Large Language Models (LLMs), built on Transformer architectures, exhibit remarkable generalization across a wide range of tasks. However, fine-tuning these models for specific tasks remains resource-intensive due to their extensive parameterization. In this paper, we explore two remarkable phenomena related to the attention mechanism during the fine-tuning of LLMs (where Wq, Wk, and Wv denote the weights of the query, key, and value layers, respectively). The first phenomenon, termed “Unequal Importance of Attention Matrices”, highlights the impact of fine-tuning different weight matrices. It shows that optimizing the Wv matrix yields significantly better performance than optimizing the Wk matrix. Fine-tuning only the Wq and Wv matrices is computationally efficient while delivering results comparable to, or even better than fine-tuning all three matrices (Wq, Wk, and Wv). The second phenomenon, “Attention Matrices with Customized Learning Rate Lead to Better Convergence”, emphasizes the importance of assigning distinct learning rates to these matrices. Specifically, a higher learning rate for the Wv matrix compared to Wq and Wk accelerates convergence and improves performance. Building on these insights, we propose a new strategy that improves fine-tuning efficiency in terms of both storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach, supporting our theoretical findings. Our analysis lays the theoretical groundwork for configuring and improving algorithms in LLMs fine-tuning.
论文检索
输入标题、作者或关键词，从 1,280 篇学术成果中精准定位
From General Relation Patterns to Task-Specific Decision-Making in Continual Multi-Agent Coordination
PDF ↗Continual Multi-Agent Reinforcement Learning (Co-MARL) requires agents to address catastrophic forgetting issues while learning new coordination policies with the dynamics team. In this paper, we delve into the core of Co-MARL, namely Relation Patterns, which refer to agents’ general understanding of interactions. In addition to generality, relation patterns exhibit task-specificity when mapped to different action spaces. To this end, we propose a novel method called General Relation Patterns-Guided Task-specific Decision-Maker (RPG). In RPG, agents extract relation patterns from dynamic observation spaces using a relation capturer. These task-agnostic relation patterns are then mapped to different action spaces via a task-specific decision-maker generated by a conditional hypernetwork. To combat forgetting, we further introduce regularization items on both the relation capturer and the conditional hypernetwork. Results on SMAC and LBF demonstrate that RPG effectively prevents catastrophic forgetting when learning new tasks and achieves zero-shot generalization to unseen tasks.
Smoothed online convex optimization (SOCO), in which the online player incurs both a hitting cost and a switching cost for changing its decisions, has garnered significant attention in recent years. While existing studies typically assume that the gradient information is revealed immediately, such an assumption may not hold in some real-world applications. To overcome this limitation, we investigate SOCO with delayed feedback, and develop two online algorithms that can minimize the dynamic regret with switching cost. Firstly, we extend Mild-OGD, an existing algorithm that adopts the meta-expert framework for online convex optimization with delayed feedback, to account for switching cost. Specifically, we analyze the switching cost in the expert-algorithm of Mild-OGD, and then modify its meta-algorithm to incorporate this cost when assigning the weight to each expert. We demonstrate that our proposed method, Smelt-DOGD can achieve an O(√(dT(P_T+1))) dynamic regret bound with switching cost, where d is the maximum delay and P_T is the path-length. Secondly, we develop an efficient variant to reduce the number of projections per round from O(log T) to 1, yet maintaining the same theoretical guarantee. The key idea is to construct a new surrogate loss defined over a simpler domain for expert-algorithms so that these experts do not need to perform the complex projection operations in each round. Finally, we conduct experiments to validate the effectiveness and efficiency of our algorithms.
Multimodal sentiment analysis (MSA) has shown promising results but often poses significant challenges in real-world applications due to its dependence on the complete and aligned multimodal sequences. While existing approaches attempt to address missing modalities through feature reconstruction, they often neglect the complex interplay between homogeneous and heterogeneous relationships in multimodal features. To address this problem, we propose Decoupled-Adaptive Reconstruction (DAR), a novel framework that explicitly addresses these limitations through two key components: (1) a mutual information-based decoupling module that decomposes features into common and independent representations, and (2) a reconstruction module that independently processes these decoupled features before fusion for downstream tasks. Extensive experiments on two benchmark datasets demonstrate that DAR significantly outperforms existing methods in both modality reconstruction and sentiment analysis tasks, particularly in scenarios with missing or unaligned modalities. Our results show improvements of 2.21% in bi-classification accuracy and 3.9% in regression error compared to state-of-the-art baselines on the MOSEI dataset.
