arXiv subjects

Junming Shao

Publications and source records attributed to Junming Shao.

13 recordsLinked to original sources

arXiv2026

SPEAR: A System for Post-Quantization Error-Adaptive Recovery Enabling Efficient Low-Bit LLM Serving

Efficient large language model (LLM) serving is increasingly constrained by deployment cost. Quantization is a key technique for reducing serving cost, yet even state-of-the-art 4-bit quantizers exhibit a noticeable quality gap from FP16, particularly for smaller models where low-bit serving is most beneficial. We identify a fundamental cause of this gap: quantization error is highly input-dependent and varies substantially across tokens, while existing post-quantization compensation methods are static and apply identical corrections to all inputs. As a result, easy tokens are over-corrected while hard tokens remain under-corrected. We present SPEAR, a system for post-quantization error-adaptive recovery that improves low-bit LLM serving. SPEAR introduces lightweight Error Compensators (ECs) modulated by per-token gates and places them only at the most error-sensitive layers identified through a CKA-guided entropy-aware diagnostic. This focuses a small parameter budget where it is most effective. Efficient deployment of ECs presents several systems challenges, including additional computation, tensor-parallel synchronization caused by input-dependent gating, and latency instability across configurations. SPEAR addresses these issues through adaptive kernel-fusion dispatch, combining an epilogue-integrated peer-reduction kernel with P2P dual-write to fuse the post-EC computation into low-bit GEMMs, and an SLO-constrained EC-aware scheduler for predictable serving performance. Across challenging per-channel quantization settings, SPEAR recovers 56-75% of the perplexity gap between W4 and FP16 while adding less than 1% model memory overhead and maintaining latency comparable to a widely used 4-bit serving deployment.

cs.AR↗

arXiv2026

The Geometry of Compromise: Unlocking Generative Capabilities via Controllable Modality Alignment

Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the underlying distributional mismatch intact. We decompose the modality gap into a Centroid Gap and a Distribution Gap, and demonstrate that the Distribution Gap is the true predictor of cross-modal task quality ($R^2 = 0.986$), whereas the commonly used Raw Gap is misleading ($R^2 = 0.691$). Motivated by this observation, we propose TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment), a fine-tuning framework that explicitly reduces both components. The proposed CMA jointly mitigates centroid offsets and reshapes the distributional structure, while a three-phase curriculum with gradient-aware scheduling progressively introduces alignment during training to enable stable optimization. Experiments demonstrate that our method significantly improves cross-modal alignment. With $\alpha_{\text{target}}{=}0.05$, the modality gap is reduced by 66.6\% with only 4.84\% accuracy drop. Under stronger alignment ($\alpha_{\text{target}}{=}0.5$), the gap is reduced by 82.3\%, clustering ARI improves from 0.318 to 0.516, and captioning CIDEr increases by 57.1\% over the original model. Our code and pre-trained models will be made publicly available upon acceptance.

cs.CV↗

arXiv2024

Exploiting Fine-Grained Prototype Distribution for Boosting Unsupervised Class Incremental Learning

The dynamic nature of open-world scenarios has attracted more attention to class incremental learning (CIL). However, existing CIL methods typically presume the availability of complete ground-truth labels throughout the training process, an assumption rarely met in practical applications. Consequently, this paper explores a more challenging problem of unsupervised class incremental learning (UCIL). The essence of addressing this problem lies in effectively capturing comprehensive feature representations and discovering unknown novel classes. To achieve this, we first model the knowledge of class distribution by exploiting fine-grained prototypes. Subsequently, a granularity alignment technique is introduced to enhance the unsupervised class discovery. Additionally, we proposed a strategy to minimize overlap between novel and existing classes, thereby preserving historical knowledge and mitigating the phenomenon of catastrophic forgetting. Extensive experiments on the five datasets demonstrate that our approach significantly outperforms current state-of-the-art methods, indicating the effectiveness of the proposed method.

cs.LG↗

arXiv2023

Seeing Is Not Always Believing: Invisible Collision Attack and Defence on Pre-Trained Models

