arXiv subjects

Lei Yu

Publications and source records attributed to Lei Yu.

At least 127 records · Page 7Linked to original sources

arXiv2024

First Activations Matter: Training-Free Methods for Dynamic Activation in Large Language Models

Dynamic activation (DA) techniques, such as DejaVu and MoEfication, have demonstrated their potential to significantly enhance the inference efficiency of large language models (LLMs). However, these techniques often rely on ReLU activation functions or require additional parameters and training to maintain performance. This paper introduces a training-free Threshold-based Dynamic Activation(TDA) method that leverage sequence information to exploit the inherent sparsity of models across various architectures. This method is designed to accelerate generation speed by 18-25\% without significantly compromising task performance, thereby addressing the limitations of existing DA techniques. Moreover, we delve into the root causes of LLM sparsity and theoretically analyze two of its critical features: history-related activation uncertainty and semantic-irrelevant activation inertia. Our comprehensive analyses not only provide a robust theoretical foundation for DA methods but also offer valuable insights to guide future research in optimizing LLMs for greater efficiency and effectiveness.

cs.CL↗

arXiv2024

A Reparameterized Discrete Diffusion Model for Text Generation

This work studies discrete diffusion probabilistic models with applications to natural language generation. We derive an alternative yet equivalent formulation of the sampling from discrete diffusion processes and leverage this insight to develop a family of reparameterized discrete diffusion models. The derived generic framework is highly flexible, offers a fresh perspective of the generation process in discrete diffusion models, and features more effective training and decoding techniques. We conduct extensive experiments to evaluate the text generation capability of our model, demonstrating significant improvements over existing diffusion models.

cs.CL↗

arXiv2024

POA: Pre-training Once for Models of All Sizes

Large-scale self-supervised pre-training has paved the way for one foundation model to handle many different vision tasks. Most pre-training methodologies train a single model of a certain size at one time. Nevertheless, various computation or storage constraints in real-world scenarios require substantial efforts to develop a series of models with different sizes to deploy. Thus, in this study, we propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All), to tackle this aforementioned issue. Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. At each pre-training step, we randomly sample a sub-network from the original student to form the elastic student and train all branches in a self-distilling fashion. Once pre-trained, POA allows the extraction of pre-trained models of diverse sizes for downstream tasks. Remarkably, the elastic student facilitates the simultaneous pre-training of multiple models with different sizes, which also acts as an additional ensemble of models of various sizes to enhance representation learning. Extensive experiments, including k-nearest neighbors, linear probing evaluation and assessments on multiple downstream tasks demonstrate the effectiveness and advantages of our POA. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones, producing around a hundred models with different sizes through a single pre-training session. The code is available at: https://github.com/Qichuzyy/POA.

cs.CV↗

arXiv2024

The most distant HI galaxies discovered by the 500 m dish FAST

Neutral hydrogen (HI) is the primary component of the cool interstellar medium (ISM) and is the reservoir of fuel for star formation. Owing to the sensitivity of existing radio telescopes, our understanding of the evolution of the ISM in galaxies remains limited, as it is based on only a few hundred galaxies detected in HI beyond the local Universe. With the high sensitivity of the Five-hundred-meter Aperture Spherical radio Telescope (FAST), we carried out a blind HI search, the FAST Ultra-Deep Survey (FUDS), which extends to redshifts up to 0.42 and a sensitivity of 50 $\rm μJy \cdot beam^{-1}$. Here, we report the first discovery of six galaxies in HI at $z>0.38$. For these galaxies, the FAST angular resolution of $\sim\,4'$ corresponds to a mean linear size of $\sim1.3\,h_{70}^{-1}\,$Mpc. These galaxies are among the most distant HI emission detections known, with one having the most massive HI content ($10^{10.93 \pm 0.04}~h_{70}^{-2}\, \rm M_\odot$). Using recent data from the DESI survey, and new observations with the Hale, BTA, and Keck telescopes, optical counterparts are detected for all galaxies within the 3-$σ$ positional uncertainty ($0.5\,h_{70}^{-1}\,$Mpc) and $\rm 200\,km \cdot s^{-1}$ in recession velocity. Assuming that the dominant source of HI is the identified optical counterpart, we find an evidence of evolution in the HI content of galaxies over the last 4.2 Gyr. Our new high-redshift HI galaxy sample provides the opportunity to better investigate the evolution of cool gas in galaxies. A larger sample size in the future will allow us to refine our knowledge of the formation and evolution of galaxies.

