Deploying Large Language Models (LLMs) on edge devices is increasingly important, as it eliminates reliance on network connections, reduces expensive API calls, and enhances user privacy. However, on-device deployment is challenging due to the limited computational resources of edge devices. In particular, the key bottleneck stems from memory bandwidth constraints related to weight loading. Weight-only quantization effectively reduces memory access, yet often induces significant accuracy degradation. Recent efforts to incorporate sub-branches have shown promise for mitigating quantization errors, but these methods either lack robust optimization strategies or rely on suboptimal objectives. To address these gaps, we propose FeedBack Quantization (FBQuant), a novel approach inspired by negative feedback mechanisms in automatic control. FBQuant inherently ensures that the reconstructed weights remain bounded by the quantization process, thereby reducing the risk of overfitting. To further offset the additional latency introduced by sub-branches, we develop an efficient CUDA kernel that decreases 60% of extra inference time. Comprehensive experiments demonstrate the efficiency and effectiveness of FBQuant across various LLMs. Notably, for 3-bit Llama2-7B, FBQuant improves zero-shot accuracy by 1.2%.
论文检索
输入标题、作者或关键词，从 1,235 篇学术成果中精准定位
Fast Second-Order Online Kernel Learning Through Incremental Matrix Sketching and Decomposition
PDF ↗Second-order Online Kernel Learning (OKL) has attracted considerable research interest due to its promising predictive performance in streaming environments. However, existing second-order OKL approaches suffer from at least quadratic time complexity with respect to the pre-set budget, rendering them unsuitable for large-scale datasets. Moreover, the singular value decomposition required to obtain explicit feature mapping is computationally expensive due to the complete decomposition process. To address these issues, we propose FORKS, a fast incremental matrix sketching and decomposition approach tailored for second-order OKL. FORKS constructs an incremental maintenance paradigm for second-order kernelized gradient descent, which includes incremental matrix sketching for kernel approximation and incremental matrix decomposition for explicit feature mapping construction. Theoretical analysis demonstrates that FORKS achieves a logarithmic regret guarantee on par with other second-order approaches while maintaining a linear time complexity w.r.t. the budget, significantly enhancing efficiency over existing methods. We validate the performance of our method through extensive experiments conducted on real-world datasets, demonstrating its superior scalability and robustness against adversarial attacks.
The learnware paradigm focuses on leveraging numerous established high-performing models to solve machine learning tasks instead of starting from scratch. As the key concept of this paradigm, a learnware consists of a well-trained model of any structure and a specification that characterizes the model's capabilities, allowing it to be identified and reused for future tasks. Given the existence of numerous real-world models trained on diverse label spaces, effectively identifying and combining these models to address tasks involving previously unseen label spaces represents a critical challenge in this paradigm. In this paper, we make the first attempt to identify and reuse effective learnware combinations for tackling learning tasks across different label spaces, extending their applicability beyond the original purposes of individual learnwares. To this end, we introduce a statistical class-wise specification for establishing similarity relations between various label spaces. Leveraging these relations, we model the utility of a learnware combination as a minimum-cost maximum-flow problem, and further develop fine-grained learnware identification and assembly methods. Extensive experiments with thousands of heterogeneous models validate our approach, demonstrating that reusing identified learnware combinations can outperform both training from scratch and fine-tuning a generic pre-trained model.
Graph domain adaptation (GDA), which transfers knowledge from a labeled source domain to an unlabeled target graph domain, attracts considerable attention in numerous fields. However, existing methods commonly employ message-passing neural networks (MPNNs) to learn domain-invariant representations by aligning the entire domain distribution, inadvertently neglecting category-level distribution alignment and potentially causing category confusion. To address the problem, we propose an effective framework named Coupling Category Alignment (CoCA) for GDA, which effectively addresses the category alignment issue with theoretical guarantees. CoCA incorporates a graph convolutional network branch and a graph kernel network branch, which explore graph topology in implicit and explicit manners. To mitigate category-level domain shifts, we leverage knowledge from both branches, iteratively filtering highly reliable samples from the target domain using one branch and fine-tuning the other accordingly. Furthermore, with these reliable target domain samples, we incorporate the coupled branches into a holistic contrastive learning framework. This framework includes multi-view contrastive learning to ensure consistent representations across the dual branches, as well as cross-domain contrastive learning to achieve category-level domain consistency. Theoretically, we establish a sharper generalization bound, which ensures the effectiveness of category alignment. Extensive experiments on benchmark datasets validate the superiority of the proposed CoCA compared with baselines.
