang

Longhui Yu, Weisen Jiang, Han Shi et al. · cambridge

Large language models (LLMs) have pushed the limits of natural language understanding and exhibited excellent problem-solving ability. Despite the great success, most existing open-source LLMs (e.g., LLaMA-2) are still far away from satisfactory for solving mathematical problem due to the complex reasoning procedures. To bridge this gap, we propose MetaMath, a fine-tuned language model that specializes in mathematical reasoning. Specifically, we start by bootstrapping mathematical questions by rewriting the question from multiple perspectives without extra knowledge, which results in a new dataset called MetaMathQA. Then we fine-tune the LLaMA-2 models on MetaMathQA. Experimental results on two popular benchmarks (i.e., GSM8K and MATH) for mathematical reasoning demonstrate that MetaMath outperforms a suite of open-source LLMs by a significant margin. Our MetaMath-7B model achieves 66.4% on GSM8K and 19.4% on MATH, exceeding the state-of-the-art models of the same size by 11.5% and 8.7%. Particularly, MetaMath-70B achieves an accuracy of 82.3% on GSM8K, slightly better than GPT-3.5-Turbo. We release all the MetaMathQA dataset, the MetaMath models with different model sizes and the training code for public use.

Zhijing Shao, Duotun Wang, Qing-Yao Tian et al.

Although neural rendering has made significant advances in creating lifelike, animatable full-body and head avatars, incorporating detailed expressions into full-body avatars remains largely unexplored. We present DEGAS, the first 3D Gaussian Splatting (3DGS)-based modeling method for full-body avatars with rich facial expressions. Trained on multiview videos of a given subject, our method learns a conditional variational autoencoder that takes both the body motion and facial expression as driving signals to generate Gaussian maps in the UV layout. To drive the facial expressions, instead of the commonly used 3D Morphable Models (3DMMs) in 3D head avatars, we propose to adopt the expression latent space trained solely on 2D portrait images, bridging the gap between 2D talking faces and 3D avatars. Leveraging the rendering capability of 3DGS and the rich expressiveness of the expression latent space, the learned avatars can be reenacted to reproduce photorealistic rendering images with subtle and accurate facial expressions. Experiments on an existing dataset and our newly proposed dataset of full-body talking avatars demonstrate the efficacy of our method. We also propose an audio-driven extension of our method with the help of 2D talking faces, opening new possibilities for interactive AI agents.

Yun Long, Haifeng Luo, Yu Zhang

This study explores the application of Large Language Models (LLMs), specifically GPT-4, in the analysis of classroom dialogue, a crucial research task for both teaching diagnosis and quality improvement. Recognizing the knowledge-intensive and labor-intensive nature of traditional qualitative methods in educational research, this study investigates the potential of LLM to streamline and enhance the analysis process. The study involves datasets from a middle school, encompassing classroom dialogues across mathematics and Chinese classes. These dialogues were manually coded by educational experts and then analyzed using a customised GPT-4 model. This study focuses on comparing manual annotations with the outputs of GPT-4 to evaluate its efficacy in analyzing educational dialogues. Time efficiency, inter-coder agreement, and inter-coder reliability between human coders and GPT-4 are evaluated. Results indicate substantial time savings with GPT-4, and a high degree of consistency in coding between the model and human coders, with some discrepancies in specific codes. These findings highlight the strong potential of LLM in teaching evaluation and facilitation.

Zelin Gao, Weichen Dai, Yu Zhang

Neural Radiance Fields (NeRF) have garnered considerable attention as a paradigm for novel view synthesis by learning scene representations from discrete observations. Nevertheless, NeRF exhibit pronounced performance degradation when confronted with sparse view inputs, consequently curtailing its further applicability. In this work, we introduce Hierarchical Geometric, Semantic, and Photometric Guided NeRF (HG3-NeRF), a novel methodology that can address the aforementioned limitation and enhance consistency of geometry, semantic content, and appearance across different views. We propose Hierarchical Geometric Guidance (HGG) to incorporate the attachment of Structure from Motion (SfM), namely sparse depth prior, into the scene representations. Different from direct depth supervision, HGG samples volume points from local-to-global geometric regions, mitigating the misalignment caused by inherent bias in the depth prior. Furthermore, we draw inspiration from notable variations in semantic consistency observed across images of different resolutions and propose Hierarchical Semantic Guidance (HSG) to learn the coarse-to-fine semantic content, which corresponds to the coarse-to-fine scene representations. Experimental results demonstrate that HG3-NeRF can outperform other state-of-the-art methods on different standard benchmarks and achieve high-fidelity synthesis results for sparse view inputs.

