arXiv subjects

Hao Luo

Publications and source records attributed to Hao Luo.

At least 127 records · Page 7Linked to original sources

arXiv2023

Revisiting Vision Transformer from the View of Path Ensemble

Vision Transformers (ViTs) are normally regarded as a stack of transformer layers. In this work, we propose a novel view of ViTs showing that they can be seen as ensemble networks containing multiple parallel paths with different lengths. Specifically, we equivalently transform the traditional cascade of multi-head self-attention (MSA) and feed-forward network (FFN) into three parallel paths in each transformer layer. Then, we utilize the identity connection in our new transformer form and further transform the ViT into an explicit multi-path ensemble network. From the new perspective, these paths perform two functions: the first is to provide the feature for the classifier directly, and the second is to provide the lower-level feature representation for subsequent longer paths. We investigate the influence of each path for the final prediction and discover that some paths even pull down the performance. Therefore, we propose the path pruning and EnsembleScale skills for improvement, which cut out the underperforming paths and re-weight the ensemble components, respectively, to optimize the path combination and make the short paths focus on providing high-quality representation for subsequent paths. We also demonstrate that our path combination strategies can help ViTs go deeper and act as high-pass filters to filter out partial low-frequency signals. To further enhance the representation of paths served for subsequent paths, self-distillation is applied to transfer knowledge from the long paths to the short paths. This work calls for more future research to explain and design ViTs from new perspectives.

cs.CV↗

arXiv2023

A unified differential equation solver approach for separable convex optimization: splitting, acceleration and nonergodic rate

This paper provides a self-contained ordinary differential equation solver approach for separable convex optimization problems. A novel primal-dual dynamical system with built-in time rescaling factors is introduced, and the exponential decay of a tailored Lyapunov function is established. Then several time discretizations of the continuous model are considered and analyzed via a unified discrete Lyapunov function. Moreover, two families of accelerated proximal alternating direction methods of multipliers are obtained, and nonergodic optimal mixed-type convergence rates shall be proved for the primal objective residual, the feasibility violation and the Lagrangian gap. Finally, numerical experiments are provided to validate the practical performances.

math.OC↗

arXiv2023

MimCo: Masked Image Modeling Pre-training with Contrastive Teacher

Recent masked image modeling (MIM) has received much attention in self-supervised learning (SSL), which requires the target model to recover the masked part of the input image. Although MIM-based pre-training methods achieve new state-of-the-art performance when transferred to many downstream tasks, the visualizations show that the learned representations are less separable, especially compared to those based on contrastive learning pre-training. This inspires us to think whether the linear separability of MIM pre-trained representation can be further improved, thereby improving the pre-training performance. Since MIM and contrastive learning tend to utilize different data augmentations and training strategies, combining these two pretext tasks is not trivial. In this work, we propose a novel and flexible pre-training framework, named MimCo, which combines MIM and contrastive learning through two-stage pre-training. Specifically, MimCo takes a pre-trained contrastive learning model as the teacher model and is pre-trained with two types of learning targets: patch-level and image-level reconstruction losses. Extensive transfer experiments on downstream tasks demonstrate the superior performance of our MimCo pre-training framework. Taking ViT-S as an example, when using the pre-trained MoCov3-ViT-S as the teacher model, MimCo only needs 100 epochs of pre-training to achieve 82.53% top-1 finetuning accuracy on Imagenet-1K, which outperforms the state-of-the-art self-supervised learning counterparts.

cs.CV↗

arXiv2023

Revisit Parameter-Efficient Transfer Learning: A Two-Stage Paradigm

Parameter-Efficient Transfer Learning (PETL) aims at efficiently adapting large models pre-trained on massive data to downstream tasks with limited task-specific data. In view of the practicality of PETL, previous works focus on tuning a small set of parameters for each downstream task in an end-to-end manner while rarely considering the task distribution shift issue between the pre-training task and the downstream task. This paper proposes a novel two-stage paradigm, where the pre-trained model is first aligned to the target distribution. Then the task-relevant information is leveraged for effective adaptation. Specifically, the first stage narrows the task distribution shift by tuning the scale and shift in the LayerNorm layers. In the second stage, to efficiently learn the task-relevant information, we propose a Taylor expansion-based importance score to identify task-relevant channels for the downstream task and then only tune such a small portion of channels, making the adaptation to be parameter-efficient. Overall, we present a promising new direction for PETL, and the proposed paradigm achieves state-of-the-art performance on the average accuracy of 19 downstream tasks.