Multi-view clustering aims to integrate complementary information from multiple views to improve clustering performance. However, existing ensemble-based methods suffer from information loss due to their reliance on single-granularity labels, limiting the discriminative capability of learned representations. Meanwhile, representation and graph fusion-based approaches face challenges such as explicit view alignment and manual weight tuning, making them less effective for heterogeneous views with varying data distributions. To address these limitations, we propose a novel multi-view clustering framework via Multi-granularity Ensemble (MGE), fully using the multi-granularity information across diverse views for accurate and consistent clustering. Specifically, MGE first modifies the hierarchical clustering and then leverages it on each view (including the fused view) to achieve multi-granularity labels. Moreover, the cross-view and cross-granularity fusion strategy is designed to learn a robust co-association similarity matrix, which effectively preserves the fine-grained and coarse-grained structures of multi-view data and facilitates subsequent clustering. Therefore, MGE can provide a comprehensive representation of local and global patterns within data, eliminating the requirement for view alignment and weight tuning. Experiments demonstrate that MGE consistently outperforms state-of-the-art methods across multiple datasets, validating its effectiveness and superiority in handling heterogeneous views.
Empowering Vision Transformers with Multi-Scale Causal Intervention for Long-Tailed Image Classification
PDF ↗Causal inference has emerged as a promising approach to mitigate long-tail classification by handling the biases introduced by class imbalance. However, along with the change of advanced backbone models from Convolutional Neural Networks (CNNs) to Visual Transformers (ViT), existing causal models may not achieve an expected performance gain. This paper investigates the influence of existing causal models on CNNs and ViT variants, highlighting that ViT's global feature representation makes it hard for causal methods to model associations between fine-grained features and predictions, which leads to difficulties in classifying tail classes with similar visual appearance. To address these issues, this paper proposes TSCNet, a two-stage causal modeling method to discover fine-grained causal associations through multi-scale causal interventions. Specifically, in the hierarchical causal representation learning stage (HCRL), it decouples the background and objects, applying backdoor interventions at both the patch and feature level to prevent model from using class-irrelevant areas to infer labels which enhances fine-grained causal representation. In the counterfactual logits' bias calibration stage (CLBC), it refines the optimization of model's decision boundary by adaptive constructing counterfactual balanced data distribution to remove the spurious associations in the logits caused by data distribution. Extensive experiments conducted on various long-tail benchmarks demonstrate that the proposed TSCNet can eliminate multiple biases introduced by data imbalance, which outperforms existing methods.
This paper studies lexicographic multi-armed bandits (MAB), where after selecting an arm, the agent observes a reward vector including multiple objectives, each with a different level of importance. Although previous literature has proposed the algorithm for lexicographic MAB, their algorithm suffers from several limitations: (1) it exhibits poor adversarial robustness due to its reliance on stochastic rewards, (2) its regret bound is suboptimal compared to single-objective counterparts, and (3) the regret bound does not adapt to specific problem instances. To address these limitations, we study lexicographic MAB with adversarial corruptions, where an adversary might corrupt the stochastic rewards with a corruption budget of C. First, when the value of C is known, we propose an algorithm achieving a problem-dependent regret bound of O(∑(log T / Δⁱ(a) + C)) for the i-th objective (i ∈ [M]), where Δⁱ(a) is the reward gap for arm a on the i-th objective, and M is the number of objectives. In the purely stochastic setting (C=0), this regret bound approaches optimality. Second, we introduce another algorithm that does not require value of C but incurs a less favorable regret bound of O(∑(γ_T / Δⁱ(a) + γ_T)) for the i-th objective, where γ_T = O((log T)² + KC(log T)²). Finally, we conduct experiments on both synthetic and real-world datasets to verify the effectiveness of our algorithms.
We study online task scheduling problems where tasks arrive sequentially and are processed by the platform or server. The service processes for tasks are multi-stage and are modeled as episodic Markov Decision Processes (MDPs). While processing a task, the system acquires rewards by consuming resources. The goal of the platform is to maximize the reward-to-cost ratio over a sequence of K tasks. Online scheduling with multi-stage tasks faces two major challenges: intra-dependence among the different stages within a task and inter-dependence among different tasks. These challenges are further exacerbated by the unknown rewards, costs, and task arrival distribution. To address these challenges, we propose the Robbins-Monro-based Value Iteration for Ratio Maximization (RM^2VI) algorithm. Specifically,RM^2VI addresses ``intra-dependence'' through optimistic value iteration and handles ``inter-dependence'' using the Robbins-Monro method. The algorithm has a greedy structure and achieves a sub-linear regret of O(K^(3/4)), establishing the no-regret property (per-task). We test RM^2VI in two synthetic experiments of sale promotion in E-commerce and machine learning job training in cloud computing. The results show RM^2VI achieves the best reward-to-cost ratio compared with the baselines.
