James T. Kwok

Longhui Yu, Weisen Jiang, Han Shi et al. · cambridge

Large language models (LLMs) have pushed the limits of natural language understanding and exhibited excellent problem-solving ability. Despite the great success, most existing open-source LLMs (e.g., LLaMA-2) are still far away from satisfactory for solving mathematical problem due to the complex reasoning procedures. To bridge this gap, we propose MetaMath, a fine-tuned language model that specializes in mathematical reasoning. Specifically, we start by bootstrapping mathematical questions by rewriting the question from multiple perspectives without extra knowledge, which results in a new dataset called MetaMathQA. Then we fine-tune the LLaMA-2 models on MetaMathQA. Experimental results on two popular benchmarks (i.e., GSM8K and MATH) for mathematical reasoning demonstrate that MetaMath outperforms a suite of open-source LLMs by a significant margin. Our MetaMath-7B model achieves 66.4% on GSM8K and 19.4% on MATH, exceeding the state-of-the-art models of the same size by 11.5% and 8.7%. Particularly, MetaMath-70B achieves an accuracy of 82.3% on GSM8K, slightly better than GPT-3.5-Turbo. We release all the MetaMathQA dataset, the MetaMath models with different model sizes and the training code for public use.

Shuhao Chen, Weisen Jiang, Yeqi Gong et al.

Fine-tuning large language models often undermines their safety alignment, a problem further amplified by harmful fine-tuning attacks in which adversarial data removes safeguards and induces unsafe behaviors. We propose SPARD, a defense framework that integrates Safety-Projected Alternating optimization with Relevance-Diversity aware data selection. SPARD employs SPAG, which optimizes alternatively between utility updates and explicit safety projections with a set of safe data to enforce safety constraints. To curate safe data, we introduce a Relevance-Diversity Determinantal Point Process to select compact safe data, balancing task relevance and safety coverage. Experiments on GSM8K and OpenBookQA under four harmful fine-tuning attacks demonstrate that SPARD consistently achieves the lowest average attack success rates, substantially outperforming state-of-the-art defense methods, while maintaining high task accuracy. Code is available at https://github.com/shuhao02/SPARD.

Quanming Yao, James T. Kwok · tsinghua

Matrix and tensor completion aim to recover a low-rank matrix / tensor from limited observations and have been commonly used in applications such as recommender systems and multi-relational data mining. A state-of-the-art matrix completion algorithm is Soft-Impute, which exploits the special "sparse plus low-rank" structure of the matrix iterates to allow efficient SVD in each iteration. Though Soft-Impute is a proximal algorithm, it is generally believed that acceleration destroys the special structure and is thus not useful. In this paper, we show that Soft-Impute can indeed be accelerated without comprising this structure. To further reduce the iteration time complexity, we propose an approximate singular value thresholding scheme based on the power method. Theoretical analysis shows that the proposed algorithm still enjoys the fast $O(1/T^2)$ convergence rate of accelerated proximal algorithms. We further extend the proposed algorithm to tensor completion with the scaled latent nuclear norm regularizer. We show that a similar "sparse plus low-rank" structure also exists, leading to low iteration complexity and fast $O(1/T^2)$ convergence rate. Extensive experiments demonstrate that the proposed algorithm is much faster than Soft-Impute and other state-of-the-art matrix and tensor completion algorithms.

Kai Chen, Yunhao Gou, Runhui Huang et al.

GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging for the open-source community. Existing vision-language models rely on external tools for speech processing, while speech-language models still suffer from limited or totally without vision-understanding capabilities. To address this gap, we propose the EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech abilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we surprisingly notice that omni-modal alignment can further enhance vision-language and speech abilities compared with the bi-modal aligned counterparts. Moreover, a lightweight style module is introduced for the flexible speech style controls including emotions and pitches. For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.

Hui Zhang, Quanming Yao, James T. Kwok et al. · tsinghua

Feature extractor plays a critical role in text recognition (TR), but customizing its architecture is relatively less explored due to expensive manual tweaking. In this work, inspired by the success of neural architecture search (NAS), we propose to search for suitable feature extractors. We design a domain-specific search space by exploring principles for having good feature extractors. The space includes a 3D-structured space for the spatial model and a transformed-based space for the sequential model. As the space is huge and complexly structured, no existing NAS algorithms can be applied. We propose a two-stage algorithm to effectively search in the space. In the first stage, we cut the space into several blocks and progressively train each block with the help of an auxiliary head. We introduce the latency constraint into the second stage and search sub-network from the trained supernet via natural gradient descent. In experiments, a series of ablation studies are performed to better understand the designed space, search algorithm, and searched architectures. We also compare the proposed method with various state-of-the-art ones on both hand-written and scene TR tasks. Extensive results show that our approach can achieve better recognition performance with less latency.

Shuhao Chen, Weisen Jiang, Baijiong Lin et al.

Recent works show that assembling multiple off-the-shelf large language models (LLMs) can harness their complementary abilities. To achieve this, routing is a promising method, which learns a router to select the most suitable LLM for each query. However, existing routing models are ineffective when multiple LLMs perform well for a query. To address this problem, in this paper, we propose a method called query-based Router by Dual Contrastive learning (RouterDC). The RouterDC model consists of an encoder and LLM embeddings, and we propose two contrastive learning losses to train the RouterDC model. Experimental results show that RouterDC is effective in assembling LLMs and largely outperforms individual top-performing LLMs as well as existing routing methods on both in-distribution (+2.76\%) and out-of-distribution (+1.90\%) tasks. Source code is available at https://github.com/shuhao02/RouterDC.

