Weizhong Zhang

Xiao Zhou, Yong Lin, Renjie Pi et al.

Distributionally robust optimization (DRO) and invariant risk minimization (IRM) are two popular methods proposed to improve out-of-distribution (OOD) generalization performance of machine learning models. While effective for small models, it has been observed that these methods can be vulnerable to overfitting with large overparameterized models. This work proposes a principled method, \textbf{M}odel \textbf{A}gnostic sam\textbf{PL}e r\textbf{E}weighting (\textbf{MAPLE}), to effectively address OOD problem, especially in overparameterized scenarios. Our key idea is to find an effective reweighting of the training samples so that the standard empirical risk minimization training of a large model on the weighted training data leads to superior OOD generalization performance. The overfitting issue is addressed by considering a bilevel formulation to search for the sample reweighting, in which the generalization complexity depends on the search space of sample weights instead of the model size. We present theoretical analysis in linear case to prove the insensitivity of MAPLE to model size, and empirically verify its superiority in surpassing state-of-the-art methods by a large margin. Code is available at \url{https://github.com/x-zho14/MAPLE}.

Zhichao Huang, Yanbo Fan, Chen Liu et al.

While adversarial training and its variants have shown to be the most effective algorithms to defend against adversarial attacks, their extremely slow training process makes it hard to scale to large datasets like ImageNet. The key idea of recent works to accelerate adversarial training is to substitute multi-step attacks (e.g., PGD) with single-step attacks (e.g., FGSM). However, these single-step methods suffer from catastrophic overfitting, where the accuracy against PGD attack suddenly drops to nearly 0% during training, destroying the robustness of the networks. In this work, we study the phenomenon from the perspective of training instances. We show that catastrophic overfitting is instance-dependent and fitting instances with larger gradient norm is more likely to cause catastrophic overfitting. Based on our findings, we propose a simple but effective method, Adversarial Training with Adaptive Step size (ATAS). ATAS learns an instancewise adaptive step size that is inversely proportional to its gradient norm. The theoretical analysis shows that ATAS converges faster than the commonly adopted non-adaptive counterparts. Empirically, ATAS consistently mitigates catastrophic overfitting and achieves higher robust accuracy on CIFAR10, CIFAR100 and ImageNet when evaluated on various adversarial budgets.

Xiao Zhou, Renjie Pi, Weizhong Zhang et al.

The goal of coreset selection in supervised learning is to produce a weighted subset of data, so that training only on the subset achieves similar performance as training on the entire dataset. Existing methods achieved promising results in resource-constrained scenarios such as continual learning and streaming. However, most of the existing algorithms are limited to traditional machine learning models. A few algorithms that can handle large models adopt greedy search approaches due to the difficulty in solving the discrete subset selection problem, which is computationally costly when coreset becomes larger and often produces suboptimal results. In this work, for the first time we propose a continuous probabilistic bilevel formulation of coreset selection by learning a probablistic weight for each training sample. The overall objective is posed as a bilevel optimization problem, where 1) the inner loop samples coresets and train the model to convergence and 2) the outer loop updates the sample probability progressively according to the model's performance. Importantly, we develop an efficient solver to the bilevel optimization problem via unbiased policy gradient without trouble of implicit differentiation. We provide the convergence property of our training procedure and demonstrate the superiority of our algorithm against various coreset selection methods in various tasks, especially in more challenging label-noise and class-imbalance scenarios.

Yong Lin, Lu Tan, Yifan Hao et al.

Generalization to out-of-distribution (OOD) data is a critical challenge in machine learning. Ensemble-based methods, like weight space ensembles that interpolate model parameters, have been shown to achieve superior OOD performance. However, the underlying mechanism for their effectiveness remains unclear. In this study, we closely examine WiSE-FT, a popular weight space ensemble method that interpolates between a pre-trained and a fine-tuned model. We observe an unexpected ``FalseFalseTrue" phenomenon, in which WiSE-FT successfully corrects many cases where each individual model makes incorrect predictions, which contributes significantly to its OOD effectiveness. To gain further insights, we conduct theoretical analysis in a multi-class setting with a large number of spurious features. Our analysis predicts the above phenomenon and it further shows that ensemble-based models reduce prediction errors in the OOD settings by utilizing a more diverse set of spurious features. Contrary to the conventional wisdom that focuses on learning invariant features for better OOD performance, our findings suggest that incorporating a large number of diverse spurious features weakens their individual contributions, leading to improved overall OOD generalization performance. Additionally, our findings provide the first explanation for the mysterious phenomenon of weight space ensembles outperforming output space ensembles in OOD. Empirically we demonstrate the effectiveness of utilizing diverse spurious features on a MultiColorMNIST dataset, and our experimental results are consistent with the theoretical analysis. Building upon the new theoretical insights into the efficacy of ensemble methods, we further propose a novel averaging method called BAlaNced averaGing (BANG) which significantly enhances the OOD performance of WiSE-FT.

