Ziyi Guan, Jason Chun Lok Li, Zhijian Hou, Pingping Zhang, Donglai Xu, Yuzhi Zhao, Mengyang Wu, Jinpeng Chen, Thanh-Toan Nguyen, Pengfei Xian, Wenao Ma, Shengchao Qin, Graziano Chesi, Ngai Wong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large language models (LLMs) have seen substantial growth, necessitating efficient model pruning techniques. Existing post-training pruning methods primarily measure weight importance in converged dense models, often overlooking changes in weight significance during the pruning process, leading to performance degradation. To address this issue, we present LLM-Barber (Block-Aware Rebuilder for Sparsity Mask in One-Shot), a novel one-shot pruning framework that rebuilds the sparsity mask of pruned models without any retraining or weight reconstruction. LLM-Barber incorporates block-aware error optimization across Self-Attention and MLP blocks, facilitating global performance optimization. We are the first to employ the product of weights and gradients as a pruning metric in the context of LLM post-training pruning. This enables accurate identification of weight importance in massive models and significantly reduces computational complexity compared to methods using second-order information. Our experiments show that LLM-Barber efficiently prunes models from LLaMA and OPT families (7B to 13B) on a single A100 GPU in just 30 minutes, achieving state-of-the-art results in both perplexity and zero-shot performance across various language benchmarks.

3D Gaussian inpainting, a critical technique for numerous applications in virtual reality and multimedia, has made significant progress with pretrained diffusion models. However, ensuring multi-view consistency, an essential requirement for high-quality inpainting, remains a key challenge. In this work, we present PAInpainter, a novel approach designed to advance 3D Gaussian inpainting by leveraging perspective-aware content propagation and consistency verification across multi-view inpainted images. Our method iteratively refines inpainting and optimizes the 3D Gaussian representation with multiple views adaptively sampled from a perspective graph. By propagating inpainted images as prior information and verifying consistency across neighboring views, PAInpainter substantially enhances global consistency and texture fidelity in restored 3D scenes. Extensive experiments demonstrate the superiority of PAInpainter over existing methods. Our approach achieves superior 3D inpainting quality, with PSNR scores of 26.03 dB and 29.51 dB on the SPIn-NeRF and NeRFiller datasets, respectively, highlighting its effectiveness and generalization capability.

In this paper, we focus on the problem of distributed state estimation for a target system driven by continuous-time linear time-invariant (LTI) system with unknown inputs, where the communication channels are disturbed by stochastic noise and subjected to impulsive sequential attacks. In multi-agent systems (MASs), agents usually move along relatively smooth paths, but occasionally take evasive actions or make sudden turns. This pattern of motion results in relatively small communication disturbances most of the time, but large communication disturbances can occasionally occur. First, this communication disturbances are modeled as a hybrid form: stochastic noise and impulsive sequential attacks, where stochastic noise stands for relatively small communication disturbances most of the time, and impulsive sequential attacks stand for occasionally large communication disturbances. Then, a practical/bounded distributed unknown input observer (DUIO) under stochastic noise and impulsive sequential attack is proposed. Furthermore, the sufficient conditions for the realization of the practical/bounded DUIO in the sense of mean square convergence are presented by graph theory and stochastic analysis. The error bound that depends on the coupling gain, stochastic noise intensity, and average impulsive interval is provided. Finally, the validity of the designed practical/bounded DUIO is confirmed by numerical simulations. Note to Practitioners—Information exchange among agents is a critical for achieving coordination in MASs. Most existing works usually assume that the communication is ideal, i.e., each agent is able to accurately obtain information from its neighbors or is only disturbed by stochastic noise. However, agents generally move along relatively smooth paths but may occasionally execute evasive maneuvers or abrupt turns in MASs. This motion pattern usually results in relatively small communication disturbances, though in certain instances, significant disruptions may occur. How to effectively deal with the impact of such non-ideal communication environments on system performance and ensure that agents can still maintain stable cooperation in complex dynamic environments has become one of the critical topics in the field of control. Specifically, this problem involves several challenges: First, the random and non-Gaussian characteristics of communication disturbances render traditional filtering and estimation methods difficult to apply directly; Second, the burstiness and uncertainty of impulsive noise can lead to severe fluctuations in system states, thereby affecting coordination among agents. In light of this, this paper models such communication disturbances as stochastic impulsive noise and proposes a practical/bounded DUIO under stochastic impulsive noise.

