Recent Newton-type federated learning algorithms have demonstrated linear convergence with respect to the communication rounds. However, communicating Hessian matrices is often unfeasible due to their quadratic communication complexity. In this paper, we introduce a novel approach to tackle this issue while still achieving fast convergence rates. Our proposed method, named as Federated Newton Sketch methods (FedNS), approximates the centralized Newton's method by communicating the sketched square-root Hessian instead of the exact Hessian. To enhance communication efficiency, we reduce the sketch size to match the effective dimension of the Hessian matrix. We provide convergence analysis based on statistical learning for the federated Newton sketch approaches. Specifically, our approaches reach super-linear convergence rates w.r.t. the communication rounds for the first time. We validate the effectiveness of our algorithms through various experiments, which coincide with our theoretical findings.
论文检索
输入标题、作者或关键词，从 2,866 篇学术成果中精准定位
Overparameterization often leads to benign overfitting, where deep neural networks can be trained to overfit the training data but still generalize well on unseen data. However, it lacks a generalized asymptotic framework for nonlinear regressions and connections to conventional complexity notions. In this paper, we propose a generalized high-dimensional analysis for nonlinear regression models, including various nonlinear feature mapping methods and subsampling. Specifically, we first provide an implicit regularization parameter and asymptotic equivalents related to a classical complexity notion, i.e., effective dimension. We then present a high-dimensional analysis for nonlinear ridge regression and extend it to ridgeless regression in the under-parameterized and over-parameterized regimes, respectively. We find that the limiting risks decrease with the effective dimension. Motivated by these theoretical findings, we propose an algorithm, namely RFRed, to improve generalization ability. Finally, we validate our theoretical findings and the proposed algorithm through several experiments.
Existing approaches to Implicit Neural Representation (INR) can be interpreted as a global scene representation via a linear combination of Fourier bases of different frequencies. However, such universal basis functions can limit the representation capability in local regions where a specific component is unnecessary, resulting in unpleasant artifacts. To this end, we introduce a learnable spatial mask that effectively dispatches distinct Fourier bases into respective regions. This translates into collaging Fourier patches, thus enabling an accurate representation of complex signals. Comprehensive experiments demonstrate the superior reconstruction quality of the proposed approach over existing baselines across various INR tasks, including image fitting, video representation, and 3D shape representation. Our method outperforms all other baselines, improving the image fitting PSNR by over 3dB and 3D reconstruction to 98.81 IoU and 0.0011 Chamfer Distance.
Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency
PDF ↗Although recent methods in Unsupervised Domain Adaptation (UDA) have achieved success in segmenting rainy or snowy scenes by improving consistency, they face limitations when dealing with more challenging scenarios like foggy and night scenes. We argue that these prior methods excessively focus on weather-specific features in adverse scenes, which exacerbates the existing domain gaps. To address this issue, we propose a new metric to evaluate the severity of all adverse scenes and offer a novel perspective that enables task unification across all adverse scenarios. Our method focuses on Severity, allowing our model to learn more consistent features and facilitate domain distribution alignment, thereby alleviating domain gaps. Unlike the vague descriptions of consistency in previous methods, we introduce Cross-domain Consistency, which is quantified using the Structure Similarity Index Measure (SSIM) to measure the distance between the source and target domains. Specifically, our unified model consists of two key modules: the Merging Style Augmentation Module (MSA) and the Severity Perception Mask Module (SPM). The MSA module transforms all adverse scenes into augmented scenes, effectively eliminating weather-specific features and enhancing Cross-domain Consistency. The SPM module incorporates a Severity Perception mechanism, guiding a Mask operation that enables our model to learn highly consistent features from the augmented scenes. Our unified framework, named PASS (Parsing All adverSe Scenes), achieves significant performance improvements over state-of-the-art methods on widely-used benchmarks for all adverse scenes. Notably, the performance of PASS is superior to Semi-Unified models and even surpasses weather-specific models.