Yueqi Xie, Weizhong Zhang, Renjie Pi et al.

Non-IID data distribution across clients and poisoning attacks are two main challenges in real-world federated learning (FL) systems. While both of them have attracted great research interest with specific strategies developed, no known solution manages to address them in a unified framework. To universally overcome both challenges, we propose SmartFL, a generic approach that optimizes the server-side aggregation process with a small amount of proxy data collected by the service provider itself via a subspace training technique. Specifically, the aggregation weight of each participating client at each round is optimized using the server-collected proxy data, which is essentially the optimization of the global model in the convex hull spanned by client models. Since at each round, the number of tunable parameters optimized on the server side equals the number of participating clients (thus independent of the model size), we are able to train a global model with massive parameters using only a small amount of proxy data (e.g., around one hundred samples). With optimized aggregation, SmartFL ensures robustness against both heterogeneous and malicious clients, which is desirable in real-world FL where either or both problems may occur. We provide theoretical analyses of the convergence and generalization capacity for SmartFL. Empirically, SmartFL achieves state-of-the-art performance on both FL with non-IID data distribution and FL with malicious clients. The source code will be released.

Jipeng Zhang, Yaxuan Qin, Renjie Pi et al.

Instruction tuning has achieved unprecedented success in NLP, turning large language models into versatile chatbots. However, the increasing variety and volume of instruction datasets demand significant computational resources. To address this, it is essential to extract a small and highly informative subset (i.e., Coreset) that achieves comparable performance to the full dataset. Achieving this goal poses non-trivial challenges: 1) data selection requires accurate data representations that reflect the training samples' quality, 2) considering the diverse nature of instruction datasets, and 3) ensuring the efficiency of the coreset selection algorithm for large models. To address these challenges, we propose Task-Agnostic Gradient Clustered COreset Selection (TAGCOS). Specifically, we leverage sample gradients as the data representations, perform clustering to group similar data, and apply an efficient greedy algorithm for coreset selection. Experimental results show that our algorithm, selecting only 5% of the data, surpasses other unsupervised methods and achieves performance close to that of the full dataset.

Yibin Wang, Weizhong Zhang, Jianwei Zheng et al.

Current subject-driven image generation methods encounter significant challenges in person-centric image generation. The reason is that they learn the semantic scene and person generation by fine-tuning a common pre-trained diffusion, which involves an irreconcilable training imbalance. Precisely, to generate realistic persons, they need to sufficiently tune the pre-trained model, which inevitably causes the model to forget the rich semantic scene prior and makes scene generation over-fit to the training data. Moreover, even with sufficient fine-tuning, these methods can still not generate high-fidelity persons since joint learning of the scene and person generation also lead to quality compromise. In this paper, we propose Face-diffuser, an effective collaborative generation pipeline to eliminate the above training imbalance and quality compromise. Specifically, we first develop two specialized pre-trained diffusion models, i.e., Text-driven Diffusion Model (TDM) and Subject-augmented Diffusion Model (SDM), for scene and person generation, respectively. The sampling process is divided into three sequential stages, i.e., semantic scene construction, subject-scene fusion, and subject enhancement. The first and last stages are performed by TDM and SDM respectively. The subject-scene fusion stage, that is the collaboration achieved through a novel and highly effective mechanism, Saliency-adaptive Noise Fusion (SNF). Specifically, it is based on our key observation that there exists a robust link between classifier-free guidance responses and the saliency of generated images. In each time step, SNF leverages the unique strengths of each model and allows for the spatial blending of predicted noises from both models automatically in a saliency-aware manner. Extensive experiments confirm the impressive effectiveness and robustness of the Face-diffuser.