Resistive random-access memory (RRAM), with its programmable and nonvolatile conductance, permits compute-in-memory (CIM) at a much higher energy efficiency than the traditional von Neumann architecture, making it a promising candidate for edge AI. Nonetheless, the fixed-size crossbar tiles on RRAM are inherently unfit for conventional pyramid-shape convolutional neural networks (CNNs) that incur low crossbar utilization. To this end, we recognize the mixed-signal (digital-analog) nature in RRAM circuits and customize an isotropic shift-pointwise network that exploits digital shift operations for efficient spatial mixing and analog pointwise operations for channel mixing. To fast ablate various shift-pointwise topologies, a new recon-figurable energy-efficient shift module is designed and packaged into a seamless mixed-domain simulator. The optimized design achieves a near-100% crossbar utilization, providing a state-of-the-art INT8 accuracy of 94.88% (76.55%) on the CIFAR-10 (CIFAR-100) dataset with 1.6M parameters, which sets a new standard for RRAM-based AI accelerators.

This paper addresses the problem of determining the peak of the response to a linear time-invariant (LTI) signal of a linear system whose system matrices are rational functions of an uncertainty vector constrained into a convex bounded polytope. The uncertainty can be time-invariant, bounded-rate time-varying or arbitrarily time-varying. A novel approach based on linear matrix inequalities (LMIs) is proposed for obtaining upper bounds of the sought peak based on the construction of a structured polynomial Lyapunov function in the state and in the uncertainty. A priori and a posteriori conditions for establishing optimality of the obtained upper bounds are also provided. As shown by some numerical examples, which includes the model of an electric circuit, the proposed approach may have significant advantages with respect to the existing methods in terms of conservatism or computational burden.

We introduce a new kind of linear transform named Deformable Butterfly (DeBut) that generalizes the conventional butterfly matrices and can be adapted to various input-output dimensions. It inherits the fine-to-coarse-grained learnable hierarchy of traditional butterflies and when deployed to neural networks, the prominent structures and sparsity in a DeBut layer constitutes a new way for network compression. We apply DeBut as a drop-in replacement of standard fully connected and convolutional layers, and demonstrate its superiority in homogenizing a neural network and rendering it favorable properties such as light weight and low inference complexity, without compromising accuracy. The natural complexity-accuracy tradeoff arising from the myriad deformations of a DeBut layer also opens up new rooms for analytical and practical research. The codes and Appendix are publicly available at: https://github.com/ruilin0212/DeBut.

Elasticities in depth, width, kernel size and resolution have been explored in compressing deep neural networks (DNNs). Recognizing that the kernels in a convolutional neural network (CNN) are 4-way tensors, we further exploit a new elasticity dimension along the input-output channels. Specifically, a novel nuclear-norm rank minimization factorization (NRMF) approach is proposed to dynamically and globally search for the reduced tensor ranks during training. Correlation between tensor ranks across multiple layers is revealed, and a graceful tradeoff between model size and accuracy is obtained. Experiments then show the superiority of NRMF over the previous non-elastic variational Bayesian matrix factorization (VBMF) scheme.

This paper addresses the problem of establishing whether a matrix with rational dependence on a complex parameter and its conjugate is Hurwitz (i.e., has all eigenvalues with negative real part) over the complex unit circumference. A necessary and sufficient condition is proposed, which requires to test stability of a constant matrix and to solve a semidefinite program (SDP) built through the use of determinants. Some numerical examples show the application of the proposed condition for stability analysis of 2D systems with mixed signals, and that the computational burden of the proposed condition may be dramatically smaller than that of existing conditions.