cs.CV↗

arXiv2023

MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID

Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both in-domain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINet.

cs.CV↗

arXiv2023

Model-Based Decentralized Policy Optimization

Decentralized policy optimization has been commonly used in cooperative multi-agent tasks. However, since all agents are updating their policies simultaneously, from the perspective of individual agents, the environment is non-stationary, resulting in it being hard to guarantee monotonic policy improvement. To help the policy improvement be stable and monotonic, we propose model-based decentralized policy optimization (MDPO), which incorporates a latent variable function to help construct the transition and reward function from an individual perspective. We theoretically analyze that the policy optimization of MDPO is more stable than model-free decentralized policy optimization. Moreover, due to non-stationarity, the latent variable function is varying and hard to be modeled. We further propose a latent variable prediction method to reduce the error of the latent variable function, which theoretically contributes to the monotonic policy improvement. Empirically, MDPO can indeed obtain superior performance than model-free decentralized policy optimization in a variety of cooperative multi-agent tasks.

cs.LG↗

arXiv2022

A catalog of quasar candidates identified by astrometric and mid-infrared methods in Gaia EDR3

Quasars are very important in materializing the reference frame. The excess emission of AGNs (active galactic nuclei) in the mid infrared band can be used to identify quasar candidates. As extremely distant and point like objects, quasars also could be further selected by astrometry method. Increasing the number of reliable quasar candidates is necessary in characterizing the properties of Gaia astrometric solution and evaluating the reliability of Gaia own quasars classification. We identify quasars by using appropriate AllWISE [W1-W2] color and different combinations of astrometric criteria. Together with the contamination and completeness, the magnitude, astrometric properties, density distribution, and the morphological indexes of these selected quasars are evaluated. We obtain a quasar candidate catalog of 1,503,373 sources, which contains 1,186,690 candidates (78.9%) in common with the Gaia EDR3_AGN catalog and 316,683 newly identified quasar candidates. The completeness of this catalog is around 80% compared to LQAC5, and the purity of the overall catalog is about 90%. We also found that the purity of quasar candidates selected by this method will decrease in the crowded sky area and the region with less WISE observations.

astro-ph.GA↗

arXiv2022

Reconfigurable Intelligent Surface Aided Wireless Sensing for Scene Depth Estimation

Current scene depth estimation approaches mainly rely on optical sensing, which carries privacy concerns and suffers from estimation ambiguity for distant, shiny, and transparent surfaces/objects. Reconfigurable intelligent surfaces (RISs) provide a path for employing a massive number of antennas using low-cost and energy-efficient architectures. This has the potential for realizing RIS-aided wireless sensing with high spatial resolution. In this paper, we propose to employ RIS-aided wireless sensing systems for scene depth estimation. We develop a comprehensive framework for building accurate depth maps using RIS-aided mmWave sensing systems. In this framework, we propose a new RIS interaction codebook capable of creating a sensing grid of reflected beams that meets the desirable characteristics of efficient scene depth map construction. Using the designed codebook, the received signals are processed to build high-resolution depth maps. Simulation results compare the proposed solution against RGB-based approaches and highlight the promise of adopting RIS-aided mmWave sensing in scene depth perception.

eess.SP↗

arXiv2022

A universal accelerated primal-dual method for convex optimization problems

This work presents a universal accelerated first-order primal-dual method for affinely constrained convex optimization problems. It can handle both Lipschitz and Hölder gradients but does not need to know the smoothness level of the objective function. In line search part, it uses dynamically decreasing parameters and produces approximate Lipschitz constant with moderate magnitude. In addition, based on a suitable discrete Lyapunov function and tight decay estimates of some differential/difference inequalities, a universal optimal mixed-type convergence rate is established. Some numerical tests are provided to confirm the efficiency of the proposed method.