Weisen Jiang, Yu Zhang, James T. Kwok

Prompt tuning for pre-trained masked language models (MLM) has shown promising performance in natural language processing tasks with few labeled examples. It tunes a prompt for the downstream task, and a verbalizer is used to bridge the predicted token and label prediction. Due to the limited training data, prompt initialization is crucial for prompt tuning. Recently, MetaPrompting (Hou et al., 2022) uses meta-learning to learn a shared initialization for all task-specific prompts. However, a single initialization is insufficient to obtain good prompts for all tasks and samples when the tasks are complex. Moreover, MetaPrompting requires tuning the whole MLM, causing a heavy burden on computation and memory as the MLM is usually large. To address these issues, we use a prompt pool to extract more task knowledge and construct instance-dependent prompts via attention. We further propose a novel soft verbalizer (RepVerb) which constructs label embedding from feature embeddings directly. Combining meta-learning the prompt pool and RepVerb, we propose MetaPrompter for effective structured prompting. MetaPrompter is parameter-efficient as only the pool is required to be tuned. Experimental results demonstrate that MetaPrompter performs better than the recent state-of-the-arts and RepVerb outperforms existing soft verbalizers.

Baijiong Lin, Weisen Jiang, Feiyang Ye et al.

Multi-task learning aims to learn multiple related tasks simultaneously and has achieved great success in various fields. However, the disparity in loss and gradient scales among tasks often leads to performance compromises, and the balancing of tasks remains a significant challenge. In this paper, we propose Dual-Balancing Multi-Task Learning (DB-MTL) to achieve task balancing from both the loss and gradient perspectives. Specifically, DB-MTL achieves loss-scale balancing by performing logarithm transformation on each task loss, and rescales gradient magnitudes by normalizing all task gradients to comparable magnitudes using the maximum gradient norm. Extensive experiments on a number of benchmark datasets demonstrate that DB-MTL consistently performs better than the current state-of-the-art.

Yulong Zhang, Shuhao Chen, Weisen Jiang et al.

Limited transferability hinders the performance of deep learning models when applied to new application scenarios. Recently, Unsupervised Domain Adaptation (UDA) has achieved significant progress in addressing this issue via learning domain-invariant features. However, the performance of existing UDA methods is constrained by the large domain shift and limited target domain data. To alleviate these issues, we propose DomAin-guided Conditional Diffusion Model (DACDM) to generate high-fidelity and diversity samples for the target domain. In the proposed DACDM, by introducing class information, the labels of generated samples can be controlled, and a domain classifier is further introduced in DACDM to guide the generated samples for the target domain. The generated samples help existing UDA methods transfer from the source domain to the target domain more easily, thus improving the transfer performance. Extensive experiments on various benchmarks demonstrate that DACDM brings a large improvement to the performance of existing UDA methods.

Tao Li, Weisen Jiang, Fanghui Liu et al.

Pre-training followed by fine-tuning is widely adopted among practitioners. The performance can be improved by "model soups"~\cite{wortsman2022model} via exploring various hyperparameter configurations.The Learned-Soup, a variant of model soups, significantly improves the performance but suffers from substantial memory and time costs due to the requirements of (i) having to load all fine-tuned models simultaneously, and (ii) a large computational graph encompassing all fine-tuned models. In this paper, we propose Memory Efficient Hyperplane Learned Soup (MEHL-Soup) to tackle this issue by formulating the learned soup as a hyperplane optimization problem and introducing block coordinate gradient descent to learn the mixing coefficients. At each iteration, MEHL-Soup only needs to load a few fine-tuned models and build a computational graph with one combined model. We further extend MEHL-Soup to MEHL-Soup+ in a layer-wise manner. Experimental results on various ViT models and data sets show that MEHL-Soup(+) outperforms Learned-Soup(+) in terms of test accuracy, and also reduces memory usage by more than $13\times$. Moreover, MEHL-Soup(+) can be run on a single GPU and achieves $9\times$ speed up in soup construction compared with the Learned-Soup. The code is released at https://github.com/nblt/MEHL-Soup.

Xiaofeng Cao, Weixin Bu, Shengjun Huang et al.

Learning on big data brings success for artificial intelligence (AI), but the annotation and training costs are expensive. In future, learning on small data that approximates the generalization ability of big data is one of the ultimate purposes of AI, which requires machines to recognize objectives and scenarios relying on small data as humans. A series of learning topics is going on this way such as active learning and few-shot learning. However, there are few theoretical guarantees for their generalization performance. Moreover, most of their settings are passive, that is, the label distribution is explicitly controlled by finite training resources from known distributions. This survey follows the agnostic active sampling theory under a PAC (Probably Approximately Correct) framework to analyze the generalization error and label complexity of learning on small data in model-agnostic supervised and unsupervised fashion. Considering multiple learning communities could produce small data representation and related topics have been well surveyed, we thus subjoin novel geometric representation perspectives for small data: the Euclidean and non-Euclidean (hyperbolic) mean, where the optimization solutions including the Euclidean gradients, non-Euclidean gradients, and Stein gradient are presented and discussed. Later, multiple learning communities that may be improved by learning on small data are summarized, which yield data-efficient representations, such as transfer learning, contrastive learning, graph representation learning. Meanwhile, we find that the meta-learning may provide effective parameter update policies for learning on small data. Then, we explore multiple challenging scenarios for small data, such as the weak supervision and multi-label. Finally, multiple data applications that may benefit from efficient small data representation are surveyed.

Mingwei Xu, Xiaofeng Cao, Ivor W. Tsang et al.

Federated Learning (FL) typically aggregates client model parameters using a weighting approach determined by sample proportions. However, this naive weighting method may lead to unfairness and degradation in model performance due to statistical heterogeneity and the inclusion of noisy data among clients. Theoretically, distributional robustness analysis has shown that the generalization performance of a learning model with respect to any shifted distribution is bounded. This motivates us to reconsider the weighting approach in federated learning. In this paper, we replace the aforementioned weighting method with a new strategy that considers the generalization bounds of each local model. Specifically, we estimate the upper and lower bounds of the second-order origin moment of the shifted distribution for the current local model, and then use these bounds disagreements as the aggregation proportions for weightings in each communication round. Experiments demonstrate that the proposed weighting strategy significantly improves the performance of several representative FL algorithms on benchmark datasets.

Xiaofeng Cao, Yaming Guo, Ivor W. Tsang et al.