This paper addresses the problem of determining parametric linear quadratic regulators (LQRs) for polytopic discrete-time systems. Specifically, it is supposed that the matrices of the system are linear functions of a vector of parameters constrained over the simplex. It is shown that a candidate for the sought parametric LQR can be obtained by solving a semidefinite program (SDP) built through homogeneous polynomially-dependent quadratic Lyapunov functions (HPD-QLFs) of chosen degree. In particular, it is shown that the found candidate is guaranteed to approximate arbitrarily well the true parametric LQR by using a degree sufficiently large.

The problem of visual servoing to reach the desired location keeping elliptical projections in the camera field of view (FOV) while following a straight line is considered. The proposed approach is representing the whole path with seven polynomials of a path abscise: variables in polynomial coefficients for translational path being zero to represent a minimum path length and for rotational part being adjustable satisfying the FOV limit. The planned elliptical trajectories are tracked by an image-based visual servoing (IBVS) controller. The proposed strategy is verified by a simulational case with a circle and a superposed point, where a traditional IBVS controller directs the camera a detour to the ground, the proposed approach however keeps straight the camera trajectory and also the circle visible. In addition, a six degrees of freedom (6-DoF) articulated arm mounted with a pinhole camera is used to validate the proposed method by taking three Christmas balls as the target.

Vision feedback control techniques are desirable for a wide range of robotics applications due to their robustness to image noise and modeling errors. However in the case of a robot-mounted camera, they encounter difficulties when the camera traverses large displacements. This scenario necessitates continuous visual target feedback during the robot motion, while simultaneously considering the robot's self- and external-constraints. Herein, we propose to combine workspace (Cartesian space) path-planning with robot teach-by-demonstration to address the visibility constraint, joint limits and “whole arm” collision avoidance for vision-based control of a robot manipulator. User demonstration data generates safe regions for robot motion with respect to joint limits and potential “whole arm” collisions. Our algorithm uses these safe regions to generate new feasible trajectories under a visibility constraint that achieves the desired view of the target (e.g., a pre-grasping location) in new, undemonstrated locations. Experiments with a 7-DOF articulated arm validate the proposed method.

This paper investigates robust consensus for a class of uncertain multi-agent dynamical systems. Specifically, it is supposed that the system is described by a weighted adjacency matrix whose entries are polynomial functions of an uncertain vector constrained in a semi-algebraic set. For this uncertain topology, we provide necessary and sufficient conditions for ensuring robust first-order consensus and robust second-order consensus, in both cases of positive and non-positive weighted adjacency matrices. Moreover, we show how these conditions can be investigated through convex programming by using standard software. Some numerical examples illustrate the proposed results.

This correspondence addresses the analysis of the steady states of uncertain genetic regulatory networks (GRNs). The uncertainty is represented as a vector constrained in a given set that affects the coefficients of the mathematical model of the GRN. It is shown how regions containing all possible steady states can be estimated via an iterative strategy that progressively splits the concentration space into smaller sets, discarding those that are guaranteed not to contain equilibrium points of the considered model. This strategy is based on worst case evaluations of some appropriate functions of the uncertainty via linear matrix inequality optimization.

Image noise unavoidably affects the available image points that are used in visual-servoing schemes to steer a robot end-effector toward a desired location. As a consequence, letting the image points in the current view converge to those in the desired view does not ensure that the camera converges accurately to the desired location. This paper investigates the selection of object configurations to minimize the worst-case positioning error due to the presence of image noise. In particular, a strategy based on linear matrix inequalities (LMIs) and barrier functions is proposed to compute upper and lower bounds of this error for a given maximum error of the image points. This strategy can be applied to problems such as selecting an optimal subset of object points or determining an optimal position of an object in the scene. Some examples illustrate the use of the proposed strategy in such problems.

Estimating the camera pose in stereo vision systems is an important issue in computer vision and robotics. One popular way to handle this problem consists of determining the essential matrix which minimizes the algebraic error obtained from image point correspondences. Unfortunately, this search amounts to solving a nonconvex optimization, and the existing methods either rely on some approximations in order to get rid of the non-convexity or provide a solution that may be affected by the presence of local minima. This paper proposes a new approach to address this search without presenting such problems. In particular, we show that the sought essential matrix can be obtained by solving a convex optimization built through a suitable reformulation of the considered minimization via appropriate techniques for representing polynomials. Numerical results show the proposed approach compares favorably with some standard methods in both cases of synthetic data and real data.