math.OC↗

arXiv2022

An efficient semismooth Newton-AMG-based inexact primal-dual algorithm for generalized transport problems

This work is concerned with the efficient optimization method for solving a large class of optimal mass transport problems. An inexact primal-dual algorithm is presented from the time discretization of a proper dynamical system, and by using the tool of Lyapunov function, the global (super-)linear convergence rate is established for function residual and feasibility violation. The proposed algorithm contains an inner problem that possesses strong semismoothness property and motivates the use of the semismooth Newton iteration. By exploring the hidden structure of the problem itself, the linear system arising from the Newton iteration is transferred equivalently into a graph Laplacian system, for which a robust algebraic multigrid method is proposed and also analyzed via the famous Xu--Zikatanov identity. Finally, numerical experiments are provided to validate the efficiency of our method.

math.OC↗

arXiv2022

Dynamic Gradient Reactivation for Backward Compatible Person Re-identification

We study the backward compatible problem for person re-identification (Re-ID), which aims to constrain the features of an updated new model to be comparable with the existing features from the old model in galleries. Most of the existing works adopt distillation-based methods, which focus on pushing new features to imitate the distribution of the old ones. However, the distillation-based methods are intrinsically sub-optimal since it forces the new feature space to imitate the inferior old feature space. To address this issue, we propose the Ranking-based Backward Compatible Learning (RBCL), which directly optimizes the ranking metric between new features and old features. Different from previous methods, RBCL only pushes the new features to find best-ranking positions in the old feature space instead of strictly alignment, and is in line with the ultimate goal of backward retrieval. However, the sharp sigmoid function used to make the ranking metric differentiable also incurs the gradient vanish issue, therefore stems the ranking refinement during the later period of training. To address this issue, we propose the Dynamic Gradient Reactivation (DGR), which can reactivate the suppressed gradients by adding dynamic computed constant during forward step. To further help targeting the best-ranking positions, we include the Neighbor Context Agents (NCAs) to approximate the entire old feature space during training. Unlike previous works which only test on the in-domain settings, we make the first attempt to introduce the cross-domain settings (including both supervised and unsupervised), which are more meaningful and difficult. The experimental results on all five settings show that the proposed RBCL outperforms previous state-of-the-art methods by large margins under all settings.

cs.CV↗

arXiv2022

Nitrogen decoration of basal plane dislocations in 4H-SiC

Basal-plane dislocations (BPDs) pose a great challenge to the reliability of bipolar power devices based on the 4H silicon carbide (4H-SiC). It is well established that heavy nitrogen (N) doping promotes the conversion of BPDs to threading edge dislocations (TEDs) and improves the reliability of 4H-SiC-based bipolar power devices. However, the interaction between N and BPDs, and the effect of N on the electronic properties of BPDs are still ambiguous, which significantly hinder the understanding on the electron-transport mechanism of 4H-SiC-based bipolar power devices. Combining molten-alkali etching and the Kelvin probe force microscopy (KPFM) analysis, we demonstrate that BPDs create acceptor-like states in undoped 4H-SiC, while acting as donors in N-doped 4H-SiC. First-principles calculations verify that BPDs create occupied defect states above the valence band maximum (VBM) and unoccupied defect states under the conduction-band minimum (CBM) of undoped 4H-SiC. The electron transfer from the defect states of intrinsic defects and native impurities to the unoccupied defect states of BPDs gives rise to the acceptor-like behavior of BPDs in undoped 4H-SiC. Defect formation energies indicate that N atoms can spontaneously decorate BPDs during the N doping of 4H-SiC. The binding between N and BPD is strong against decomposition. The accumulation of N dopants at the core of BPDs results in the accumulation of donor-like states at the core of BPDs in N-doped 4H-SiC. This work not only enriches the understanding on the electronic behavior of BPDs in N-doped 4H-SiC, but also helps understand the electron transport mechanism of 4H-SiC-based bipolar power devices.

cond-mat.mtrl-sci↗

arXiv2022

Accelerated primal-dual methods for linearly constrained convex optimization problems