Hypothesis-pruning maximizes the hypothesis updates for active learning to find those desired unlabeled data. An inherent assumption is that this learning manner can derive those updates into the optimal hypothesis. However, its convergence may not be guaranteed well if those incremental updates are negative and disordered. In this paper, we introduce a black-box teaching hypothesis $h^\mathcal{T}$ employing a tighter slack term $\left(1+\mathcal{F}^{\mathcal{T}}(\widehat{h}_t)\right)Δ_t$ to replace the typical $2Δ_t$ for pruning. Theoretically, we prove that, under the guidance of this teaching hypothesis, the learner can converge into a tighter generalization error and label complexity bound than those non-educated learners who do not receive any guidance from a teacher:1) the generalization error upper bound can be reduced from $R(h^*)+4Δ_{T-1}$ to approximately $R(h^{\mathcal{T}})+2Δ_{T-1}$, and 2) the label complexity upper bound can be decreased from $4 θ\left(TR(h^{*})+2O(\sqrt{T})\right)$ to approximately $2θ\left(2TR(h^{\mathcal{T}})+3 O(\sqrt{T})\right)$. To be strict with our assumption, self-improvement of teaching is firstly proposed when $h^\mathcal{T}$ loosely approximates $h^*$. Against learning, we further consider two teaching scenarios: teaching a white-box and black-box learner. Experiments verify this idea and show better generalization performance than the fundamental active learning strategies, such as IWAL, IWAL-D, etc.

Weisen Jiang, Han Shi, Longhui Yu et al.

Self-Consistency samples diverse reasoning chains with answers and chooses the final answer by majority voting. It is based on forward reasoning and cannot further improve performance by sampling more reasoning chains when saturated. To further boost performance, we introduce backward reasoning to verify candidate answers. Specifically, for mathematical tasks, we mask a number in the question and ask the LLM to answer a backward question created by a simple template, i.e., to predict the masked number when a candidate answer is provided. Instead of using forward or backward reasoning alone, we propose FOBAR to combine FOrward and BAckward Reasoning for verification. Extensive experiments on six standard mathematical data sets and three LLMs show that FOBAR achieves state-of-the-art performance. In particular, FOBAR outperforms Self-Consistency, which uses forward reasoning alone, demonstrating that combining forward and forward reasoning is better. In addition, FOBAR performs better than existing verification methods, showing the effectiveness of the simple template used in backward reasoning and the proposed combination. Extensions to non-mathematical problems are also discussed and validated empirically.

Weisen Jiang, Baijiong Lin, Han Shi et al.

Recently, various merging methods have been proposed to build a multi-task model from task-specific finetuned models without retraining. However, existing methods suffer from a large performance deterioration compared to using multiple task-specific models. In this paper, we propose to inject task-specific knowledge into the merged model and design two parameter-efficient approaches (BYOM-FFT and BYOM-LoRA) to Build Your Own Multi-task model. BYOM-FFT is for merging fully finetuned models, while BYOM-LoRA is for LoRA-finetuned models. Both methods are data-free and computation-efficient. Extensive experiments on computer vision and natural language processing tasks show that the proposed BYOM methods outperform existing merging methods by a large margin. Moreover, BYOM-FFT is general and can be integrated into existing merging methods to further boost performance.

Weiyu Chen, Baijiong Lin, Xiaoyuan Zhang et al.

Many modern deep learning applications require balancing multiple objectives that are often conflicting. Examples include multi-task learning, fairness-aware learning, and the alignment of Large Language Models (LLMs). This leads to multi-objective deep learning, which tries to find optimal trade-offs or Pareto-optimal solutions by adapting mathematical principles from the field of Multi-Objective Optimization (MOO). However, directly applying gradient-based MOO techniques to deep neural networks presents unique challenges, including high computational costs, optimization instability, and the difficulty of effectively incorporating user preferences. This paper provides a comprehensive survey of gradient-based techniques for multi-objective deep learning. We systematically categorize existing algorithms based on their outputs: (i) methods that find a single, well-balanced solution, (ii) methods that generate a finite set of diverse Pareto-optimal solutions, and (iii) methods that learn a continuous Pareto set of solutions. In addition to this taxonomy, the survey covers theoretical analyses, key applications, practical resources, and highlights open challenges and promising directions for future research. A comprehensive list of multi-objective deep learning algorithms is available at https://github.com/Baijiong-Lin/Awesome-Multi-Objective-Deep-Learning.

Weiyu Chen, James T. Kwok

Multi-task learning, which optimizes performance across multiple tasks, is inherently a multi-objective optimization problem. Various algorithms are developed to provide discrete trade-off solutions on the Pareto front. Recently, continuous Pareto front approximations using a linear combination of base networks have emerged as a compelling strategy. However, it suffers from scalability issues when the number of tasks is large. To address this issue, we propose a novel approach that integrates a main network with several low-rank matrices to efficiently learn the Pareto manifold. It significantly reduces the number of parameters and facilitates the extraction of shared features. We also introduce orthogonal regularization to further bolster performance. Extensive experimental results demonstrate that the proposed approach outperforms state-of-the-art baselines, especially on datasets with a large number of tasks.

Shuhao Chen, Weisen Jiang, Changmiao Wang et al.

Inpatient medication recommendation requires clinicians to repeatedly select specific medications, doses, and routes as a patient's condition evolves. Existing benchmarks formulate this task as admission-level prediction over coarse drug codes with multi-hot diagnostic and procedure code inputs, failing to capture the per-timepoint, information-rich nature of real prescribing. We propose RxEval, a prescription-level benchmark that evaluates LLM prescribing capability by multiple-choice questions: each question presents a detailed patient profile and time-ordered clinical trajectory, requiring selection of specific medication-dose-route triples from real prescriptions and patient-specific distractors generated via reasoning-chain perturbation. RxEval comprises 1,547 questions spanning 584 patients, 18 diagnostic categories, and 969 unique medications. Evaluation of 16 LLMs shows that RxEval is both challenging and discriminative: F1 ranges from 45.18 to 77.10 across models, and the best Exact Match is only 46.10%. Error analysis reveals that even frontier models may overlook stated patient information and fail to derive clinical conclusions.

Yunhao Gou, Zhili Liu, Kai Chen et al.