Deep learning methods have demonstrated remarkable performance across various communication signal processing tasks. However, most signal classification methods require a substantial amount of labeled samples for training, posing significant challenges in the field of communication signals, as labeling necessitates expert knowledge. This paper proposes a novel self-supervised signal classification method called Spectral-Guided Self-Supervised Signal Classification (SGSSC). Specifically, to leverage frequency-domain information with modulation semantics as prior knowledge for the model, we design a previously unexplored pretext task tailored to the format of signal data. This task involves predicting spectral information from masked time-domain signals, enabling the model to learn implicit signal features through cross-domain pattern transformation. Furthermore, the pretext task in the SGSSC method is relevant to the downstream classification task, and using traditional fine-tuning strategies on the downstream task may lead to the loss of certain features associated with the pretext task. Therefore, we propose an attention mechanism-based fine-tuning strategy that adaptively integrates pre-trained features from different levels. Extensive experimental results validate the superiority of the SGSSC method. For instance, when the proportion of labeled samples is only 0.5%, our method achieves an average improvement of 2.3% in downstream classification tasks compared to the best-performing self-supervised training strategies.
A Priori Estimation of the Approximation, Optimization and Generalization Errors of Random Neural Networks for Solving Partial Differential Equations
PDF ↗In recent years, neural networks have achieved remarkable progress in various fields and have also drawn much attention in applying them on scientific problems. A line of methods involving neural networks for solving partial differential equations (PDEs), such as Physics-Informed Neural Networks (PINNs) and the Deep Ritz Method (DRM), has emerged. Although these methods outperform classical numerical methods in certain cases, the optimization problems involving neural networks are typically non-convex and non-smooth, which can result in unsatisfactory solutions for PDEs. In contrast to deterministic neural networks, the hidden weights of random neural networks are sampled from some prior distribution and only the output weights participate in training. This makes training much simpler, but it remains unclear how to select the prior distribution. In this paper, we focus on Barron type functions and approximate them under Sobolev norms by random neural networks with clear prior distribution. In addition to the approximation error, we also derive bounds for the optimization and generalization errors of random neural networks for solving PDEs when the solutions are Barron type functions.
Ensemble clustering learns a consensus result by integrating a set of base clustering results. Recently, anchor-based methods construct an anchor similarity matrix to represent the affinity relationships among samples, significantly improving computational efficiency. However, these methods struggle with fixed anchors generated by static anchor learning strategies, which lead to low-quality anchor similarity matrix and poor clustering accuracy. To address this issue, we propose a novel method named dynamic anchor-based ensemble clustering via hypergraph reconstruction (YACHT). Specifically, YACHT first transforms the base clustering results into a hypergraph and designs a novel hypergraph enhancement strategy to improve the reliability of the initial hypergraph. YACHT reconstructs the hypergraph through matrix factorization and introduces a mapping matrix to filter out redundant information, capturing a high-quality anchor similarity matrix. Then, YACHT attempts to incorporate the hypergraph into the optimization objective to achieve hypergraph updates. To ensure the accuracy of hypergraph updates, we impose a hypergraph regularizer and a local consensus information alignment term. The alignment term is implemented by minimizing the discrepancy between the label partition derived from the hypergraph regularizer and the local consensus information indicator matrix extracted from the base clustering results. Extensive experimental results demonstrate the outstanding performance of the proposed YACHT. The code is available at https://github.com/scu-kdde/YACHT.
Although previous studies have applied diffusion models to time series forecasting, these efforts have struggled to preserve the intrinsic temporal correlations within the series, leading to suboptimal predictive outcomes. This failure primarily results from the introduction of independent, identically distributed (i.i.d.) noise. In the forward process, the addition of i.i.d. noise to the time series gradually diminishes these temporal correlations. The reverse process starts with i.i.d. noise and lacks priors related to temporal correlations, which can result in directional biases during sampling. From a frequency-domain perspective, noise disrupts the low-frequency-dominated structure of trend components, making it difficult for the model to learn long-term temporal dependencies. To address these limitations, we introduce a decomposition prediction framework to complement the novel Temporal Correlation-Empowered Diffusion Model. Overall, We decompose the time series into trend and residual components, predict them using a base model and a diffusion model, and then combine the results. Specifically, a frequency-domain MLP model was adopted as the base model due to its not distorting the original sequence, and better the capture of long-range temporal dependencies. The diffusion model incorporates two key modules to capture short- and mid-range temporal correlations: the Maintaining Temporal Correlation Module and the Redesigned Initial Module. Extensive experiments across multiple datasets demonstrate that the proposed method significantly outperforms related strong baselines.