astro-ph.GA↗

arXiv2024

QQQ: Quality Quattuor-Bit Quantization for Large Language Models

Quantization is a proven effective method for compressing large language models. Although popular techniques like W8A8 and W4A16 effectively maintain model performance, they often fail to concurrently speed up the prefill and decoding stages of inference. W4A8 is a promising strategy to accelerate both of them while usually leads to a significant performance degradation. To address these issues, we present QQQ, a Quality Quattuor-bit Quantization method with 4-bit weights and 8-bit activations. QQQ employs adaptive smoothing and Hessian-based compensation, significantly enhancing the performance of quantized models without extensive training. Furthermore, we meticulously engineer W4A8 GEMM kernels to increase inference speed. Our specialized per-channel W4A8 GEMM and per-group W4A8 GEMM achieve impressive speed increases of 3.67$\times$ and 3.29 $\times$ over FP16 GEMM. Our extensive experiments show that QQQ achieves performance on par with existing state-of-the-art LLM quantization methods while significantly accelerating inference, achieving speed boosts up to 2.24 $\times$, 2.10$\times$, and 1.25$\times$ compared to FP16, W8A8, and W4A16, respectively.

cs.LG↗

arXiv2024

Rényi--Sobolev Inequalities and Connections to Spectral Graph Theory

In this paper, we generalize the log-Sobolev inequalities to Rényi--Sobolev inequalities by replacing the entropy with the two-parameter entropy, which is a generalized version of entropy and closely related to Rényi divergences. We derive the sharp nonlinear dimension-free version of this kind of inequalities. Interestingly, the resultant inequalities show a transition phenomenon depending on the parameters. We then connect Rényi--Sobolev inequalities to contractive and data-processing inequalities, concentration inequalities, and spectral graph theory. Our proofs in this paper are based on the information-theoretic characterization of the Rényi--Sobolev inequalities, as well as the method of types.

math.PR↗

arXiv2024

A Compact Variable Stiffness Actuator for Agile Legged Locomotion

The legged robots with variable stiffness actuators (VSAs) can achieve energy-efficient and versatile locomotion. However, equipping legged robots with VSAs in real-world application is usually restricted by (i) the redundant mechanical structure design, (ii) limited stiffness variation range and speed, (iii) high energy consumption in stiffness modulation, and (iv) the lack of online stiffness control structure with high performance. In this paper, we present a novel Variable-Length Leaf-Spring Actuator (VLLSA) designed for legged robots that aims to address the aforementioned limitations. The design is based on leaf-spring mechanism and we improve the structural design to make the proposed VSA (i) compact and lightweight in mechanical structure, (ii) precise in theoretical modeling, and (iii) capable of modulating stiffness with wide range, fast speed, low energy consumption and high control performance. Hardware experiments including in-place and forward hopping validate advantages of the proposed VLLSA.

cs.RO↗

arXiv2024

Design and Control of a Bio-inspired Wheeled Bipedal Robot

Wheeled bipedal robots (WBRs) have the capability to execute agile and versatile locomotion tasks. This paper focuses on improving the dynamic performance of WBRs through innovations in both hardware and software development. Inspired by the human barbell squat, a bionic mechanical design is proposed and implemented as shown in Fig. 1. It distributes the weight onto its hip and knee joints to improve the effectiveness of joint motors while maintaining a relatively large workspace of the base link. Meanwhile, a novel model-based controller is devised, synthesizing height-variable wheeled linear inverted pendulum (HV-wLIP) model, Control Lyapunov Function (CLF) and whole-body dynamics for theoretically guaranteed stability and efficient computation. Compared with other alternatives, as a more accurate approximation of the WBR dynamics, the HV-wLIP can enable more agile response and provide theory basis for WBR controller design. Experimental results demonstrate that the robot could perform human-like deep squat, and is capable of maintaining tracking CoM velocity while manipulating base states. Furthermore, it exhibited robustness against external disturbances and unknown terrains even in the wild.

cs.RO↗

arXiv2024

On the Robustness of Graph Reduction Against GNN Backdoor

Graph Neural Networks (GNNs) are gaining popularity across various domains due to their effectiveness in learning graph-structured data. Nevertheless, they have been shown to be susceptible to backdoor poisoning attacks, which pose serious threats to real-world applications. Meanwhile, graph reduction techniques, including coarsening and sparsification, which have long been employed to improve the scalability of large graph computational tasks, have recently emerged as effective methods for accelerating GNN training on large-scale graphs. However, the current development and deployment of graph reduction techniques for large graphs overlook the potential risks of data poisoning attacks against GNNs. It is not yet clear how graph reduction interacts with existing backdoor attacks. This paper conducts a thorough examination of the robustness of graph reduction methods in scalable GNN training in the presence of state-of-the-art backdoor attacks. We performed a comprehensive robustness analysis across six coarsening methods and six sparsification methods for graph reduction, under three GNN backdoor attacks against three GNN architectures. Our findings indicate that the effectiveness of graph reduction methods in mitigating attack success rates varies significantly, with some methods even exacerbating the attacks. Through detailed analyses of triggers and poisoned nodes, we interpret our findings and enhance our understanding of how graph reduction influences robustness against backdoor attacks. These results highlight the critical need for incorporating robustness considerations in graph reduction for GNN training, ensuring that enhancements in computational efficiency do not compromise the security of GNN systems.