The deep model training procedure requires large-scale datasets of annotated data. Due to the difficulty of annotating a large number of samples, label noise caused by incorrect annotations is inevitable, resulting in low model performance and poor model generalization. To combat label noise, current methods usually select clean samples based on the small-loss criterion and use these samples for training. Due to some noisy samples similar to clean ones, these small-loss criterion-based methods are still affected by label noise. To address this issue, in this work, we propose Regroup Median Loss (RML) to reduce the probability of selecting noisy samples and correct losses of noisy samples. RML randomly selects samples with the same label as the training samples based on a new loss processing method. Then, we combine the stable mean loss and the robust median loss through a proposed regrouping strategy to obtain robust loss estimation for noisy samples. To further improve the model performance against label noise, we propose a new sample selection strategy and build a semi-supervised method based on RML. Compared to state-of-the-art methods, for both the traditionally trained and semi-supervised models, RML achieves a significant improvement on synthetic and complex real-world datasets. The source is at https://github.com/Feng-peng-Li/Regroup-Loss-Median-to-Combat-Label-Noise.
Towards Continual Learning Desiderata via HSIC-Bottleneck Orthogonalization and Equiangular Embedding
PDF ↗Deep neural networks are susceptible to catastrophic forgetting when trained on sequential tasks. Various continual learning (CL) methods often rely on exemplar buffers or/and network expansion for balancing model stability and plasticity, which, however, compromises their practical value due to privacy and memory concerns. Instead, this paper considers a strict yet realistic setting, where the training data from previous tasks is unavailable and the model size remains relatively constant during sequential training. To achieve such desiderata, we propose a conceptually simple yet effective method that attributes forgetting to layer-wise parameter overwriting and the resulting decision boundary distortion. This is achieved by the synergy between two key components: HSIC-Bottleneck Orthogonalization (HBO) implements non-overwritten parameter updates mediated by Hilbert-Schmidt independence criterion in an orthogonal space and EquiAngular Embedding (EAE) enhances decision boundary adaptation between old and new tasks with predefined basis vectors. Extensive experiments demonstrate that our method achieves competitive accuracy performance, even with absolute superiority of zero exemplar buffer and 1.02x the base model.
The de novo generation of hit-like molecules that show bioactivity and drug-likeness is an important task in computer-aided drug discovery. Although artificial intelligence can generate molecules with desired chemical properties, most previous studies have ignored the influence of disease-related cellular environments. This study proposes a novel deep generative model called GxVAEs to generate hit-like molecules from gene expression profiles by leveraging two joint variational autoencoders (VAEs). The first VAE, ProfileVAE, extracts latent features from gene expression profiles. The extracted features serve as the conditions that guide the second VAE, which is called MolVAE, in generating hit-like molecules. GxVAEs bridge the gap between molecular generation and the cellular environment in a biological system, and produce molecules that are biologically meaningful in the context of specific diseases. Experiments and case studies on the generation of therapeutic molecules show that GxVAEs outperforms current state-of-the-art baselines and yield hit-like molecules with potential bioactivity and drug-like properties. We were able to successfully generate the potential molecular structures with therapeutic effects for various diseases from patients’ disease profiles.
In open-set recognition (OSR), a promising strategy is exploiting pseudo-unknown data outside given K known classes as an additional K+1-th class to explicitly model potential open space. However, treating unknown classes without distinction is unequal for them relative to known classes due to the category-agnostic and scale-agnostic of the unknowns. This inevitably not only disrupts the inherent distributions of unknown classes but also incurs both class-wise and instance-wise imbalances between known and unknown classes. Ideally, the OSR problem should model the whole class space as K+∞, but enumerating all unknowns is impractical. Since the core of OSR is to effectively model the boundaries of known classes, this means just focusing on the unknowns nearing the boundaries of targeted known classes seems sufficient. Thus, as a compromise, we convert the open classes from infinite to K, with a novel concept Target-Aware Universum (TAU) and propose a simple yet effective framework Dual Contrastive Learning with Target-Aware Universum (DCTAU). In details, guided by the targeted known classes, TAU automatically expands the unknown classes from the previous 1 to K, effectively alleviating the distribution disruption and the imbalance issues mentioned above. Then, a novel Dual Contrastive (DC) loss is designed, where all instances irrespective of known or TAU are considered as positives to contrast with their respective negatives. Experimental results indicate DCTAU sets a new state-of-the-art.