Decision trees and forests have achieved successes in various real applications, most working with all testing classes known in training data. In this work, we focus on learning with augmented class via forests, where an augmented class may appear in testing data yet not in training data. We incorporate information of augmented class into trees' splitting, that is, augmented Gini impurity, a new splitting criterion is introduced to exploit some unlabeled data from testing distribution. We then develop the Learning with Augmented Class via Forests (short for LACForest) approach, which constructs shallow forests according to the augmented Gini impurity and then splits forests with pseudo-labeled augmented instances for better performance. We also develop deep neural forests via an optimization objective based on our augmented Gini impurity, which essentially utilizes the representation power of neural networks for forests. Theoretically, we present the convergence analysis for our augmented Gini impurity, and we finally conduct experiments to evaluate our approaches. The code is available at https://github.com/nju-xuf/LACForest.
Multi-station weather prediction provides weather forecasts for specific geographical locations, playing an important role in various aspects of daily life. Existing methods consider the relationships between individual stations discretely, making it difficult to model the continuous spatiotemporal processes of atmospheric motion, which results in suboptimal prediction outcomes. This paper proposes the Continuous Diffusive Prediction Network (CDPNet) to model the real-world continuous weather change process from discrete station observation data. CDPNet consists of two core modules: the Continuous Calibrated Initialization (CCI) and the Diffusive Difference Estimation (DDE). The CCI module interpolates data between observation stations to construct a spatially continuous physical field and ensures temporal continuity by integrating directional information from a global perspective. It accurately represents the current physical state and provides a foundation for future weather prediction. Moreover, the DDE module explicitly captures the spatial diffusion process and estimates the diffusive differences between consecutive time steps, effectively modeling spatio-temporally continuous atmospheric motion. Likewise, directional information on weather changes is introduced from the entire historical series to mitigate estimation uncertainty and improve the performance of weather prediction. Extensive experiments on the Weather2K and Global Wind/Temp datasets demonstrate that CDPNet outperforms state-of-the-art models.
In many real applications, the data attributes are incremental and the samples are stored with accumulated feature spaces gradually. Although there are several elegant approaches to tackling this problem, the theoretical analysis is still limited. There exist at least two challenges and fundamental questions. 1) How to derive the generalization bounds of these approaches? 2) Under what conditions do these approaches have a strong generalization guarantee? To solve these crucial but rarely studied problems, we provide a comprehensive theoretical analysis in this paper. We begin by summarizing and refining four strategies for addressing feature incremental data. Subsequently, we derive their generalization bounds, providing rigorous and quantitative insights. The theoretical findings highlight the key factors influencing the generalization abilities of different strategies. In tackling the above two fundamental problems, we also provide valuable guidance for exploring other learning challenges in dynamic environments. Finally, the comprehensive experimental and theoretical results mutually validate each other, underscoring the reliability of our conclusions.
Beyond Statistical Analysis: Multimodal Framework for Time Series Forecasting with LLM-Driven Temporal Pattern
PDF ↗Accurate forecasting of time series is crucial for many applications in the real world. Conventional methods primarily rely on statistical analysis of historical data, often leading to overfitting and failing to account for background information and constraints imposed by external events. Therefore, introducing large language models (LLMs) with robust textual capabilities holds significant potential. However, due to the inherent limitations of LLMs in handling numerical data, they do not exhibit advantages in precise numerical prediction tasks. Therefore, we propose a framework to integrate LLMs with conventional methods synergistically. Rather than directly outputting numerical predictions, we leverage the capabilities of the LLMs to generate textual temporal patterns, thereby fully utilizing their inherent knowledge and reasoning abilities. Additionally, we introduce a memory network designed to decode these textual representations into a format that numerical models can effectively interpret. This approach not only capitalizes on the strengths of the LLM in text processing but also bridges the gap between textual and numerical data, enhancing the overall predictive performance of the model. Our experimental results demonstrate the framework's effectiveness, achieving state-of-the-art performance on various benchmark datasets.