cs.LG↗

arXiv2024

Rényi Common Information for Doubly Symmetric Binary Sources

In this note, we provide analytic expressions for the Rényi common information of orders in $(1,\infty)$ for the doubly symmetric binary source (DSBS). Until now, analytic expressions for the Rényi common information of all orders in $[0,\infty]$ have been completely known for this source. We also consider the Rényi common information of all orders in $[-\infty,0)$ and evaluate it for the DSBS. We provide a sufficient condition under which the Rényi common information of such orders coincides with Wyner's common information for the DSBS. Based on numerical analysis, we conjecture that there is a certain phase transition as the crossover probability increasing for the Rényi common information of negative orders for the DSBS. Our proofs are based on a lemma on splitting of the entropy and the analytic expression of relaxed Wyner's common information.

cs.IT↗

arXiv2024

MOYU: A Theoretical Study on Massive Over-activation Yielded Uplifts in LLMs

Massive Over-activation Yielded Uplifts(MOYU) is an inherent property of large language models, and dynamic activation(DA) based on the MOYU property is a clever yet under-explored strategy designed to accelerate inference in these models. Existing methods that utilize MOYU often face a significant 'Impossible Trinity': struggling to simultaneously maintain model performance, enhance inference speed, and extend applicability across various architectures. Due to the theoretical ambiguities surrounding MOYU, this paper elucidates the root cause of the MOYU property and outlines the mechanisms behind two primary limitations encountered by current DA methods: 1) history-related activation uncertainty, and 2) semantic-irrelevant activation inertia. Our analysis not only underscores the limitations of current dynamic activation strategies within large-scale LLaMA models but also proposes opportunities for refining the design of future sparsity schemes.

cs.LG↗

arXiv2024

Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations

State-of-the-art language models (LMs) sometimes generate non-factual hallucinations that misalign with world knowledge. To explore the mechanistic causes of these hallucinations, we create diagnostic datasets with subject-relation queries and adapt interpretability methods to trace hallucinations through internal model representations. We discover two general and distinct mechanistic causes of hallucinations shared across LMs (Llama-2, Pythia, GPT-J): 1) knowledge enrichment hallucinations: insufficient subject attribute knowledge in lower layer MLPs, and 2) answer extraction hallucinations: failure to select the correct object attribute in upper layer attention heads. We also found these two internal mechanistic causes of hallucinations are reflected in external manifestations. Based on insights from our mechanistic analysis, we propose a novel hallucination mitigation method through targeted restoration of the LM's internal fact recall pipeline, demonstrating superior performance compared to baselines.

cs.CL↗

arXiv2024

E-SAGE: Explainability-based Defense Against Backdoor Attacks on Graph Neural Networks

Graph Neural Networks (GNNs) have recently been widely adopted in multiple domains. Yet, they are notably vulnerable to adversarial and backdoor attacks. In particular, backdoor attacks based on subgraph insertion have been shown to be effective in graph classification tasks while being stealthy, successfully circumventing various existing defense methods. In this paper, we propose E-SAGE, a novel approach to defending GNN backdoor attacks based on explainability. We find that the malicious edges and benign edges have significant differences in the importance scores for explainability evaluation. Accordingly, E-SAGE adaptively applies an iterative edge pruning process on the graph based on the edge scores. Through extensive experiments, we demonstrate the effectiveness of E-SAGE against state-of-the-art graph backdoor attacks in different attack settings. In addition, we investigate the effectiveness of E-SAGE against adversarial attacks.

cs.CR↗

arXiv2024

Exact Exponents for Concentration and Isoperimetry in Product Polish Spaces

In this paper, we derive variational formulas for the asymptotic exponents (i.e., convergence rates) of the concentration and isoperimetric functions in the product Polish probability space under certain mild assumptions. These formulas are expressed in terms of relative entropies (which are from information theory) and optimal transport cost functionals (which are from optimal transport theory). Hence, our results verify an intimate connection among information theory, optimal transport, and concentration of measure or isoperimetric inequalities. In the concentration regime, the corresponding variational formula is in fact a dimension-free bound in the sense that this bound is valid for any dimension. A cardinality bound for the alphabet of the auxiliary random variable in the expression of the asymptotic isoperimetric exponent is provided, which makes the expression computable by a finite-dimensional program for the finite alphabet case. We lastly apply our results to obtain an isoperimetric inequality in the classic isoperimetric setting, which is asymptotically sharp under certain conditions. The proofs in this paper are based on information-theoretic and optimal transport techniques.