Yu Zhang, Dong Guo, Fang Wu et al.

Large Language Models (LLMs) with extended context lengths face significant computational challenges during the pre-filling phase, primarily due to the quadratic complexity of self-attention. Existing methods typically employ dynamic pattern matching and block-sparse low-level implementations. However, their reliance on local information for pattern identification fails to capture global contexts, and the coarse granularity of blocks leads to persistent internal sparsity, resulting in suboptimal accuracy and efficiency. To address these limitations, we propose \textbf{AnchorAttention}, a difference-aware, dynamic sparse attention mechanism that efficiently identifies critical attention regions at a finer stripe granularity while adapting to global contextual information, achieving superior speed and accuracy. AnchorAttention comprises three key components: (1) \textbf{Pattern-based Anchor Computation}, leveraging the commonalities present across all inputs to rapidly compute a set of near-maximum scores as the anchor; (2) \textbf{Difference-aware Stripe Sparsity Identification}, performing difference-aware comparisons with the anchor to quickly obtain discrete coordinates of significant regions in a stripe-like sparsity pattern; (3) \textbf{Fine-grained Sparse Computation}, replacing the traditional contiguous KV block loading approach with simultaneous discrete KV position loading to maximize sparsity rates while preserving full hardware computational potential. With its finer-grained sparsity strategy, \textbf{AnchorAttention} achieves higher sparsity rates at the same recall level, significantly reducing computation time. Compared to previous state-of-the-art methods, at a text length of 128k, it achieves a speedup of 1.44$\times$ while maintaining higher recall rates.

Songlin Yang, Bailin Wang, Yu Zhang et al.

Transformers with linear attention (i.e., linear transformers) and state-space models have recently been suggested as a viable linear-time alternative to transformers with softmax attention. However, these models still underperform transformers especially on tasks that require in-context retrieval. While more expressive variants of linear transformers which replace the additive update in linear transformers with the delta rule (DeltaNet) have been found to be more effective at associative recall, existing algorithms for training such models do not parallelize over sequence length and are thus inefficient to train on modern hardware. This work describes a hardware-efficient algorithm for training linear transformers with the delta rule, which exploits a memory-efficient representation for computing products of Householder matrices. This algorithm allows us to scale up DeltaNet to standard language modeling settings. We train a 1.3B model for 100B tokens and find that it outperforms recent linear-time baselines such as Mamba and GLA in terms of perplexity and zero-shot performance on downstream tasks. We also experiment with two hybrid models which combine DeltaNet layers with (1) sliding-window attention layers every other layer or (2) two global attention layers, and find that these hybrids outperform strong transformer baselines.

Zeyu Wan, Yu Zhang, Bin He et al.

Currently, the improvement of LiDAR poses estimation accuracy is an urgent need for mobile robots. Research indicates that diverse LiDAR points have different influences on the accuracy of pose estimation. This study aimed to select a good point set to enhance accuracy. Accordingly, the sensitivity and uncertainty of LiDAR point residuals were formulated as a fundamental basis for derivation and analysis. High-sensitivity and low -uncertainty point residual terms are preferred to achieve higher pose estimation accuracy. The proposed selection method has been theoretically proven to be capable of achieving a global statistical optimum. It was tested on artificial data and compared with the KITTI benchmark. It was also implemented in LiDAR odometry (LO) and LiDAR inertial odometry (LIO), both indoors and outdoors. The experiments revealed that utilizing selected LiDAR point residuals simultaneously enhances optimization accuracy, decreases residual terms, and guarantees real-time performance.

Yu'an Chen, Guangda Chen, Lifan Pan et al.

In this paper, we propose a map-based end-to-end DRL approach for three-dimensional (3D) obstacle avoidance in a partially observed environment, which is applied to achieve autonomous navigation for an indoor mobile robot using a depth camera with a narrow field of view. We first train a neural network with LSTM units in a 3D simulator of mobile robots to approximate the Q-value function in double DRQN. We also use a curriculum learning strategy to accelerate and stabilize the training process. Then we deploy the trained model to a real robot to perform 3D obstacle avoidance in its navigation. We evaluate the proposed approach both in the simulated environment and on a robot in the real world. The experimental results show that the approach is efficient and easy to be deployed, and it performs well for 3D obstacle avoidance with a narrow observation angle, which outperforms other existing DRL-based models by 15.5% on success rate.

Yingjie Wang, Qiuyu Mao, Hanqi Zhu et al.