Explainable artificial intelligence (XAI) is an important area in the AI community, and interpretability is crucial for building robust and trustworthy AI models. While previous work has explored model-level and instance-level explainable graph learning, there has been limited investigation into explainable graph representation learning. In this paper, we focus on representation-level explainable graph learning and ask a fundamental question: What specific information about a graph is captured in graph representations? Our approach is inspired by graph kernels, which evaluate graph similarities by counting substructures within specific graph patterns. Although the pattern counting vector can serve as an explainable representation, it has limitations such as ignoring node features and being high-dimensional. To address these limitations, we introduce a framework (PXGL-GNN) for learning and explaining graph representations through graph pattern analysis. We start by sampling graph substructures of various patterns. Then, we learn the representations of these patterns and combine them using a weighted sum, where the weights indicate the importance of each graph pattern's contribution. We also provide theoretical analyses of our methods, including robustness and generalization. In our experiments, we show how to learn and explain graph representations for real-world data using pattern analysis. Additionally, we compare our method against multiple baselines in both supervised and unsupervised learning tasks to demonstrate its effectiveness.
In this paper, we propose a family of novel Deep Hierarchical Transitive-Aligned Graph Kernels (DHTAGK) for graph classification. To this end, we commence by developing a new Hierarchical Aligned Graph Auto-Encoder (HA-GAE) to construct transitive-aligned embedding graphs that encapsulate the structural correspondence information between graphs. The DHTAGK kernels then measure either the Jensen-Shannon Divergence between the adjacency matrices or the Gaussian kernel between the node feature matrices of the embedding graphs. Unlike the classical R-convolution kernels and node-based alignment kernels, the DHTAGK kernels can capture the transitive structural correspondence information and thus ensure the positive definiteness. Furthermore, the HA-GAE enables the DHTAGK kernels to simultaneously reflect both local and global graph structures and identify common structural patterns. Experimental results show that the DHTAGK kernels outperform state-of-the-art graph kernels and deep learning methods on benchmark datasets.
Graph Neural Networks (GNNs) have demonstrated remarkable effectiveness across various tasks but are often hindered by their high computational overhead. GNN-to-MLP distillation provides a promising remedy by transferring knowledge from complex GNNs to lightweight MLPs. However, existing methods largely overlook the differences in aggregation mechanisms and heterogeneous architectures. Simplifying such intricate information into MLP potentially causes information loss or distortion, ultimately resulting in suboptimal performance. This paper proposes an aggregation mechanism enhanced GNN distillation framework (AMEND). AMEND introduces multi-scope aggregation context preservation to replicate the teacher's broad aggregation scopes and an aggregation-enhanced centered kernel alignment method to match the teacher's aggregation patterns. To ensure efficient and robust knowledge transfer, we integrate a manifold mixup strategy, enabling the student to capture the teacher's insights into mixed data distributions. Experimental results on 8 standard and 4 large-scale datasets demonstrate that AMEND consistently outperforms state-of-the-art distillation methods.
In this paper, we propose a new model to learn Adaptive Kernel-based Representations (AKBR) for graph classification. Unlike state-of-the-art R-convolution graph kernels that are defined by merely counting any pair of isomorphic substructures between graphs and cannot provide an end-to-end learning mechanism for the classifier, the proposed AKBR approach aims to define an end-to-end representation learning model to construct an adaptive kernel matrix for graphs. To this end, we commence by leveraging a novel feature-channel attention mechanism to capture the interdependencies between different substructure invariants of original graphs. The proposed AKBR model can thus effectively identify the structural importance of different substructures, and compute the R-convolution kernel between pairwise graphs associated with the more significant substructures specified by their structural attentions. Furthermore, the proposed AKBR model employs all sample graphs as the prototype graphs, naturally providing an end-to-end learning architecture between the kernel computation as well as the classifier. Experimental results show that the proposed AKBR model outperforms existing state-of-the-art graph kernels and deep learning methods on standard graph benchmarks.