Yuan Gao, Zujing Liu, Weizhong Zhang et al.

Recent Large-Language Models (LLMs) pruning methods typically operate at the post-training phase without the expensive weight finetuning, however, their pruning criteria often rely on heuristically hand-crafted metrics, potentially leading to suboptimal performance. We instead propose a novel optimization-based structural pruning that learns the pruning masks in a probabilistic space directly by optimizing the loss of the pruned model. To preserve efficiency, our method eliminates the back-propagation through the LLM per se during optimization, requiring only the forward pass of the LLM. We achieve this by learning an underlying Bernoulli distribution to sample binary pruning masks, where we decouple the Bernoulli parameters from LLM loss, facilitating efficient optimization via policy gradient estimator without back-propagation. Thus, our method can 1) support global and heterogeneous pruning (i.e., automatically determine different redundancy for different layers), and 2) optionally initialize with a metric-based method (for our Bernoulli distributions). Extensive experiments conducted on LLaMA, LLaMA-2, LLaMA-3, Vicuna, and Mistral models using the C4 and WikiText2 datasets demonstrate the promising performance of our method in efficiency and effectiveness. Code is available at https://github.com/ethanygao/backprop-free_LLM_pruning.

Zhiwei Zhang, Weizhong Zhang, Yaowei Huang et al.

Traffic forecasting is a challenging task due to the complex spatio-temporal correlations among traffic series. In this paper, we identify an underexplored problem in multivariate traffic series prediction: extreme events. Road congestion and rush hours can result in low correlation in vehicle speeds at various intersections during adjacent time periods. Existing methods generally predict future series based on recent observations and entirely discard training data during the testing phase, rendering them unreliable for forecasting highly nonlinear multivariate time series. To tackle this issue, we propose a test-time compensated representation learning framework comprising a spatio-temporal decomposed data bank and a multi-head spatial transformer model (CompFormer). The former component explicitly separates all training data along the temporal dimension according to periodicity characteristics, while the latter component establishes a connection between recent observations and historical series in the data bank through a spatial attention matrix. This enables the CompFormer to transfer robust features to overcome anomalous events while using fewer computational resources. Our modules can be flexibly integrated with existing forecasting methods through end-to-end training, and we demonstrate their effectiveness on the METR-LA and PEMS-BAY benchmarks. Extensive experimental results show that our method is particularly important in extreme events, and can achieve significant improvements over six strong baselines, with an overall improvement of up to 28.2%.

Yibin Wang, Weizhong Zhang, Jianwei Zheng et al.

Image composition involves seamlessly integrating given objects into a specific visual context. Current training-free methods rely on composing attention weights from several samplers to guide the generator. However, since these weights are derived from disparate contexts, their combination leads to coherence confusion and loss of appearance information. These issues worsen with their excessive focus on background generation, even when unnecessary in this task. This not only impedes their swift implementation but also compromises foreground generation quality. Moreover, these methods introduce unwanted artifacts in the transition area. In this paper, we formulate image composition as a subject-based local editing task, solely focusing on foreground generation. At each step, the edited foreground is combined with the noisy background to maintain scene consistency. To address the remaining issues, we propose PrimeComposer, a faster training-free diffuser that composites the images by well-designed attention steering across different noise levels. This steering is predominantly achieved by our Correlation Diffuser, utilizing its self-attention layers at each step. Within these layers, the synthesized subject interacts with both the referenced object and background, capturing intricate details and coherent relationships. This prior information is encoded into the attention weights, which are then integrated into the self-attention layers of the generator to guide the synthesis process. Besides, we introduce a Region-constrained Cross-Attention to confine the impact of specific subject-related tokens to desired regions, addressing the unwanted artifacts shown in the prior method thereby further improving the coherence in the transition area. Our method exhibits the fastest inference efficiency and extensive experiments demonstrate our superiority both qualitatively and quantitatively.

Yibin Wang, Weizhong Zhang, Honghui Xu et al.