Recently, many carefully designed graph representation learning methods have achieved impressive performance on either strong heterophilic or homophilic graphs, but not both. Therefore, they are incapable of generalizing well across real-world graphs with different levels of homophily. This is attributed to their neglect of homophily in heterophilic graphs, and vice versa. In this paper, we propose a two-fold filtering mechanism to mine homophily in heterophilic graphs, and vice versa. In particular, we extend the graph heat equation to perform heterophilic aggregation of global information from a long distance. The resultant filter can be exactly approximated by the Possion-Charlier (PC) polynomials. To further exploit information at multiple orders, we introduce a powerful graph convolution PC-Conv and its instantiation PCNet for the node classification task. Compared to the state-of-the-art GNNs, PCNet shows competitive performance on well-known homophilic and heterophilic graphs. Our implementation is available at https://github.com/uestclbh/PC-Conv.
In this paper, we address the performance degradation of efficient diffusion models by introducing Multi-architecturE Multi-Expert diffusion models (MEME). We identify the need for tailored operations at different time-steps in diffusion processes and leverage this insight to create compact yet high-performing models. MEME assigns distinct architectures to different time-step intervals, balancing convolution and self-attention operations based on observed frequency characteristics. We also introduce a soft interval assignment strategy for comprehensive training. Empirically, MEME operates 3.3 times faster than baselines while improving image generation quality (FID scores) by 0.62 (FFHQ) and 0.37 (CelebA). Though we validate the effectiveness of assigning more optimal architecture per time-step, where efficient models outperform the larger models, we argue that MEME opens a new design choice for diffusion models that can be easily applied in other scenarios, such as large multi-expert models.
Transformer-based models have demonstrated remarkable performance in various domains, including natural language processing, image processing and generative modeling. The most significant contributor to the successful performance of Transformer models is the self-attention mechanism, which allows for a comprehensive understanding of the interactions between tokens in the input sequence. However, there is a well-known scalability issue, the quadratic dependency (i.e. O(n^2)) of self-attention operations on the input sequence length n, making the handling of lengthy sequences challenging. To address this limitation, there has been a surge of research on efficient transformers, aiming to alleviate the quadratic dependency on the input sequence length. Among these, the Nyströmformer, which utilizes the Nyström method to decompose the attention matrix, achieves superior performance in both accuracy and throughput. However, its landmark selection exhibits redundancy, and the model incurs computational overhead when calculating the pseudo-inverse matrix. We propose a novel Nyström method-based transformer, called Proxyformer. Unlike the traditional approach of selecting landmarks from input tokens, the Proxyformer utilizes trainable neural memory, called proxy tokens, for landmarks. By integrating contrastive learning, input injection, and a specialized dropout for the decomposed matrix, Proxyformer achieves top-tier performance for long sequence tasks in the Long Range Arena benchmark.
We study a novel variant of a contextual bandit problem with multi-dimensional reward feedback formulated as a mixed-effects model, where the correlations between multiple feedback are induced by sharing stochastic coefficients called random effects. We propose a novel algorithm, Mixed-Effects Contextual UCB (ME-CUCB), achieving tildeO(d sqrt(mT)) regret bound after T rounds where d is the dimension of contexts and m is the dimension of outcomes, with either known or unknown covariance structure. This is a tighter regret bound than that of the naive canonical linear bandit algorithm ignoring the correlations among rewards. We prove a lower bound of Omega(d sqrt(mT)) matching the upper bound up to logarithmic factors. To our knowledge, this is the first work providing a regret analysis for mixed-effects models and algorithms involving weighted least-squares estimators. Our theoretical analysis faces a significant technical challenge in that the error terms do not constitute martingales since the weights depend on the rewards. We overcome this challenge by using covering numbers, of theoretical interest in its own right. We provide numerical experiments demonstrating the advantage of our proposed algorithm, supporting the theoretical claims.
Although meta-learning seems promising performance in the realm of rapid adaptability, it is constrained by fixed cardinality. When faced with tasks of varying cardinalities that were unseen during training, the model lacks its ability. In this paper, we address and resolve this challenge by harnessing `label equivalence' emerged from stochastic numeric label assignments during episodic task sampling. Questioning what defines ``true" meta-learning, we introduce the ``any-way" learning paradigm, an innovative model training approach that liberates model from fixed cardinality constraints. Surprisingly, this model not only matches but often outperforms traditional fixed-way models in terms of performance, convergence speed, and stability. This disrupts established notions about domain generalization. Furthermore, we argue that the inherent label equivalence naturally lacks semantic information. To bridge this semantic information gap arising from label equivalence, we further propose a mechanism for infusing semantic class information into the model. This would enhance the model's comprehension and functionality. Experiments conducted on renowned architectures like MAML and ProtoNet affirm the effectiveness of our method.