Optical Flow Estimation for Tiny Objects: New Problem, Specialized Benchmark, and Bioinspired Scheme
PDF ↗Optical flow is pivotal in video-based tasks, yet existing methods mostly focus on medium-/large-size objects, while underperforming when characterizing the motion of tiny objects. To bridge this gap, we introduce the On-off Time-delay with Hassenstein-Reichardt correlator (OTHR), a computationally efficient scheme inspired by the primate visual cortex's direction selectivity mechanism. OTHR kernels, applied across multiple frames, discern bright/dark luminance changes along a specific direction over a time delay, effectively estimating motion of tiny objects amidst noise and static backgrounds. Notably, OTHR integrates seamlessly with leading deep learning flow estimation models such as RAFT and FlowFormer. We also propose refined evaluation metrics for tiny objects and contribute a new dataset featuring such objects to aid algorithm development. Our experiments confirm OTHR's superiority over competing methods, particularly in enhancing state-of-the-art models' performance on tiny object motion estimation at minimal cost. Specifically, for objects less than 100 pixels, OTHR reduces RAFT and FlowFormer's errors by 22.03% and 83.50%, respectively. The codes will be accessible at https://github.com/JaneEliot/OTHR.
In order to boost the performance of a convolutional neural network (CNN), several approaches have shown the benefit of enhancing the spatial encoding of feature maps. However, few works paid attention to the positional properties of convolutional kernels. In this paper, we demonstrate that different kernel positions are of different importance, which depends on the task, dataset and architecture, and adaptively emphasizing the informative parts in convolutional kernels can lead to considerable improvement. Therefore, we propose a novel structural re-parameterization Position Boosting Convolution (PBConv) to exploit and enhance the position information in the convolutional kernel. PBConv consists of several concurrent small convolutional kernels, which can be equivalently converted to the original kernel and bring no extra inference cost. Different from existing structural re-parameterization methods, PBconv searches for the optimal re-parameterized structure by a fast heuristic algorithm based on the dispersion of kernel weights. Such heuristic search is efficient yet effective, well adapting the varying kernel weight distribution. As a result, PBConv can significantly improve the representational power of a model, especially its ability to extract fine-grained low-level features. Importantly, PBConv is orthogonal to procedural re-parameterization methods and can further boost performance based on them.
LensNet: An End-to-End Learning Framework for Empirical Point Spread Function Modeling and Lensless Imaging Reconstruction
PDF ↗Lensless imaging stands out as a promising alternative to conventional lens-based systems, particularly in scenarios demanding ultracompact form factors and cost-effective architectures. However, such systems are fundamentally governed by the Point Spread Function (PSF), which dictates how a point source contributes to the final captured signal. Traditional lensless techniques often require explicit calibrations and extensive pre-processing, relying on static or approximate PSF models. These rigid strategies can result in limited adaptability to real-world challenges, including noise, system imperfections, and dynamic scene variations, thus impeding high-fidelity reconstruction. In this paper, we propose LensNet, an end-to-end deep learning framework that integrates spatial-domain and frequency-domain representations in a unified pipeline. Central to our approach is a learnable Coded Mask Simulator (CMS) that enables dynamic, data-driven estimation of the PSF during training, effectively mitigating the shortcomings of fixed or sparsely calibrated kernels. By embedding a Wiener filtering component, LensNet refines global structure and restores fine-scale details, thus alleviating the dependency on multiple handcrafted pre-processing steps. Extensive experiments demonstrate LensNet's robust performance and superior reconstruction quality compared to state-of-the-art methods, particularly in preserving high-frequency details and attenuating noise. The proposed framework establishes a novel convergence between physics-based modeling and data-driven learning, paving the way for more accurate, flexible, and practical lensless imaging solutions for applications ranging from miniature sensors to medical diagnostics. The link of code is https://github.com/baijiesong/Lensnet.
L2M2: A Hierarchical Framework Integrating Large Language Model and Multi-agent Reinforcement Learning
PDF ↗Multi-agent reinforcement learning (MARL) has demonstrated remarkable success in collaborative tasks, yet faces significant challenges in scaling to complex scenarios requiring sustained planning and coordination across long horizons. While hierarchical approaches help decompose these tasks, they typically rely on hand-crafted subtasks and domain-specific knowledge, limiting their generalizability. We present L2M2, a novel hierarchical framework that leverages large language models (LLMs) for high-level strategic planning and MARL for low-level execution. L2M2 enables zero-shot planning that supports both end-to-end training and direct integration with pre-trained MARL models. Experiments in the VMAS environment demonstrate that L2M2's LLM-guided MARL achieves superior performance while requiring less than 20% of the training samples compared to baseline methods. In the MOSMAC environment, L2M2 demonstrates strong performance with pre-defined subgoals and maintains substantial effectiveness without subgoals - scenarios where baseline methods consistently fail. Analysis through kernel density estimation reveals L2M2's ability to automatically generate appropriate navigation plans, demonstrating its potential for addressing complex multi-agent coordination tasks.