Instruction tuning of Large Vision-language Models (LVLMs) has revolutionized the development of versatile models with zero-shot generalization across a wide range of downstream vision-language tasks. However, the diversity of training tasks of different sources and formats would lead to inevitable task conflicts, where different tasks conflict for the same set of model parameters, resulting in sub-optimal instruction-following abilities. To address that, we propose the Mixture of Cluster-conditional LoRA Experts (MoCLE), a novel Mixture of Experts (MoE) architecture designed to activate the task-customized model parameters based on the instruction clusters. A separate universal expert is further incorporated to improve generalization capabilities of MoCLE for novel instructions. Extensive experiments on InstructBLIP and LLaVA demonstrate the effectiveness of MoCLE.

Yanbin Wei, Qiushi Huang, James T. Kwok et al.

Knowledge Graph Completion (KGC) is crucial for addressing knowledge graph incompleteness and supporting downstream applications. Many models have been proposed for KGC. They can be categorized into two main classes: triple-based and text-based approaches. Triple-based methods struggle with long-tail entities due to limited structural information and imbalanced entity distributions. Text-based methods alleviate this issue but require costly training for language models and specific finetuning for knowledge graphs, which limits their efficiency. To alleviate these limitations, in this paper, we propose KICGPT, a framework that integrates a large language model (LLM) and a triple-based KGC retriever. It alleviates the long-tail problem without incurring additional training overhead. KICGPT uses an in-context learning strategy called Knowledge Prompt, which encodes structural knowledge into demonstrations to guide the LLM. Empirical results on benchmark datasets demonstrate the effectiveness of KICGPT with smaller training overhead and no finetuning.

Zhili Liu, Kai Chen, Jianhua Han et al.

Masked Autoencoder~(MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. However, when the various downstream tasks have data distributions different from the pre-training data, the semantically irrelevant pre-training information might result in negative transfer, impeding MAE's scalability. To address this issue, we propose a novel MAE-based pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE), which can be trained once but provides customized pre-training models for diverse downstream tasks. Different from the mixture of experts (MoE), our MoCE trains each expert only with semantically relevant images by using cluster-conditional gates. Thus, each downstream task can be allocated to its customized model pre-trained with data most similar to the downstream data. Experiments on a collection of 11 downstream tasks show that MoCE outperforms the vanilla MAE by 2.45\% on average. It also obtains new state-of-the-art self-supervised learning results on detection and segmentation.

Yanbin Wei, Shuai Fu, Weisen Jiang et al.

Large Language Models (LLMs) are increasingly used for various tasks with graph structures. Though LLMs can process graph information in a textual format, they overlook the rich vision modality, which is an intuitive way for humans to comprehend structural information and conduct general graph reasoning. The potential benefits and capabilities of representing graph structures as visual images (i.e., $\textit{visual graph}$) are still unexplored. To fill the gap, we innovatively propose an end-to-end framework, called $\textbf{G}$raph to v$\textbf{I}$sual and $\textbf{T}$extual Integr$\textbf{A}$tion (GITA), which firstly incorporates visual graphs into general graph reasoning. Besides, we establish $\textbf{G}$raph-based $\textbf{V}$ision-$\textbf{L}$anguage $\textbf{Q}$uestion $\textbf{A}$nswering (GVLQA) dataset from existing graph data, which is the first vision-language dataset for general graph reasoning purposes. Extensive experiments on the GVLQA dataset and five real-world datasets show that GITA outperforms mainstream LLMs in terms of general graph reasoning capabilities. Moreover, We highlight the effectiveness of the layout augmentation on visual graphs and pretraining on the GVLQA dataset.

Zhili Liu, Yunhao Gou, Kai Chen et al.

As the capabilities of large language models (LLMs) continue to expand, aligning these models with human values remains a significant challenge. Recent studies show that reasoning abilities contribute significantly to model safety, while integrating Mixture-of-Experts (MoE) architectures can further enhance alignment. In this work, we address a fundamental question: How to effectively incorporate reasoning abilities and MoE architectures into self-alignment process in LLMs? We propose Mixture of insighTful Experts (MoTE), a novel framework that synergistically combines reasoning chains and expert mixtures to improve self-alignments. From a data perspective, MoTE employs a structured reasoning chain comprising four key stages: Question Analysis, Answer Guidance, Safe Answer, and Safety Checking. This approach enhances safety through multi-step reasoning and proves effective even for smaller and less powerful LLMs (e.g., 7B models). From an architectural perspective, MoTE adopts a multi-LoRA framework with step-level routing, where each expert is dedicated to a specific reasoning step. This design eliminates the need for balance losses, ensures stable training, and supports adaptive inference lengths. Experimental results demonstrate that MoTE significantly improves model safety, jailbreak resistance, and over-refusal capabilities, achieving performance comparable to OpenAI's state-of-the-art o1 model.

Yunhao Gou, Kai Chen, Zhili Liu et al.

Recent breakthroughs in reasoning language models have significantly advanced text-based reasoning. On the other hand, Multi-modal Large Language Models (MLLMs) still lag behind, hindered by their outdated internal LLMs. Upgrading these is often prohibitively expensive, as it requires complete vision-language alignment retraining which is costly. To address this issue, we introduce Perception-Reasoning Decoupling, which modularizes the MLLM's reasoning component and makes it easily replaceable. This approach redefines the MLLM's role to convert multi-modal inputs into detailed textual outputs that can be processed by any powerful, external, text-only LLM reasoners. To align the MLLM's perceptual output with the final reasoning task, we propose a novel reinforcement learning algorithm called Visual Perception Optimization (VPO). VPO rewards the MLLM based on the correctness of answers generated by the external reasoner to produce faithful and query-relevant captions. Together, this decoupling pipeline and VPO form our Reasoning-Aligned PerceptIon Decoupling (RAPID) approach. Empirical results show that RAPID achieves significant performance gains on multi-modal reasoning benchmarks. Crucially, RAPID enables a novel inference-time scaling paradigm: Once trained with VPO, the MLLM can be paired with any state-of-the-art LLM reasoner for consistent performance improvement without retraining.

Xiaofeng Cao, Mingwei Xu, Xin Yu et al.