Scene text synthesis involves rendering specified texts onto arbitrary images. Current methods typically formulate this task in an end-to-end manner but lack effective character-level guidance during training. Besides, their text encoders, pre-trained on a single font type, struggle to adapt to the diverse font styles encountered in practical applications. Consequently, these methods suffer from character distortion, repetition, and absence, particularly in polystylistic scenarios. To this end, this paper proposes DreamText for high-fidelity scene text synthesis. Our key idea is to reconstruct the diffusion training process, introducing more refined guidance tailored to this task, to expose and rectify the model's attention at the character level and strengthen its learning of text regions. This transformation poses a hybrid optimization challenge, involving both discrete and continuous variables. To effectively tackle this challenge, we employ a heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize the diverse font present in the training dataset. This joint training is seamlessly integrated into the alternate optimization process, fostering a synergistic relationship between learning character embedding and re-estimating character attention. Specifically, in each step, we first encode potential character-generated position information from cross-attention maps into latent character masks. These masks are then utilized to update the representation of specific characters in the current step, which, in turn, enables the generator to correct the character's attention in the subsequent steps. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art.

Kaiyi Zhang, Yang Chen, Ximing Yang et al.

Ideal part editing should guarantee the diversity of edited parts, the fidelity to the remaining parts, and the quality of the results. However, previous methods do not disentangle each part completely, which means the edited parts will affect the others, resulting in poor diversity and fidelity. In addition, some methods lack constraints between parts, which need manual selections of edited results to ensure quality. Therefore, we propose a four-stage process for point cloud part editing: Segmentation, Generation, Assembly, and Selection. Based on this process, we introduce SGAS, a model for part editing that employs two strategies: feature disentanglement and constraint. By independently fitting part-level feature distributions, we realize the feature disentanglement. By explicitly modeling the transformation from object-level distribution to part-level distributions, we realize the feature constraint. Considerable experiments on different datasets demonstrate the efficiency and effectiveness of SGAS on point cloud part editing. In addition, SGAS can be pruned to realize unsupervised part-aware point cloud generation and achieves state-of-the-art results.

Zhibo Zhang, Ximing Yang, Weizhong Zhang et al.

This paper presents a robust fine-tuning method designed for pre-trained 3D point cloud models, to enhance feature robustness in downstream fine-tuned models. We highlight the limitations of current fine-tuning methods and the challenges of learning robust models. The proposed method, named Weight-Space Ensembles for Fine-Tuning then Linear Probing (WiSE-FT-LP), integrates the original pre-training and fine-tuning models through weight space integration followed by Linear Probing. This approach significantly enhances the performance of downstream fine-tuned models under distribution shifts, improving feature robustness while maintaining high performance on the target distribution. We apply this robust fine-tuning method to mainstream 3D point cloud pre-trained models and evaluate the quality of model parameters and the degradation of downstream task performance. Experimental results demonstrate the effectiveness of WiSE-FT-LP in enhancing model robustness, effectively balancing downstream task performance and model feature robustness without altering the model structures.

Zhibo Zhang, Ximing Yang, Weizhong Zhang et al.

Cross-modal knowledge transfer enhances point cloud representation learning in LiDAR semantic segmentation. Despite its potential, the \textit{weak teacher challenge} arises due to repetitive and non-diverse car camera images and sparse, inaccurate ground truth labels. To address this, we propose the Efficient Image-to-LiDAR Knowledge Transfer (ELiTe) paradigm. ELiTe introduces Patch-to-Point Multi-Stage Knowledge Distillation, transferring comprehensive knowledge from the Vision Foundation Model (VFM), extensively trained on diverse open-world images. This enables effective knowledge transfer to a lightweight student model across modalities. ELiTe employs Parameter-Efficient Fine-Tuning to strengthen the VFM teacher and expedite large-scale model training with minimal costs. Additionally, we introduce the Segment Anything Model based Pseudo-Label Generation approach to enhance low-quality image labels, facilitating robust semantic representations. Efficient knowledge transfer in ELiTe yields state-of-the-art results on the SemanticKITTI benchmark, outperforming real-time inference models. Our approach achieves this with significantly fewer parameters, confirming its effectiveness and efficiency.

Weilin Wan, Weizhong Zhang, Quan Zhou et al.