This work proposes an accelerated primal-dual dynamical system for affine constrained convex optimization and presents a class of primal-dual methods with nonergodic convergence rates. In continuous level, exponential decay of a novel Lyapunov function is established and in discrete level, implicit, semi-implicit and explicit numerical discretizations for the continuous model are considered sequentially and lead to new accelerated primal-dual methods for solving linearly constrained optimization problems. Special structures of the subproblems in those schemes are utilized to develop efficient inner solvers. In addition, nonergodic convergence rates in terms of primal-dual gap, primal objective residual and feasibility violation are proved via a tailored discrete Lyapunov function. Moreover, our method has also been applied to decentralized distributed optimization for fast and efficient solution.

math.OC↗

arXiv2022

A primal-dual flow for affine constrained convex optimization

We introduce a novel primal-dual flow for affine constrained convex optimization problems. As a modification of the standard saddle-point system, our primal-dual flow is proved to possess the exponential decay property, in terms of a tailored Lyapunov function. Then two primal-dual methods are obtained from numerical discretizations of the continuous model, and global nonergodic linear convergence rate is established via a discrete Lyapunov function. Instead of solving the subproblem of the primal variable, we apply the semi-smooth Newton iteration to the subproblem with respect to the multiplier, provided that there are some additional properties such as semi-smoothness and sparsity. Especially, numerical tests on the linearly constrained $l_1$-$l_2$ minimization and the total-variation based image denoising model have been provided.

math.OC↗

arXiv2022

Accelerated differential inclusion for convex optimization

This paper introduces a second-order differential inclusion for unconstrained convex optimization. In continuous level, solution existence in proper sense is obtained and exponential decay of a novel Lyapunov function along with the solution trajectory is derived as well. Then in discrete level, based on numerical discretizations of the continuous model, two inexact proximal point algorithms are proposed, and some new convergence rates are established via a discrete Lyapunov function.

math.OC↗

arXiv2022

From differential equation solvers to accelerated first-order methods for convex optimization

Convergence analysis of accelerated first-order methods for convex optimization problems are presented from the point of view of ordinary differential equation solvers. A new dynamical system, called Nesterov accelerated gradient flow, has been derived from the connection between acceleration mechanism and $A$-stability of ODE solvers, and the exponential decay of a tailored Lyapunov function along with the solution trajectory is proved. Numerical discretizations are then considered and convergence rates are established via a unified discrete Lyapunov function. The proposed differential equation solver approach can not only cover existing accelerated methods, such as FISTA, Güler's proximal algorithm and Nesterov's accelerated gradient method, but also produce new algorithms for composite convex optimization that possess accelerated convergence rates.

math.OC↗

arXiv2022

Error estimation of a discontinuous Galerkin method for time fractional subdiffusion problems with nonsmooth data

This paper is devoted to the numerical analysis of a piecewise constant discontinuous Galerkin method for time fractional subdiffusion problems. The regularity of weak solution is firstly established by using variational approach and Mittag-Leffler function. Then several optimal error estimates are derived with low regularity data. Finally, numerical experiments are conducted to verify the theoretical results.

math.NA↗

arXiv2022

Scaled ReLU Matters for Training Vision Transformers

Vision transformers (ViTs) have been an alternative design paradigm to convolutional neural networks (CNNs). However, the training of ViTs is much harder than CNNs, as it is sensitive to the training parameters, such as learning rate, optimizer and warmup epoch. The reasons for training difficulty are empirically analysed in ~\cite{xiao2021early}, and the authors conjecture that the issue lies with the \textit{patchify-stem} of ViT models and propose that early convolutions help transformers see better. In this paper, we further investigate this problem and extend the above conclusion: only early convolutions do not help for stable training, but the scaled ReLU operation in the \textit{convolutional stem} (\textit{conv-stem}) matters. We verify, both theoretically and empirically, that scaled ReLU in \textit{conv-stem} not only improves training stabilization, but also increases the diversity of patch tokens, thus boosting peak performance with a large margin via adding few parameters and flops. In addition, extensive experiments are conducted to demonstrate that previous ViTs are far from being well trained, further showing that ViTs have great potential to be a better substitute of CNNs.

cs.CV↗