Learning with high-resource data has demonstrated substantial success in artificial intelligence (AI); however, the costs associated with data annotation and model training remain significant. A fundamental objective of AI research is to achieve robust generalization with limited-resource data. This survey employs agnostic active sampling theory within the Probably Approximately Correct (PAC) framework to analyze the generalization error and label complexity associated with learning from low-resource data in both model-agnostic supervised and unsupervised settings. Based on this analysis, we investigate a suite of optimization strategies tailored for low-resource data learning, including gradient-informed optimization, meta-iteration optimization, geometry-aware optimization, and LLMs-powered optimization. Furthermore, we provide a comprehensive overview of multiple learning paradigms that can benefit from low-resource data, including domain transfer, reinforcement feedback, and hierarchical structure modeling. Finally, we conclude our analysis and investigation by summarizing the key findings and highlighting their implications for learning with low-resource data.

Ziyu Zhou, Jiaxi Hu, Qingsong Wen et al.

In deep time series forecasting, the Fourier Transform (FT) is extensively employed for frequency representation learning. However, it often struggles in capturing multi-scale, time-sensitive patterns. Although the Wavelet Transform (WT) can capture these patterns through frequency decomposition, its coefficients are insensitive to change points in time series, leading to suboptimal modeling. To mitigate these limitations, we introduce the multi-order Wavelet Derivative Transform (WDT) grounded in the WT, enabling the extraction of time-aware patterns spanning both the overall trend and subtle fluctuations. Compared with the standard FT and WT, which model the raw series, the WDT operates on the derivative of the series, selectively magnifying rate-of-change cues and exposing abrupt regime shifts that are particularly informative for time series modeling. Practically, we embed the WDT into a multi-branch framework named WaveTS, which decomposes the input series into multi-scale time-frequency coefficients, refines them via linear layers, and reconstructs them into the time domain via the inverse WDT. Extensive experiments on ten benchmark datasets demonstrate that WaveTS achieves state-of-the-art forecasting accuracy while retaining high computational efficiency.

Yanbin Wei, Jiangyue Yan, Chun Kang et al.

Large Multimodal Models (LMMs) have shown generalized zero-shot capabilities in diverse domain question-answering (QA) tasks, including graph QA that involves complex graph topologies. However, most current approaches use only a single type of graph representation, namely Topology Representation Form (TRF), such as prompt-unified text descriptions or style-fixed visual styles. Those "one-size-fits-all" approaches fail to consider the specific preferences of different models or tasks, often leading to incorrect or overly long responses. To address this, we first analyze the characteristics and weaknesses of existing TRFs, and then design a set of TRFs, denoted by $F_{ZS}$, tailored to zero-shot graph QA. We then introduce a new metric, Graph Response Efficiency (GRE), which measures the balance between the performance and the brevity in graph QA. Built on these, we develop the DynamicTRF framework, which aims to improve both the accuracy and conciseness of graph QA. To be specific, DynamicTRF first creates a TRF Preference (TRFP) dataset that ranks TRFs based on their GRE scores, to probe the question-specific TRF preferences. Then it trains a TRF router on the TRFP dataset, to adaptively assign the best TRF from $F_{ZS}$ for each question during the inference. Extensive experiments across 7 in-domain algorithmic graph QA tasks and 2 out-of-domain downstream tasks show that DynamicTRF significantly enhances the zero-shot graph QA of LMMs in terms of accuracy

Yunhao Gou, Hansi Yang, Zhili Liu et al.

Visual Instruction Tuning (VIT) aims to enhance Multimodal Large Language Models (MLLMs), yet its effectiveness is often compromised by corrupted datasets with issues such as hallucinated content, incorrect responses, and poor OCR quality. Previous approaches to address these challenges have focused on refining datasets through high-quality data collection or rule-based filtering that can be costly or limited in scope. In this paper, we conduct a systematic investigation into the impact of corrupted data on MLLMs and discover that, although corrupted data degrade model performance, such adverse effects are largely reversible, and MLLMs are {\bf corrupted but not broken}. Specifically, we find that disabling a small subset of parameters can almost fully restore performance. Moreover, corrupted MLLMs inherently possess the capability to differentiate between clean and corrupted samples, facilitating dataset cleaning without external intervention. Building on these insights, we introduce a corruption-robust training paradigm that significantly surpasses existing strategies for mitigating the effects of corrupted data.

Hansi Yang, James T. Kwok

Distributed learning, which does not require gathering training data in a central location, has become increasingly important in the big-data era. In particular, random-walk-based decentralized algorithms are flexible in that they do not need a central server trusted by all clients and do not require all clients to be active in all iterations. However, existing distributed learning algorithms assume that all learning clients share the same task. In this paper, we consider the more difficult meta-learning setting, in which different clients perform different (but related) tasks with limited training data. To reduce communication cost and allow better privacy protection, we propose LoDMeta (Local Decentralized Meta-learning) with the use of local auxiliary optimization parameters and random perturbations on the model parameter. Theoretical results are provided on both convergence and privacy analysis. Empirical results on a number of few-shot learning data sets demonstrate that LoDMeta has similar meta-learning accuracy as centralized meta-learning algorithms, but does not require gathering data from each client and is able to better protect data privacy for each client.

Lifeng Shen, Jincheng Yu, Hansi Yang et al.

Mixup and its variants form a popular class of data augmentation techniques.Using a random sample pair, it generates a new sample by linear interpolation of the inputs and labels. However, generating only one single interpolation may limit its augmentation ability. In this paper, we propose a simple yet effective extension called multi-mix, which generates multiple interpolations from a sample pair. With an ordered sequence of generated samples, multi-mix can better guide the training process than standard mixup. Moreover, theoretically, this can also reduce the stochastic gradient variance. Extensive experiments on a number of synthetic and large-scale data sets demonstrate that multi-mix outperforms various mixup variants and non-mixup-based baselines in terms of generalization, robustness, and calibration.

Yunhao Gou, Kai Chen, Zhili Liu et al.