DaringFed: A Dynamic Bayesian Persuasion Pricing for Online Federated Learning Under Two-sided Incomplete Information
PDF ↗Online Federated Learning (OFL) is a real-time learning paradigm that sequentially executes parameter aggregation immediately for each random arriving client. To motivate clients to participate in OFL, it is crucial to offer appropriate incentives to offset the training resource consumption. However, the design of incentive mechanisms in OFL is constrained by the dynamic variability of Two-sided Incomplete Information (TII) concerning resources, where the server is unaware of the clients’ dynamically changing computational resources, while clients lack knowledge of the real-time communication resources allocated by the server. To incentivize clients to participate in training by offering dynamic rewards to each arriving client, we design a novel Dynamic Bayesian persuasion pricing for online Federated learning (DaringFed) under TII. Specifically, we begin by formulating the interaction between the server and clients as a dynamic signaling and pricing allocation problem within a Bayesian persuasion game, and then demonstrate the existence of a unique Bayesian persuasion Nash equilibrium. By deriving the optimal design of DaringFed under one-sided incomplete information, we further analyze the approximate optimal design of DaringFed with a specific bound under TII. Finally, extensive evaluation conducted on real datasets demonstrate that DaringFed optimizes accuracy and converges speed by 16.99%, while experiments with synthetic datasets validate the convergence of estimate unknown values and the effectiveness of DaringFed in improving the server’s utility by up to 12.6%.
Partial label learning (PLL) is a significant weakly supervised learning framework, where each training example corresponds to a set of candidate labels and only one label is the ground-truth label. For the first time, this paper investigates the partial label clustering problem, which takes advantage of the limited available partial labels to improve the clustering performance. Specifically, we first construct a weight matrix of examples based on their relationships in the feature space and disambiguate the candidate labels to estimate the ground-truth label based on the weight matrix. Then, we construct a set of must-link and cannot-link constraints based on the disambiguation results. Moreover, we propagate the initial must-link and cannot-link constraints based on an adversarial prior promoted dual-graph learning approach. Finally, we integrate weight matrix construction, label disambiguation, and pairwise constraints propagation into a joint model to achieve mutual enhancement. We also theoretically prove that a better disambiguated label matrix can help improve clustering performance. Comprehensive experiments demonstrate our method realizes superior performance when comparing with state-of-the-art constrained clustering methods, and outperforms PLL and semi-supervised PLL methods when only limited samples are annotated. The code and appendix are publicly available at https://github.com/xyt-ml/PLC.
Spatial transcriptomics, comprising spatial location and high-throughput gene expression information, provides revolutionary insights into disease discovery and cellular evolution. Spatial transcriptomic clustering, which pinpoints distinct spatial domains within tissues, reveals cellular interactions and enhances our understanding of the intricate architecture of tissues. Existing methods typically construct spatial graphs using a static radius based on spatial coordinates, which hinders the accurate identification of spatial domains and complicates the precise partitioning of boundary nodes within clusters. To address this issue, we introduce a novel spatially resolved transcriptomics data clustering network (TSstc). Specifically, we employ a tailored spatial-scale modulation approach, constructing different spatial graphs incrementally as the radius of the spatial domain expands, and a Spatiality-Aware Sampling (SAS) strategy is proposed to aggregate node representations by considering the spatial dependencies between spots. We then use GCN encoders to learn gene embedding with gene graph and multiple spatial embeddings with spatial graphs. During training, we incorporate cross-view correlation-based tailored spatial regularization constraints to preserve high-quality neighbor relationships across spatial embeddings at different scales. Finally, a zero-inflated negative binomial model is utilized to capture the global probability distribution of gene expression profiles. Extensive experimental results demonstrate that our approach surpasses existing state-of-the-art methods in clustering tasks and related downstream applications.
Hierarchical optimization is attracting significant attentions as it can be applied to a broad range of machine learning tasks. Recently, many algorithms are proposed to improve the theoretical results of minimax and bilevel optimizations. Among these works, a core issue that has not been well studies is to escape saddle point and find local minimum. In this paper, thus, we investigate the methods to achieve second-order optimality for nonconvex minimax and bilevel optimization. Specifically, we propose a new algorithm named PRGDA without the computation of second order derivative of the primal function. In nonconvex-strongly-concave minimax optimization, we prove that our algorithm can find a second-order stationary point with the gradient complexity that matches state-of-the-art result to find first-order stationary point. To our best knowledge, PRGDA is the first stochastic algorithm that is guaranteed to obtain the second-order stationary point for nonconvex minimax problems. In nonconvex-strongly-convex bilevel optimization, our method also achieves better gradient complexity to find local minimum. Finally, we conduct two numerical experiments to validate the performance of our new method.