In this survey, we first introduce the background of popular sensors used for self-driving, their data properties, and the corresponding object detection algorithms. Next, we discuss existing datasets that can be used for evaluating multi-modal 3D object detection algorithms. Then we present a review of multi-modal fusion based 3D detection networks, taking a close look at their fusion stage, fusion input and fusion granularity, and how these design choices evolve with time and technology. After the review, we discuss open challenges as well as possible solutions. We hope that this survey can help researchers to get familiar with the field and embark on investigations in the area of multi-modal 3D object detection.

Weichen Dai, Yu Zhang, Shenzhou Chen et al.

Visible images have been widely used for motion estimation. Thermal images, in contrast, are more challenging to be used in motion estimation since they typically have lower resolution, less texture, and more noise. In this paper, a novel dataset for evaluating the performance of multi-spectral motion estimation systems is presented. All the sequences are recorded from a handheld multi-spectral device. It consists of a standard visible-light camera, a long-wave infrared camera, an RGB-D camera, and an inertial measurement unit (IMU). The multi-spectral images, including both color and thermal images in full sensor resolution (640 x 480), are obtained from a standard and a long-wave infrared camera at 32Hz with hardware-synchronization. The depth images are captured by a Microsoft Kinect2 and can have benefits for learning cross-modalities stereo matching. For trajectory evaluation, accurate ground-truth camera poses obtained from a motion capture system are provided. In addition to the sequences with bright illumination, the dataset also contains dim, varying, and complex illumination scenes. The full dataset, including raw data and calibration data with detailed data format specifications, is publicly available.

Weichen Dai, Yu Zhang, Donglei Sun et al.

Multi-spectral sensors consisting of a standard (visible-light) camera and a long-wave infrared camera can simultaneously provide both visible and thermal images. Since thermal images are independent from environmental illumination, they can help to overcome certain limitations of standard cameras under complicated illumination conditions. However, due to the difference in the information source of the two types of cameras, their images usually share very low texture similarity. Hence, traditional texture-based feature matching methods cannot be directly applied to obtain stereo correspondences. To tackle this problem, a multi-spectral visual odometry method without explicit stereo matching is proposed in this paper. Bundle adjustment of multi-view stereo is performed on the visible and the thermal images using direct image alignment. Scale drift can be avoided by additional temporal observations of map points with the fixed-baseline stereo. Experimental results indicate that the proposed method can provide accurate visual odometry results with recovered metric scale. Moreover, the proposed method can also provide a metric 3D reconstruction in semi-dense density with multi-spectral information, which is not available from existing multi-spectral methods.

Yong Zhang, Yu Zhang, Zhao Zhang et al.

Traditional human activity recognition (HAR) based on time series adopts sliding window analysis method. This method faces the multi-class window problem which mistakenly labels different classes of sampling points within a window as a class. In this paper, a HAR algorithm based on U-Net is proposed to perform activity labeling and prediction at each sampling point. The activity data of the triaxial accelerometer is mapped into an image with the single pixel column and multi-channel which is input into the U-Net network for training and recognition. Our proposal can complete the pixel-level gesture recognition function. The method does not need manual feature extraction and can effectively identify short-term behaviors in long-term activity sequences. We collected the Sanitation dataset and tested the proposed scheme with four open data sets. The experimental results show that compared with Support Vector Machine (SVM), k-Nearest Neighbor (kNN), Decision Tree(DT), Quadratic Discriminant Analysis (QDA), Convolutional Neural Network (CNN) and Fully Convolutional Networks (FCN) methods, our proposal has the highest accuracy and F1-socre in each dataset, and has stable performance and high robustness. At the same time, after the U-Net has finished training, our proposal can achieve fast enough recognition speed.

Fei Han, Xue Yang, Yu Zhang et al.

Apprenticeship learning has recently attracted a wide attention due to its capability of allowing robots to learn physical tasks directly from demonstrations provided by human experts. Most previous techniques assumed that the state space is known a priori or employed simple state representations that usually suffer from perceptual aliasing. Different from previous research, we propose a novel approach named Sequence-based Multimodal Apprenticeship Learning (SMAL), which is capable to simultaneously fusing temporal information and multimodal data, and to integrate robot perception with decision making. To evaluate the SMAL approach, experiments are performed using both simulations and real-world robots in the challenging search and rescue scenarios. The empirical study has validated that our SMAL approach can effectively learn plans for robots to make decisions using sequence of multimodal observations. Experimental results have also showed that SMAL outperforms the baseline methods using individual images.