Multimodal large language models (MLLMs) have shown impressive reasoning abilities. However, they are also more vulnerable to jailbreak attacks than their LLM predecessors. Although still capable of detecting the unsafe responses, we observe that safety mechanisms of the pre-aligned LLMs in MLLMs can be easily bypassed with the introduction of image features. To construct robust MLLMs, we propose ECSO (Eyes Closed, Safety On), a novel training-free protecting approach that exploits the inherent safety awareness of MLLMs, and generates safer responses via adaptively transforming unsafe images into texts to activate the intrinsic safety mechanism of pre-aligned LLMs in MLLMs. Experiments on five state-of-the-art (SoTA) MLLMs demonstrate that ECSO enhances model safety significantly (e.g.,, 37.6% improvement on the MM-SafetyBench (SD+OCR) and 71.3% on VLSafe with LLaVA-1.5-7B), while consistently maintaining utility results on common MLLM benchmarks. Furthermore, we show that ECSO can be used as a data engine to generate supervised-finetuning (SFT) data for MLLM alignment without extra human intervention.

Qianli Ma, Zhen Liu, Zhenjing Zheng et al.

Time-Series Mining (TSM) is an important research area since it shows great potential in practical applications. Deep learning models that rely on massive labeled data have been utilized for TSM successfully. However, constructing a large-scale well-labeled dataset is difficult due to data annotation costs. Recently, pre-trained models have gradually attracted attention in the time series domain due to their remarkable performance in computer vision and natural language processing. In this survey, we provide a comprehensive review of Time-Series Pre-Trained Models (TS-PTMs), aiming to guide the understanding, applying, and studying TS-PTMs. Specifically, we first briefly introduce the typical deep learning models employed in TSM. Then, we give an overview of TS-PTMs according to the pre-training techniques. The main categories we explore include supervised, unsupervised, and self-supervised TS-PTMs. Further, extensive experiments involving 27 methods, 434 datasets, and 679 transfer learning scenarios are conducted to analyze the advantages and disadvantages of transfer learning strategies, Transformer-based models, and representative TS-PTMs. Finally, we point out some potential directions of TS-PTMs for future work.

Han Shi, Jiahui Gao, Hang Xu et al.

Recently over-smoothing phenomenon of Transformer-based models is observed in both vision and language fields. However, no existing work has delved deeper to further investigate the main cause of this phenomenon. In this work, we make the attempt to analyze the over-smoothing problem from the perspective of graph, where such problem was first discovered and explored. Intuitively, the self-attention matrix can be seen as a normalized adjacent matrix of a corresponding graph. Based on the above connection, we provide some theoretical analysis and find that layer normalization plays a key role in the over-smoothing issue of Transformer-based models. Specifically, if the standard deviation of layer normalization is sufficiently large, the output of Transformer stacks will converge to a specific low-rank subspace and result in over-smoothing. To alleviate the over-smoothing problem, we consider hierarchical fusion strategies, which combine the representations from different layers adaptively to make the output more diverse. Extensive experiment results on various data sets illustrate the effect of our fusion method.

Zac Wellmer, James T. Kwok

A World Model is a generative model used to simulate an environment. World Models have proven capable of learning spatial and temporal representations of Reinforcement Learning environments. In some cases, a World Model offers an agent the opportunity to learn entirely inside of its own dream environment. In this work we explore improving the generalization capabilities from dream environments to real environments (Dream2Real). We present a general approach to improve a controller's ability to transfer from a neural network dream environment to reality at little additional cost. These improvements are gained by drawing on inspiration from Domain Randomization, where the basic idea is to randomize as much of a simulator as possible without fundamentally changing the task at hand. Generally, Domain Randomization assumes access to a pre-built simulator with configurable parameters but oftentimes this is not available. By training the World Model using dropout, the dream environment is capable of creating a nearly infinite number of different dream environments. Previous use cases of dropout either do not use dropout at inference time or averages the predictions generated by multiple sampled masks (Monte-Carlo Dropout). Dropout's Dream Land leverages each unique mask to create a diverse set of dream environments. Our experimental results show that Dropout's Dream Land is an effective technique to bridge the reality gap between dream environments and reality. Furthermore, we additionally perform an extensive set of ablation studies.

Huapeng Wu, Jie Gui, Jun Zhang et al.

Recently, deep convolutional neural network methods have achieved an excellent performance in image superresolution (SR), but they can not be easily applied to embedded devices due to large memory cost. To solve this problem, we propose a pyramidal dense attention network (PDAN) for lightweight image super-resolution in this paper. In our method, the proposed pyramidal dense learning can gradually increase the width of the densely connected layer inside a pyramidal dense block to extract deep features efficiently. Meanwhile, the adaptive group convolution that the number of groups grows linearly with dense convolutional layers is introduced to relieve the parameter explosion. Besides, we also present a novel joint attention to capture cross-dimension interaction between the spatial dimensions and channel dimension in an efficient way for providing rich discriminative feature representations. Extensive experimental results show that our method achieves superior performance in comparison with the state-of-the-art lightweight SR methods.

Huapeng Wu, Jie Gui, Jun Zhang et al.

Recently, convolutional neural network (CNN) based image super-resolution (SR) methods have achieved significant performance improvement. However, most CNN-based methods mainly focus on feed-forward architecture design and neglect to explore the feedback mechanism, which usually exists in the human visual system. In this paper, we propose feedback pyramid attention networks (FPAN) to fully exploit the mutual dependencies of features. Specifically, a novel feedback connection structure is developed to enhance low-level feature expression with high-level information. In our method, the output of each layer in the first stage is also used as the input of the corresponding layer in the next state to re-update the previous low-level filters. Moreover, we introduce a pyramid non-local structure to model global contextual information in different scales and improve the discriminative representation of the network. Extensive experimental results on various datasets demonstrate the superiority of our FPAN in comparison with the state-of-the-art SR methods.

Haoang Chi, Feng Liu, Wenjing Yang et al.

In few-shot domain adaptation (FDA), classifiers for the target domain are trained with accessible labeled data in the source domain (SD) and few labeled data in the target domain (TD). However, data usually contain private information in the current era, e.g., data distributed on personal phones. Thus, the private information will be leaked if we directly access data in SD to train a target-domain classifier (required by FDA methods). In this paper, to thoroughly prevent the privacy leakage in SD, we consider a very challenging problem setting, where the classifier for the TD has to be trained using few labeled target data and a well-trained SD classifier, named few-shot hypothesis adaptation (FHA). In FHA, we cannot access data in SD, as a result, the private information in SD will be protected well. To this end, we propose a target orientated hypothesis adaptation network (TOHAN) to solve the FHA problem, where we generate highly-compatible unlabeled data (i.e., an intermediate domain) to help train a target-domain classifier. TOHAN maintains two deep networks simultaneously, where one focuses on learning an intermediate domain and the other takes care of the intermediate-to-target distributional adaptation and the target-risk minimization. Experimental results show that TOHAN outperforms competitive baselines significantly.