In reinforcement learning, temporal abstraction in the action space, exemplified by action repetition, is a technique to facilitate policy learning through extended actions. However, a primary limitation in previous studies of action repetition is its potential to degrade performance, particularly when sub-optimal actions are repeated. This issue often negates the advantages of action repetition. To address this, we propose a novel algorithm named Uncertainty-aware Temporal Extension (UTE). UTE employs ensemble methods to accurately measure uncertainty during action extension. This feature allows policies to strategically choose between emphasizing exploration or adopting an uncertainty-averse approach, tailored to their specific needs. We demonstrate the effectiveness of UTE through experiments in Gridworld and Atari 2600 environments. Our findings show that UTE outperforms existing action repetition algorithms, effectively mitigating their inherent limitations and significantly enhancing policy learning efficiency.
Thompson sampling (TS) has been known for its outstanding empirical performance supported by theoretical guarantees across various reward models in the classical stochastic multi-armed bandit problems. Nonetheless, its optimality is often restricted to specific priors due to the common observation that TS is fairly insensitive to the choice of the prior when it comes to asymptotic regret bounds. However, when the model contains multiple parameters, the optimality of TS highly depends on the choice of priors, which casts doubt on the generalizability of previous findings to other models. To address this gap, this study explores the impact of selecting noninformative priors, offering insights into the performance of TS when dealing with new models that lack theoretical understanding. We first extend the regret analysis of TS to the model of uniform distributions with unknown supports, which would be the simplest non-regular model. Our findings reveal that changing noninformative priors can significantly affect the expected regret, aligning with previously known results in other multiparameter bandit models. Although the uniform prior is shown to be optimal, we highlight the inherent limitation of its optimality, which is limited to specific parameterizations and emphasizes the significance of the invariance property of priors. In light of this limitation, we propose a slightly modified TS-based policy, called TS with Truncation (TS-T), which can achieve the asymptotic optimality for the Gaussian models and the uniform models by using the reference prior and the Jeffreys prior that are invariant under one-to-one reparameterizations. This policy provides an alternative approach to achieving optimality by employing fine-tuned truncation, which would be much easier than hunting for optimal priors in practice.
Spear and Shield: Adversarial Attacks and Defense Methods for Model-Based Link Prediction on Continuous-Time Dynamic Graphs
PDF ↗Real-world graphs are dynamic, constantly evolving with new interactions, such as financial transactions in financial networks. Temporal Graph Neural Networks (TGNNs) have been developed to effectively capture the evolving patterns in dynamic graphs. While these models have demonstrated their superiority, being widely adopted in various important fields, their vulnerabilities against adversarial attacks remain largely unexplored. In this paper, we propose T-SPEAR, a simple and effective adversarial attack method for link prediction on continuous-time dynamic graphs, focusing on investigating the vulnerabilities of TGNNs. Specifically, before the training procedure of a victim model, which is a TGNN for link prediction, we inject edge perturbations to the data that are unnoticeable in terms of the four constraints we propose, and yet effective enough to cause malfunction of the victim model. Moreover, we propose a robust training approach T-SHIELD to mitigate the impact of adversarial attacks. By using edge filtering and enforcing temporal smoothness to node embeddings, we enhance the robustness of the victim model. Our experimental study shows that T-SPEAR significantly degrades the victim model's performance on link prediction tasks, and even more, our attacks are transferable to other TGNNs, which differ from the victim model assumed by the attacker. Moreover, we demonstrate that T-SHIELD effectively filters out adversarial edges and exhibits robustness against adversarial attacks, surpassing the link prediction performance of the naive TGNN by up to 11.2% under T-SPEAR. The code and datasets are available at https://github.com/wooner49/T-spear-shield
The generation of energetically optimal 3D molecular conformers is crucial in cheminformatics and drug discovery. While deep generative models have been utilized for direct generation in Euclidean space, this approach encounters challenges, including the complexity of navigating a vast search space. Recent generative models that implement simplifications to circumvent these challenges have achieved state-of-the-art results, but this simplified approach unavoidably creates a gap between the generated conformers and the ground-truth conformational landscape. To bridge this gap, we introduce DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization, a novel diffusion framework that enables direct learning of nonlinear diffusion processes in prior-constrained Euclidean space for the optimization of 3D molecular conformers. Through the incorporation of an SE(3)-equivariant Schrödinger bridge, we establish the roto-translational equivariance of the generated conformers. Our framework is model-agnostic and offers an easily implementable solution for the post hoc optimization of conformers produced by any generation method. Through comprehensive evaluations and analyses, we establish the strengths of our framework, substantiating the application of the Schrödinger bridge for molecular conformer optimization. First, our approach consistently outperforms four baseline approaches, producing conformers with higher diversity and improved quality. Then, we show that the intermediate conformers generated during our diffusion process exhibit valid and chemically meaningful characteristics. We also demonstrate the robustness of our method when starting from conformers of diverse quality, including those unseen during training. Lastly, we show that the precise generation of low-energy conformers via our framework helps in enhancing the downstream prediction of molecular properties. The code is available at https://github.com/Danyeong-Lee/DiSCO.
OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models
PDF ↗Large language models (LLMs) with hundreds of billions of parameters require powerful server-grade GPUs for inference, limiting their practical deployment. To address this challenge, we introduce the outlier-aware weight quantization (OWQ) method, which aims to minimize LLM's footprint through low-precision representation. OWQ prioritizes a small subset of structured weights sensitive to quantization, storing them in high-precision, while applying highly tuned quantization to the remaining dense weights. This sensitivity-aware mixed-precision scheme reduces the quantization error notably, and extensive experiments demonstrate that 3.1-bit models using OWQ perform comparably to 4-bit models optimized by OPTQ. Furthermore, OWQ incorporates a parameter-efficient fine-tuning for task-specific adaptation, called weak column tuning (WCT), enabling accurate task-specific LLM adaptation with minimal memory overhead in the optimized format. OWQ represents a notable advancement in the flexibility, efficiency, and practicality of LLM optimization literature. The source code is available at https://github.com/xvyaward/owq.
Task-free online continual learning (TF-CL) is a challenging problem where the model incrementally learns tasks without explicit task information. Although training with entire data from the past, present as well as future is considered as the gold standard, naive approaches in TF-CL with the current samples may be conflicted with learning with samples in the future, leading to catastrophic forgetting and poor plasticity. Thus, a proactive consideration of an unseen future sample in TF-CL becomes imperative. Motivated by this intuition, we propose a novel TF-CL framework considering future samples and show that injecting adversarial perturbations on both input data and decision-making is effective. Then, we propose a novel method named Doubly Perturbed Continual Learning (DPCL) to efficiently implement these input and decision-making perturbations. Specifically, for input perturbation, we propose an approximate perturbation method that injects noise into the input data as well as the feature vector and then interpolates the two perturbed samples. For decision-making process perturbation, we devise multiple stochastic classifiers. We also investigate a memory management scheme and learning rate scheduling reflecting our proposed double perturbations. We demonstrate that our proposed method outperforms the state-of-the-art baseline methods by large margins on various TF-CL benchmarks.
In the machine learning field, the challenge of effectively learning with limited data has become increasingly crucial. Active Learning (AL) algorithms play a significant role in this by enhancing model performance. We introduce a novel AL algorithm, termed Co-learning (CoLAL), designed to select the most diverse and representative samples within a training dataset. This approach utilizes noisy labels and predictions made by the primary model on unlabeled data. By leveraging a probabilistic graphical model, we combine two multi-class classifiers into a binary one. This classifier determines if both the main and the peer models agree on a prediction. If they do, the unlabeled sample is assumed to be easy to classify and is thus not beneficial to increase the target model's performance. We prioritize data that represents the unlabeled set without overlapping decision boundaries. The discrepancies between these boundaries can be estimated by the probability that two models result in the same prediction. Through theoretical analysis and experimental validation, we reveal that the integration of noisy labels into the peer model effectively identifies target model's potential inaccuracies. We evaluated the CoLAL method across seven benchmark datasets: four text datasets (AGNews, DBPedia, PubMed, SST-2) and text-based state-of-the-art (SOTA) baselines, and three image datasets (CIFAR100, MNIST, OpenML-155) and computer vision SOTA baselines. The results show that our CoLAL method significantly outperforms existing SOTA in text-based AL, and is competitive with SOTA image-based AL techniques.