Path-planning allows one to steer a camera to a desired location while taking into account the presence of constraints such as visibility, workspace, and joint limits. Unfortunately, the planned path can be significantly different from the real path due to the presence of uncertainty on the available data, with the consequence that some constraints may be not fulfilled by the real path even if they are satisfied by the planned path. In this paper we address the problem of performing robust path-planning, i.e. computing a path that satisfies the required constraints not only for the nominal model as in traditional path-planning but rather for a family of admissible models. Specifically, we consider an uncertain model where the point correspondences between the initial and desired views and the camera intrinsic parameters are affected by unknown random uncertainties with known bounds. The difficulty we have to face is that traditional path-planning schemes applied to different models lead to different paths rather than to a common and robust path. To solve this problem we propose a technique based on polynomial optimization where the required constraints are imposed on a number of trajectories corresponding to admissible camera poses and parameterized by a common design variable. The planned image trajectory is then followed by using an IBVS controller. Simulations carried out with all typical uncertainties that characterize a real experiment illustrate the proposed strategy and provide promising results.

Visual servoing consists of positioning a robot end-effector based on the matching of some object features in the image. However, due to the presence of image noise, this matching can never be ensured, hence introducing an error on the final location of the robot. This paper addresses the problem of estimating the worst-case location error introduced by image points matching. In particular, we propose some strategies for computing upper bounds and lower bounds of such an error according to several possible measures for certain image noise intensity and camera-object configuration. These bounds provide an admissible region of the sought worst-case location error, and hence allow one to establish performance limitation of visual servo systems. Some examples are reported to illustrate the proposed strategies and their results.

Genetic networks with SUM regulatory functions are a fundamental class of models studied in systems biology. A primary issue for these models consists of establishing the number of the equilibrium points and their location. Unfortunately, this is a difficult problem, indeed existing methods very often do not allow one to solve it. This paper proposes a study of this problem, and describes an approach that exploits the properties of SUM regulatory functions in order to correctly characterize these points of interest. This is verified by some numerical examples, which illustrate the proposed solution and show the advantages with respect to existing methods.

This paper proposes a new approach to estimate the camera displacement of stereo vision systems via minimization of the algebraic error over the essential matrices manifold. The proposed approach is based on the use of homogeneous forms and linear matrix inequality (LMI) optimizations, and has the advantages of not presenting local minima and not introducing approximations of nonlinear terms. Numerical investigations carried out with both synthetic and real data show that the proposed approach provides significantly better results than SVD methods as well as minimizations of the algebraic error over the essential matrices manifold via both gradient descent and simplex search algorithms.

Path planning is a useful technique for visual servoing as it allows one to take into account system constraints and achieve desired performances during the camera motion. In this paper, we propose a new framework for path planning based on the use of homogeneous forms and linear matrix inequalities (LMIs). Specifically, we introduce a general parametrization of the trajectories from the initial to the desired location based on homogeneous forms and a parameter-dependent version of the Rodrigues formula. This allows us to impose typical constraints (field of view, workspace, joint, avoidance of collision, and occlusion) via positivity conditions on suitable homogeneous forms. Then, we reformulate the problem of finding a trajectory in the 3-D space satisfying all these constraints as an LMI optimization that can handle the maximization of typical performances (e.g., visibility margin, similarity to a straight line). The planned camera path is tracked by using an image-based controller. The proposed approach is illustrated and validated through simulations and experiments.