Out-of-distribution detection (OOD) is a crucial technique for deploying machine learning models in the real world to handle the unseen scenarios. In this paper, we first propose a simple yet effective Neural Activation Prior (NAP) for OOD detection. Our neural activation prior is based on a key observation that, for a channel before the global pooling layer of a fully trained neural network, the probability of a few neurons being activated with a large response by an in-distribution (ID) sample is significantly higher than that by an OOD sample. An intuitive explanation is that for a model fully trained on ID dataset, each channel would play a role in detecting a certain pattern in the ID dataset, and a few neurons can be activated with a large response when the pattern is detected in an input sample. Then, a new scoring function based on this prior is proposed to highlight the role of these strongly activated neurons in OOD detection. Our approach is plug-and-play and does not lead to any performance degradation on ID data classification and requires no extra training or statistics from training or external datasets. Notice that previous methods primarily rely on post-global-pooling features of the neural networks, while the within-channel distribution information we leverage would be discarded by the global pooling operator. Consequently, our method is orthogonal to existing approaches and can be effectively combined with them in various applications. Experimental results show that our method achieves the state-of-the-art performance on CIFAR benchmark and ImageNet dataset, which demonstrates the power of the proposed prior. Finally, we extend our method to Transformers and the experimental findings indicate that NAP can also significantly enhance the performance of OOD detection on Transformers, thereby demonstrating the broad applicability of this prior knowledge.

Weilin Wan, Weizhong Zhang, Cheng Jin

Data selection improves computational efficiency by choosing informative subsets of training samples. However, existing methods ignore the compute budget, treating data selection and importance evaluation independently of compute budget constraints. Yet empirical studies show no algorithm can consistently outperform others (or even random selection) across varying budgets. We therefore argue that compute budget must be integral to data-selection strategies, since different budgets impose distinct requirements on data quantity, quality, and distribution for effective training. To this end, we propose a novel Computational budget-Aware Data Selection (CADS) method and naturally formulate it into a bilevel optimization framework, where the inner loop trains the model within the constraints of the computational budget on some selected subset of training data, while the outer loop optimizes data selection based on model evaluation. Our technical contributions lie in addressing two main challenges in solving this bilevel optimization problem: the expensive Hessian matrix estimation for outer-loop gradients and the computational burden of achieving inner-loop optimality during iterations. To solve the first issue, we propose a probabilistic reparameterization strategy and compute the gradient using a Hessian-free policy gradient estimator. To address the second challenge, we transform the inner optimization problem into a penalty term in the outer objective, further discovering that we only need to estimate the minimum of a one-dimensional loss to calculate the gradient, significantly improving efficiency. Extensive experiments show that our method achieves performance gains of up to 14.42% over baselines in vision and language benchmarks.

Xupeng Shi, Pengfei Zheng, A. Adam Ding et al.

Modern deep neural networks (DNNs) are vulnerable to adversarial attacks and adversarial training has been shown to be a promising method for improving the adversarial robustness of DNNs. Pruning methods have been considered in adversarial context to reduce model capacity and improve adversarial robustness simultaneously in training. Existing adversarial pruning methods generally mimic the classical pruning methods for natural training, which follow the three-stage 'training-pruning-fine-tuning' pipelines. We observe that such pruning methods do not necessarily preserve the dynamics of dense networks, making it potentially hard to be fine-tuned to compensate the accuracy degradation in pruning. Based on recent works of \textit{Neural Tangent Kernel} (NTK), we systematically study the dynamics of adversarial training and prove the existence of trainable sparse sub-network at initialization which can be trained to be adversarial robust from scratch. This theoretically verifies the \textit{lottery ticket hypothesis} in adversarial context and we refer such sub-network structure as \textit{Adversarial Winning Ticket} (AWT). We also show empirical evidences that AWT preserves the dynamics of adversarial training and achieve equal performance as dense adversarial training.