Han Shi, Jiahui Gao, Xiaozhe Ren et al.

Transformer-based models are popularly used in natural language processing (NLP). Its core component, self-attention, has aroused widespread interest. To understand the self-attention mechanism, a direct method is to visualize the attention map of a pre-trained model. Based on the patterns observed, a series of efficient Transformers with different sparse attention masks have been proposed. From a theoretical perspective, universal approximability of Transformer-based models is also recently proved. However, the above understanding and analysis of self-attention is based on a pre-trained model. To rethink the importance analysis in self-attention, we study the significance of different positions in attention matrix during pre-training. A surprising result is that diagonal elements in the attention map are the least important compared with other attention positions. We provide a proof showing that these diagonal elements can indeed be removed without deteriorating model performance. Furthermore, we propose a Differentiable Attention Mask (DAM) algorithm, which further guides the design of the SparseBERT. Extensive experiments verify our interesting findings and illustrate the effect of the proposed algorithm.

Bo Han, Quanming Yao, Tongliang Liu et al.

Classical machine learning implicitly assumes that labels of the training data are sampled from a clean distribution, which can be too restrictive for real-world scenarios. However, statistical-learning-based methods may not train deep learning models robustly with these noisy labels. Therefore, it is urgent to design Label-Noise Representation Learning (LNRL) methods for robustly training deep models with noisy labels. To fully understand LNRL, we conduct a survey study. We first clarify a formal definition for LNRL from the perspective of machine learning. Then, via the lens of learning theory and empirical study, we figure out why noisy labels affect deep models' performance. Based on the theoretical guidance, we categorize different LNRL methods into three directions. Under this unified taxonomy, we provide a thorough discussion of the pros and cons of different categories. More importantly, we summarize the essential components of robust LNRL, which can spark new directions. Lastly, we propose possible research directions within LNRL, such as new datasets, instance-dependent LNRL, and adversarial LNRL. We also envision potential directions beyond LNRL, such as learning with feature-noise, preference-noise, domain-noise, similarity-noise, graph-noise and demonstration-noise.

Yaqing Wang, Quanming Yao, James T. Kwok

Matrix learning is at the core of many machine learning problems. A number of real-world applications such as collaborative filtering and text mining can be formulated as a low-rank matrix completion problem, which recovers incomplete matrix using low-rank assumptions. To ensure that the matrix solution has a low rank, a recent trend is to use nonconvex regularizers that adaptively penalize singular values. They offer good recovery performance and have nice theoretical properties, but are computationally expensive due to repeated access to individual singular values. In this paper, based on the key insight that adaptive shrinkage on singular values improve empirical performance, we propose a new nonconvex low-rank regularizer called "nuclear norm minus Frobenius norm" regularizer, which is scalable, adaptive and sound. We first show it provably holds the adaptive shrinkage property. Further, we discover its factored form which bypasses the computation of singular values and allows fast optimization by general optimization algorithms. Stable recovery and convergence are guaranteed. Extensive low-rank matrix completion experiments on a number of synthetic and real-world data sets show that the proposed method obtains state-of-the-art recovery performance while being the fastest in comparison to existing low-rank matrix learning methods.

Han Shi, Haozheng Fan, James T. Kwok

The (variational) graph auto-encoder and its variants have been popularly used for representation learning on graph-structured data. While the encoder is often a powerful graph convolutional network, the decoder reconstructs the graph structure by only considering two nodes at a time, thus ignoring possible interactions among edges. On the other hand, structured prediction, which considers the whole graph simultaneously, is computationally expensive. In this paper, we utilize the well-known triadic closure property which is exhibited in many real-world networks. We propose the triad decoder, which considers and predicts the three edges involved in a local triad together. The triad decoder can be readily used in any graph-based auto-encoder. In particular, we incorporate this to the (variational) graph auto-encoder. Experiments on link prediction, node clustering and graph generation show that the use of triads leads to more accurate prediction, clustering and better preservation of the graph characteristics.

Han Shi, Renjie Pi, Hang Xu et al.

Neural Architecture Search (NAS) has shown great potentials in finding better neural network designs. Sample-based NAS is the most reliable approach which aims at exploring the search space and evaluating the most promising architectures. However, it is computationally very costly. As a remedy, the one-shot approach has emerged as a popular technique for accelerating NAS using weight-sharing. However, due to the weight-sharing of vastly different networks, the one-shot approach is less reliable than the sample-based approach. In this work, we propose BONAS (Bayesian Optimized Neural Architecture Search), a sample-based NAS framework which is accelerated using weight-sharing to evaluate multiple related architectures simultaneously. Specifically, we apply Graph Convolutional Network predictor as a surrogate model for Bayesian Optimization to select multiple related candidate models in each iteration. We then apply weight-sharing to train multiple candidate models simultaneously. This approach not only accelerates the traditional sample-based approach significantly, but also keeps its reliability. This is because weight-sharing among related architectures are more reliable than those in the one-shot approach. Extensive experiments are conducted to verify the effectiveness of our method over many competing algorithms.

Shuai Zheng, Ziyue Huang, James T. Kwok

Communication overhead is a major bottleneck hampering the scalability of distributed machine learning systems. Recently, there has been a surge of interest in using gradient compression to improve the communication efficiency of distributed neural network training. Using 1-bit quantization, signSGD with majority vote achieves a 32x reduction on communication cost. However, its convergence is based on unrealistic assumptions and can diverge in practice. In this paper, we propose a general distributed compressed SGD with Nesterov's momentum. We consider two-way compression, which compresses the gradients both to and from workers. Convergence analysis on nonconvex problems for general gradient compressors is provided. By partitioning the gradient into blocks, a blockwise compressor is introduced such that each gradient block is compressed and transmitted in 1-bit format with a scaling factor, leading to a nearly 32x reduction on communication. Experimental results show that the proposed method converges as fast as full-precision distributed momentum SGD and achieves the same testing accuracy. In particular, on distributed ResNet training with 7 workers on the ImageNet, the proposed algorithm achieves the same testing accuracy as momentum SGD using full-precision gradients, but with $46\%$ less wall clock time.