math.PR↗

arXiv2024

Dynamic Activation Pitfalls in LLaMA Models: An Empirical Study

In this work, we systematically investigate the efficacy of dynamic activation mechanisms within the LLaMA family of language models. Despite the potential of dynamic activation methods to reduce computation and increase speed in models using the ReLU activation function, our empirical findings have uncovered several inherent pitfalls in the current dynamic activation schemes. Through extensive experiments across various dynamic activation strategies, we demonstrate that LLaMA models usually underperform when compared to their ReLU counterparts, particularly in scenarios demanding high sparsity ratio. We attribute these deficiencies to a combination of factors: 1) the inherent complexity of dynamically predicting activation heads and neurons; 2) the inadequate sparsity resulting from activation functions; 3) the insufficient preservation of information resulting from KV cache skipping. Our analysis not only sheds light on the limitations of dynamic activation in the context of large-scale LLaMA models but also proposes roadmaps for enhancing the design of future sparsity schemes.

cs.LG↗

arXiv2024

Super-Resolving Blurry Images with Events

Super-resolution from motion-blurred images poses a significant challenge due to the combined effects of motion blur and low spatial resolution. To address this challenge, this paper introduces an Event-based Blurry Super Resolution Network (EBSR-Net), which leverages the high temporal resolution of events to mitigate motion blur and improve high-resolution image prediction. Specifically, we propose a multi-scale center-surround event representation to fully capture motion and texture information inherent in events. Additionally, we design a symmetric cross-modal attention module to fully exploit the complementarity between blurry images and events. Furthermore, we introduce an intermodal residual group composed of several residual dense Swin Transformer blocks, each incorporating multiple Swin Transformer layers and a residual connection, to extract global context and facilitate inter-block feature aggregation. Extensive experiments show that our method compares favorably against state-of-the-art approaches and achieves remarkable performance.

cs.CV↗

arXiv2024

Learning Monocular Depth from Focus with Event Focal Stack

Depth from Focus estimates depth by determining the moment of maximum focus from multiple shots at different focal distances, i.e. the Focal Stack. However, the limited sampling rate of conventional optical cameras makes it difficult to obtain sufficient focus cues during the focal sweep. Inspired by biological vision, the event camera records intensity changes over time in extremely low latency, which provides more temporal information for focus time acquisition. In this study, we propose the EDFF Network to estimate sparse depth from the Event Focal Stack. Specifically, we utilize the event voxel grid to encode intensity change information and project event time surface into the depth domain to preserve per-pixel focal distance information. A Focal-Distance-guided Cross-Modal Attention Module is presented to fuse the information mentioned above. Additionally, we propose a Multi-level Depth Fusion Block designed to integrate results from each level of a UNet-like architecture and produce the final output. Extensive experiments validate that our method outperforms existing state-of-the-art approaches.

cs.CV↗

arXiv2024

Non-Uniform Exposure Imaging via Neuromorphic Shutter Control

By leveraging the blur-noise trade-off, imaging with non-uniform exposures largely extends the image acquisition flexibility in harsh environments. However, the limitation of conventional cameras in perceiving intra-frame dynamic information prevents existing methods from being implemented in the real-world frame acquisition for real-time adaptive camera shutter control. To address this challenge, we propose a novel Neuromorphic Shutter Control (NSC) system to avoid motion blurs and alleviate instant noises, where the extremely low latency of events is leveraged to monitor the real-time motion and facilitate the scene-adaptive exposure. Furthermore, to stabilize the inconsistent Signal-to-Noise Ratio (SNR) caused by the non-uniform exposure times, we propose an event-based image denoising network within a self-supervised learning paradigm, i.e., SEID, exploring the statistics of image noises and inter-frame motion information of events to obtain artificial supervision signals for high-quality imaging in real-world scenes. To illustrate the effectiveness of the proposed NSC, we implement it in hardware by building a hybrid-camera imaging prototype system, with which we collect a real-world dataset containing well-synchronized frames and events in diverse scenarios with different target scenes and motion patterns. Experiments on the synthetic and real-world datasets demonstrate the superiority of our method over state-of-the-art approaches.

cs.CV↗