Weizhong Zhang, Shuang Qiu

Sparse Conditional Random Field (CRF) is a powerful technique in computer vision and natural language processing for structured prediction. However, solving sparse CRFs in large-scale applications remains challenging. In this paper, we propose a novel safe dynamic screening method that exploits an accurate dual optimum estimation to identify and remove the irrelevant features during the training process. Thus, the problem size can be reduced continuously, leading to great savings in the computational cost without sacrificing any accuracy on the finally learned model. To the best of our knowledge, this is the first screening method which introduces the dual optimum estimation technique -- by carefully exploring and exploiting the strong convexity and the complex structure of the dual problem -- in static screening methods to dynamic screening. In this way, we can absorb the advantages of both the static and dynamic screening methods and avoid their drawbacks. Our estimation would be much more accurate than those developed based on the duality gap, which contributes to a much stronger screening rule. Moreover, our method is also the first screening method in sparse CRFs and even structure prediction models. Experimental results on both synthetic and real-world datasets demonstrate that the speedup gained by our method is significant.

Jiaxin Li, Yan Ding, Weizhong Zhang et al.

Augmented reality technology based on image registration is becoming increasingly popular for the convenience of pre-surgery preparation and medical education. This paper focuses on the registration of forearm images and digital anatomical models. Due to the difference in texture features of forearm multi-modal images, this paper proposes a forearm feature representation curve (FFRC) based on structure compliant multi-modal image registration framework (FAM) for the forearm.

Xiao Zhou, Weizhong Zhang, Hang Xu et al.

Weight pruning is an effective technique to reduce the model size and inference time for deep neural networks in real-world deployments. However, since magnitudes and relative importance of weights are very different for different layers of a neural network, existing methods rely on either manual tuning or handcrafted heuristic rules to find appropriate pruning rates individually for each layer. This approach generally leads to suboptimal performance. In this paper, by directly working on the probability space, we propose an effective network sparsification method called {\it probabilistic masking} (ProbMask), which solves a natural sparsification formulation under global sparsity constraint. The key idea is to use probability as a global criterion for all layers to measure the weight importance. An appealing feature of ProbMask is that the amounts of weight redundancy can be learned automatically via our constraint and thus we avoid the problem of tuning pruning rates individually for different layers in a network. Extensive experimental results on CIFAR-10/100 and ImageNet demonstrate that our method is highly effective, and can outperform previous state-of-the-art methods by a significant margin, especially in the high pruning rate situation. Notably, the gap of Top-1 accuracy between our ProbMask and existing methods can be up to 10\%. As a by-product, we show ProbMask is also highly effective in identifying supermasks, which are subnetworks with high performance in a randomly weighted dense neural network.

Cong Fang, Yihong Gu, Weizhong Zhang et al.

Analysis of over-parameterized neural networks has drawn significant attention in recentyears. It was shown that such systems behave like convex systems under various restrictedsettings, such as for two-level neural networks, and when learning is only restricted locally inthe so-called neural tangent kernel space around specialized initializations. However, there areno theoretical techniques that can analyze fully trained deep neural networks encountered inpractice. This paper solves this fundamental problem by investigating such overparameterizeddeep neural networks when fully trained. We generalize a new technique called neural feature repopulation, originally introduced in (Fang et al., 2019a) for two-level neural networks, to analyze deep neural networks. It is shown that under suitable representations, overparameterized deep neural networks are inherently convex, and when optimized, the system can learn effective features suitable for the underlying learning task under mild conditions. This new analysis is consistent with empirical observations that deep neural networks are capable of learning efficient feature representations. Therefore, the highly unexpected result of this paper can satisfactorily explain the practical success of deep neural networks. Empirical studies confirm that predictions of our theory are consistent with results observed in practice.

Yunzhe Tao, Lin Ma, Weizhong Zhang et al.

Time series prediction has been studied in a variety of domains. However, it is still challenging to predict future series given historical observations and past exogenous data. Existing methods either fail to consider the interactions among different components of exogenous variables which may affect the prediction accuracy, or cannot model the correlations between exogenous data and target data. Besides, the inherent temporal dynamics of exogenous data are also related to the target series prediction, and thus should be considered as well. To address these issues, we propose an end-to-end deep learning model, i.e., Hierarchical attention-based Recurrent Highway Network (HRHN), which incorporates spatio-temporal feature extraction of exogenous variables and temporal dynamics modeling of target variables into a single framework. Moreover, by introducing the hierarchical attention mechanism, HRHN can adaptively select the relevant exogenous features in different semantic levels. We carry out comprehensive empirical evaluations with various methods over several datasets, and show that HRHN outperforms the state of the arts in time series prediction, especially in capturing sudden changes and sudden oscillations of time series.