This paper considers the problem of realizing visual servoing taking into account constraints such as visibility and workspace constraints while minimizing a cost function such as spanned image area and trajectory length. A new path-planning scheme is proposed by, first, introducing a robust object reconstruction which allows one to obtain feasible image trajectories. Second, the rotation path is parameterized through a particular extension of the Euler parameters in order to obtain an equivalent expression of the rotation matrix as a quadratic function of unconstrained variables, hence largely simplified with respect to standard parameterizations which involve transcendental functions. Then, polynomials of arbitrary degree are used to complete the parametrization and formulate a general optimization where a number of constraints and costs can be considered. The optimal trajectory is followed by tracking the image trajectories with standard IBVS controllers.

The problem of evaluating worst-case camera positioning error induced by unknown-but-bounded (UBB) image noise for a given object-camera configuration is considered. Specifically, it is shown that upper bounds to the rotation and translation worst-case error for a certain image noise intensity can be obtained through convex optimizations. These upper bounds, contrary to lower bounds provided by standard optimization tools, allow one to design robust visual servo systems.

Visual servoing consists of steering a robot from an initial to a desired location by exploiting the information provided by visual sensors. This paper deals with the problem of realizing visual servoing for robot manipulators taking into account constraints such as visibility, workspace (that is obstacle avoidance), and joint constraints, while minimizing a cost function such as spanned image area, trajectory length, and curvature. To solve this problem, a new path-planning scheme is proposed. First, a robust object reconstruction is computed from visual measurements which allows one to obtain feasible image trajectories. Second, the rotation path is parameterized through an extension of the Euler parameters that yields an equivalent expression of the rotation matrix as a quadratic function of unconstrained variables, hence, largely simplifying standard parameterizations which involve transcendental functions. Then, polynomials of arbitrary degree are used to complete the parametrization and formulate the desired constraints and costs as a general optimization problem. The optimal trajectory is followed by tracking the image trajectory with an IBVS controller combined with repulsive potential fields in order to fulfill the constraints in real conditions.

The problem of establishing if it is possible to reach the desired location keeping all features in the field of view and following a straight line is considered. The proposed approach is based on the polynomial parameterization of the camera path and allows one to find the path that follows a straight line and maximizes the distance of the image trajectories from the screen boundary.

The problem of estimating the camera pose in eye-in-hand visual servoing is considered for the case in which the available point correspondences are less than eight. Two strategies based on the properties of the essential matrix are hence described, which mainly consist of the computation of the roots of a one-variable polynomial in the seven points case and the minimization of a two-variables polynomial through a gradient algorithm in the six points case.

A new visual servoing technique consisting of generating circular-like trajectories is proposed which does not require either geometrical models of the object or points depth. For calibrated camera, the object is kept in the field of view and global stability is achieved. Then, necessary and sufficient conditions are provided for establishing tolerable errors on the estimates of the intrinsic and extrinsic parameters in order to guarantee robust field of view and robust local stability. Simulation results show that the translational trajectories obtained in presence of large displacements are significantly shorter than those produced by one of the best existing methods, in both cases of correct and bad calibration. Very satisfactory results are obtained also in presence of small displacements.

A simple technique for estimating the camera displacement from point correspondences in eye-in-hand visual servoing is presented. The idea for providing more accurate results than existing methods consists of taking into account that the point correspondences used during the camera motion correspond to stationary spatial points, hence exploiting additional information. This is done by first estimating the object Euclidean structure and then estimating the camera displacement from this estimate.

A visual servoing strategy for keeping features in the field of view is proposed which consists of a switching among position-based control strategies and backward motion. In the absence of uncertainty on the extrinsic parameters, all features are kept in the field of view. Moreover, if the intrinsic parameters are also known, the trajectory length is minimized in the rotational space and, for some cases, also minimized in the translational space. Simulation results also show a certain degree of robustness against uncertainty on the intrinsic parameters.

The first aim of any visual-servoing strategy is to avoid features being lost from the field of view and that the desired location may not be reached. However, avoiding both these system failures turns out to be very difficult, especially when the initial and desired locations are distant. Moreover, the methods that succeed in presence of large camera displacements often produce a long translational trajectory that may not be allowed by the robot workspace and/or joint limits. In this paper, a new strategy for dealing with such problems is proposed, which consists of generating circular-like trajectories that may satisfy the task requirements more naturally than other solutions. Knowledge of geometrical models of the object or points depth is not required. It is shown that system failures are avoided for a calibrated camera. Moreover, necessary and sufficient conditions are provided for establishing tolerable errors on the estimates of the intrinsic and extrinsic parameters, in order to guarantee a robust field of view and robust local asymptotic stability. Several simulation results show that the translational trajectories obtained in presence of large displacements are significantly shorter than those produced by the existing methods, in cases of both correct and bad camera calibration. Very satisfactory results are achieved also in presence of small displacements.