Shuai Zheng, James T. Kwok

Stochastic methods with coordinate-wise adaptive stepsize (such as RMSprop and Adam) have been widely used in training deep neural networks. Despite their fast convergence, they can generalize worse than stochastic gradient descent. In this paper, by revisiting the design of Adagrad, we propose to split the network parameters into blocks, and use a blockwise adaptive stepsize. Intuitively, blockwise adaptivity is less aggressive than adaptivity to individual coordinates, and can have a better balance between adaptivity and generalization. We show theoretically that the proposed blockwise adaptive gradient descent has comparable convergence rate as its counterpart with coordinate-wise adaptive stepsize, but is faster up to some constant. We also study its uniform stability and show that blockwise adaptivity can lead to lower generalization error than coordinate-wise adaptivity. Experimental results show that blockwise adaptive gradient descent converges faster and improves generalization performance over Nesterov's accelerated gradient and Adam.

Yaqing Wang, James T. Kwok, Lionel M. Ni

Convolutional sparse coding (CSC) can learn representative shift-invariant patterns from multiple kinds of data. However, existing CSC methods can only model noises from Gaussian distribution, which is restrictive and unrealistic. In this paper, we propose a general CSC model capable of dealing with complicated unknown noise. The noise is now modeled by Gaussian mixture model, which can approximate any continuous probability density function. We use the expectation-maximization algorithm to solve the problem and design an efficient method for the weighted CSC problem in maximization step. The crux is to speed up the convolution in the frequency domain while keeping the other computation involving weight matrix in the spatial domain. Besides, we simultaneously update the dictionary and codes by nonconvex accelerated proximal gradient algorithm without bringing in extra alternating loops. The resultant method obtains comparable time and space complexity compared with existing CSC methods. Extensive experiments on synthetic and real noisy biomedical data sets validate that our method can model noise effectively and obtain high-quality filters and representation.

Quanming Yao, Xiawei Guo, James T. Kwok et al.

To meet the standard of differential privacy, noise is usually added into the original data, which inevitably deteriorates the predicting performance of subsequent learning algorithms. In this paper, motivated by the success of improving predicting performance by ensemble learning, we propose to enhance privacy-preserving logistic regression by stacking. We show that this can be done either by sample-based or feature-based partitioning. However, we prove that when privacy-budgets are the same, feature-based partitioning requires fewer samples than sample-based one, and thus likely has better empirical performance. As transfer learning is difficult to be integrated with a differential privacy guarantee, we further combine the proposed method with hypothesis transfer learning to address the problem of learning across different organizations. Finally, we not only demonstrate the effectiveness of our method on two benchmark data sets, i.e., MNIST and NEWS20, but also apply it into a real application of cross-organizational diabetes prediction from RUIJIN data set, where privacy is of significant concern.

Quanming Yao, James T. Kwok

Robust matrix factorization (RMF), which uses the $\ell_1$-loss, often outperforms standard matrix factorization using the $\ell_2$-loss, particularly when outliers are present. The state-of-the-art RMF solver is the RMF-MM algorithm, which, however, cannot utilize data sparsity. Moreover, sometimes even the (convex) $\ell_1$-loss is not robust enough. In this paper, we propose the use of nonconvex loss to enhance robustness. To address the resultant difficult optimization problem, we use majorization-minimization (MM) optimization and propose a new MM surrogate. To improve scalability, we exploit data sparsity and optimize the surrogate via its dual with the accelerated proximal gradient algorithm. The resultant algorithm has low time and space complexities and is guaranteed to converge to a critical point. Extensive experiments demonstrate its superiority over the state-of-the-art in terms of both accuracy and scalability.

Shuai Zheng, James T. Kwok

Variance reduction has been commonly used in stochastic optimization. It relies crucially on the assumption that the data set is finite. However, when the data are imputed with random noise as in data augmentation, the perturbed data set be- comes essentially infinite. Recently, the stochastic MISO (S-MISO) algorithm is introduced to address this expected risk minimization problem. Though it converges faster than SGD, a significant amount of memory is required. In this pa- per, we propose two SGD-like algorithms for expected risk minimization with random perturbation, namely, stochastic sample average gradient (SSAG) and stochastic SAGA (S-SAGA). The memory cost of SSAG does not depend on the sample size, while that of S-SAGA is the same as those of variance reduction methods on un- perturbed data. Theoretical analysis and experimental results on logistic regression and AUC maximization show that SSAG has faster convergence rate than SGD with comparable space requirement, while S-SAGA outperforms S-MISO in terms of both iteration complexity and storage.

Lu Hou, James T. Kwok

The power law has been observed in the degree distributions of many biological neural networks. Sparse deep neural networks, which learn an economical representation from the data, resemble biological neural networks in many ways. In this paper, we study if these artificial networks also exhibit properties of the power law. Experimental results on two popular deep learning models, namely, multilayer perceptrons and convolutional neural networks, are affirmative. The power law is also naturally related to preferential attachment. To study the dynamical properties of deep networks in continual learning, we propose an internal preferential attachment model to explain how the network topology evolves. Experimental results show that with the arrival of a new task, the new connections made follow this preferential attachment process.

Yaqing Wang, Quanming Yao, James T. Kwok et al.

Convolutional sparse coding (CSC) has been popularly used for the learning of shift-invariant dictionaries in image and signal processing. However, existing methods have limited scalability. In this paper, instead of convolving with a dictionary shared by all samples, we propose the use of a sample-dependent dictionary in which filters are obtained as linear combinations of a small set of base filters learned from the data. This added flexibility allows a large number of sample-dependent patterns to be captured, while the resultant model can still be efficiently learned by online learning. Extensive experimental results show that the proposed method outperforms existing CSC algorithms with significantly reduced time and space requirements.