Time sequences are essential in fields such as finance, healthcare, and environmental science, where understanding temporal dependencies and making accurate predictions are crucial. These sequences often exhibit complexities like nonlinearity, noise, and concept drift. Traditional models struggle to capture the intricate dynamics of multivariate and co-evolving sequences, particularly in contexts where relationships between variables shift unpredictably. This thesis introduces a range of Kernel Representation Learning (KRL) methodologies to address these challenges. We develop kernel self-representation learning to capture the temporal dependencies and hidden structures, while identifying concept drift in co-evolving sequences. Additionally, we explore theoretical connections between KRL and advanced deep-learning models. The proposed methods are validated through real-world applications, showing improvements in predictive accuracy, interpretability, and robustness.
We develop a method for the efficient verification of neural networks against convolutional perturbations such as blurring or sharpening. To define input perturbations, we use well-known camera shake, box blur and sharpen kernels. We linearly parameterise these kernels in a way that allows for a variation of the perturbation strength while preserving desired kernel properties. To facilitate their use in neural network verification, we develop an efficient way of convolving a given input with the parameterised kernels. The result of this convolution can be used to encode the perturbation in a verification setting by prepending a linear layer to a given network. This leads to tight bounds and a high effectiveness in the resulting verification step. We add further precision by employing input splitting as a branching strategy. We demonstrate that we are able to verify robustness on a number of standard benchmarks where the baseline is unable to provide any safety certificates. To the best of our knowledge, this is the first solution for verifying robustness against specific convolutional perturbations such as camera shake.
Serving LLMs requires substantial memory due to the storage requirements of Key-Value (KV) embeddings in the KV cache, which grows with sequence length. An effective approach to compress KV cache is quantization. However, traditional quantization methods face significant memory overhead due to the need to store quantization constants (at least a zero point and a scale) in full precision per data block. Depending on the block size, this overhead can add 1 or 2 bits per quantized number. We introduce QJL, a new quantization approach that consists of a Johnson-Lindenstrauss (JL) transform followed by sign-bit quantization. In contrast to existing methods, QJL eliminates memory overheads by removing the need for storing quantization constants. We propose an asymmetric estimator for the inner product of two vectors and demonstrate that applying QJL to one vector and a standard JL transform without quantization to the other provides an unbiased estimator with minimal distortion. We have developed an efficient implementation of the QJL sketch and its corresponding inner product estimator, incorporating a lightweight CUDA kernel for optimized computation. When applied across various LLMs and NLP tasks to quantize the KV cache to only 3 bits, QJL demonstrates a more than fivefold reduction in KV cache memory usage without compromising accuracy, all while achieving faster runtime.
In practical applications, it is often necessary to transfer knowledge from large pretrained models to small ones with various architectures for tackling different tasks. The Learngene framework, proposed recently, firstly extracts one compact module termed as learngene from a large well-trained model, after which learngene is used to build descendant models for handling diverse tasks. In this paper, we aim to explore extracting and inheriting learngene which can be generalized across different model architectures and tasks, remaining understudied in previous works. Inspired by the existing observations that large kernel convolutional neural networks (CNNs) exhibit significant generalization potential across various architectures and tasks, we propose a novel two-stage Learngene method termed CLKG (Convolutional Learngene for Knowledge Generalization), which inherits convolutional kernels containing generalized knowledge as learngene to build diverse models for multiple tasks. Specifically, we construct an auxiliary model comprised of small kernels and train it through dense feature distillation to inherit the feature extraction ability from large kernel CNNs. After distillation, we select certain kernels from the auxiliary model as learngene based on three criteria: direct kernel extraction, priority to edge kernels, and continuous kernel selection. Subsequently, we adapt learngene according to the width of the descendant models and use it to initialize the backbone part of descendant models. Experiments on diverse vision tasks such as image classification, object detection and semantic segmentation demonstrate the superiority of CLKG. For example, compared with from scratch training, it brings 2.89% improvements on VOC12+SBD, and reduces around 2x training data volume and training epochs to achieve better results. Furthermore, compared to knowledge distillation method, CLKG significantly reduces negative transfer on certain datasets, e.g., achieves 1.88% performance improvements on NAO dataset despite domain differences.