The problem of estimating the camera displacement in eye-in-hand visual servoing is considered, and a simple strategy based on the idea that the estimates accuracy can be improved if the fact that the point correspondences used throughout the visual servoing are relative to the same 3D points is taken into account is presented. In particular, an accurate scaled euclidean reconstruction of the object is built in the first steps of the visual servoing by suitably using existing linear methods, and from this reconstruction the camera displacement is suitably estimated. Extensive proves performed in random conditions have shown that the proposed approach provides significantly better results with respect to the existing linear methods actually used in visual servoing.

In this paper, a visual servoing strategy for dealing with the problem of keeping the observed points in the camera field of view is proposed. The approach consists of a switching control law based on camera displacement estimation and regulated from the position of the points in the image. In absence of uncertainties on the intrinsic parameters and optical axis direction, global stability is achieved and all points are kept in the field of view. Moreover, the trajectory length is minimized in the rotational space and, for some cases, also minimized in the translational one. Robustness against uncertainties is also guaranteed.

In this paper, a new method for the estimation of the fundamental matrix from point correspondences in stereo vision is presented. The minimization of the algebraic error is performed while taking explicitly into account the rank-two constraint on the fundamental matrix. It is shown how this nonconvex optimization problem can be solved avoiding local minima by using recently developed convexification techniques. The obtained estimate of the fundamental matrix turns out to be more accurate than the one provided by the linear criterion, where the rank constraint of the matrix is imposed after its computation by setting the smallest singular value to zero. This suggests that the proposed estimate can be used to initialize nonlinear criteria, such as the distance to epipolar lines and the gradient criterion, in order to obtain a more accurate estimate of the fundamental matrix.

A visual servoing algorithm for mobile robots is proposed. The main feature of the algorithm is that it exploits object profiles rather than solving correspondence problems using object features or texture. This property is crucial for mobile robot navigation in unstructured environments where the 3D scene exhibits only surfaces whose main features are their apparent contours. The framework is based on the epipolar geometry, which is recovered from object profiles and epipolar tangencies. Special symmetry conditions of epipoles are used to generate the mobile robot control law. For the sake of simplicity, mobile robot kinematics is assumed to be holonomic and the camera intrinsic parameters are assumed partially known. Such assumption can be relaxed to extend the application field of the approach.

We present a complete system for segmenting, matching and tracking planar contours for use in visual servoing. Our system can be used with arbitrary contours of any shape and without any prior knowledge of their models. The system is first shown the target view. A selected contour is automatically extracted and its image shape is stored. The robot and object are then moved and the system automatically identifies the target. The matching step is done together with the estimation of the homography matrix between the two views of the contour. Then, a 2 1/2 D visual servoing technique is used to reposition the end-effector of a robot at the target position relative to the planar contour. The system has been successfully tested on several contours with very complex shapes such as leaves, keys and the coastal outlines of islands.

In this paper we describe a method to compute the collineation matrix between two unmatched images of an unknown planar contour described using a B-spline snake. The two images of the contour are matched and the collineation matrix is used to servo a camera mounted on the robot end-effector using a 2 1/2 D visual servoing technique. The experimental results, obtained using common planar objects, show that our method give very good results and allow the robot end-effector to be positioned with a great precision. 1 Introduction The visual servoing scheme of robot manipulators can be divided in three steps. In the first off-line learning step, the reference image of the object corresponding to a desired position of the robot is acquired and some image features are extracted. In general, objects are represented by free-form curves, i.e., arbitrary space curves of the type found in practice. A curve is usually described as a set of chained points. The reference image can be obtaine...