Hyperdimensional computing (HDC) is an approach from the cognitive science literature for solving information processing tasks using data represented as high-dimensional random vectors. The technique has a rigorous mathematical backing, and is easy to implement in energy-efficient and highly parallel hardware like FPGAs and "processing-in-memory" architectures. The effectiveness of HDC in machine learning largely depends on how raw data is mapped to high-dimensional space. In this work, we propose NysHD, a new method for constructing this mapping that is based on the Nyström method from the literature on kernel approximation. Our approach provides a simple recipe to turn any user-defined positive-semidefinite similarity function into an equivalent mapping in HDC. There is a vast literature on the design of such functions for learning problems. Our approach provides a mechanism to import them into the HDC setting, expanding the types of problems that can be tackled using HDC. Empirical evaluation against existing HDC encoding methods shows that NysHD can achieve, on average, 11% and 17% better classification accuracy on graph and string datasets respectively.
Incomplete Multi-View Clustering (IMVC) has made significant progress by optimally merging multiple pre-specified incomplete views. Most existing IMVC algorithms operate under the assumption that view alignment is known, but in practice, the coupling information between views may be absent, thereby limiting the practical applicability of these methods. Being aware of this, we propose a novel IMVC method named Kernel cOupling And eLement imputAtion induced Multi-View Clustering (KOALA), which sufficiently explores the nonlinear relationship among features and optimally processes a group of kernels with missing and unaligned elements to simultaneously resolve multi-view clustering problem under both uncoupled and incomplete scenarios. Specifically, we first introduce a cross-kernel alignment learning strategy to reconstruct the coupling relationships among multiple kernels, which effectively captures high-order nonlinear relationships among samples and enhances alignment accuracy. Additionally, a low-rank tensor constraint is imposed on the optimizable alignment kernel tensor, facilitating the effective imputation of missing kernel elements by leveraging consistency information across views. Subsequently, we develop an alternative optimization approach with promising convergence to solve the resultant optimization problem. Extensive experimental results on various multi-view datasets demonstrate that the KOALA method achieves remarkable clustering performance.
The diversity of time series applications and scarcity of domain-specific data highlight the need for time-series models with strong few-shot learning capabilities. In this work, we propose a novel training scheme and a transformer-based architecture, collectively referred to as TimePFN, for multivariate time-series (MTS) forecasting. TimePFN is based on the concept of Prior-data Fitted Networks (PFN), which aims to approximate Bayesian inference. Our approach consists of (1) generating synthetic MTS data through diverse Gaussian process kernels and the linear coregionalization method, and (2) a novel MTS architecture capable of utilizing both temporal and cross-channel dependencies across all input patches. We evaluate TimePFN on several benchmark datasets and demonstrate that it outperforms the existing state-of-the-art models for MTS forecasting in both zero-shot and few-shot settings. Notably, fine-tuning TimePFN with as few as 500 data points nearly matches full dataset training error, and even 50 data points yield competitive results. We also find that TimePFN exhibits strong univariate forecasting performance, attesting to its generalization ability. Overall, this work unlocks the power of synthetic data priors for MTS forecasting and facilitates strong zero- and few-shot forecasting performance.
In this work, we consider an online robust Markov Decision Process (MDP) where we have the information of finitely many prototypes of the underlying transition kernel. We consider an adaptively updated ambiguity set of the prototypes and propose an algorithm that efficiently identifies the true underlying transition kernel while guaranteeing the performance of the corresponding robust policy. To be more specific, we provide a sublinear regret of the subsequent optimal robust policy. We also provide an early stopping mechanism and a worst-case performance bound of the value function. In numerical experiments, we demonstrate that our method outperforms existing approaches, particularly in the early stage with limited data. This work contributes to robust MDPs by considering possible prior information about the underlying transition probability and online learning, offering both theoretical insights and practical algorithms for improved decision-making under uncertainty.