Large-scale pre-trained models (PTMs) such as BERT and GPT have achieved great success in diverse fields. The typical paradigm is to pre-train a big deep learning model on large-scale data sets, and then fine-tune the model on small task-specific data sets for downstream tasks. Although PTMs have rapidly progressed with wide real-world applications, they also pose significant risks of potential attacks. Existing backdoor attacks or data poisoning methods often build up the assumption that the attacker invades the computers of victims or accesses the target data, which is challenging in real-world scenarios. In this paper, we propose a novel framework for an invisible attack on PTMs with enhanced MD5 collision. The key idea is to generate two equal-size models with the same MD5 checksum by leveraging the MD5 chosen-prefix collision. Afterwards, the two ``same" models will be deployed on public websites to induce victims to download the poisoned model. Unlike conventional attacks on deep learning models, this new attack is flexible, covert, and model-independent. Additionally, we propose a simple defensive strategy for recognizing the MD5 chosen-prefix collision and provide a theoretical justification for its feasibility. We extensively validate the effectiveness and stealthiness of our proposed attack and defensive method on different models and data sets.

cs.CR↗

arXiv2023

Fine-tuning Happens in Tiny Subspaces: Exploring Intrinsic Task-specific Subspaces of Pre-trained Language Models

Pre-trained language models (PLMs) are known to be overly parameterized and have significant redundancy, indicating a small degree of freedom of the PLMs. Motivated by the observation, in this paper, we study the problem of re-parameterizing and fine-tuning PLMs from a new perspective: Discovery of intrinsic task-specific subspace. Specifically, by exploiting the dynamics of the fine-tuning process for a given task, the parameter optimization trajectory is learned to uncover its intrinsic task-specific subspace. A key finding is that PLMs can be effectively fine-tuned in the subspace with a small number of free parameters. Beyond, we observe some outlier dimensions emerging during fine-tuning in the subspace. Disabling these dimensions degrades the model performance significantly. This suggests that these dimensions are crucial to induce task-specific knowledge to downstream tasks.

cs.CL↗

arXiv2023

Open-world Semi-supervised Novel Class Discovery

Traditional semi-supervised learning tasks assume that both labeled and unlabeled data follow the same class distribution, but the realistic open-world scenarios are of more complexity with unknown novel classes mixed in the unlabeled set. Therefore, it is of great challenge to not only recognize samples from known classes but also discover the unknown number of novel classes within the unlabeled data. In this paper, we introduce a new open-world semi-supervised novel class discovery approach named OpenNCD, a progressive bi-level contrastive learning method over multiple prototypes. The proposed method is composed of two reciprocally enhanced parts. First, a bi-level contrastive learning method is introduced, which maintains the pair-wise similarity of the prototypes and the prototype group levels for better representation learning. Then, a reliable prototype similarity metric is proposed based on the common representing instances. Prototypes with high similarities will be grouped progressively for known class recognition and novel class discovery. Extensive experiments on three image datasets are conducted and the results show the effectiveness of the proposed method in open-world scenarios, especially with scarce known classes and labels.

cs.CV↗

arXiv2022

An Interpretable Neuron Embedding for Static Knowledge Distillation

Although deep neural networks have shown well-performance in various tasks, the poor interpretability of the models is always criticized. In the paper, we propose a new interpretable neural network method, by embedding neurons into the semantic space to extract their intrinsic global semantics. In contrast to previous methods that probe latent knowledge inside the model, the proposed semantic vector externalizes the latent knowledge to static knowledge, which is easy to exploit. Specifically, we assume that neurons with similar activation are of similar semantic information. Afterwards, semantic vectors are optimized by continuously aligning activation similarity and semantic vector similarity during the training of the neural network. The visualization of semantic vectors allows for a qualitative explanation of the neural network. Moreover, we assess the static knowledge quantitatively by knowledge distillation tasks. Empirical experiments of visualization show that semantic vectors describe neuron activation semantics well. Without the sample-by-sample guidance from the teacher model, static knowledge distillation exhibit comparable or even superior performance with existing relation-based knowledge distillation methods.

cs.LG↗

arXiv2019

Large-scale Multi-view Subspace Clustering in Linear Time

A plethora of multi-view subspace clustering (MVSC) methods have been proposed over the past few years. Researchers manage to boost clustering accuracy from different points of view. However, many state-of-the-art MVSC algorithms, typically have a quadratic or even cubic complexity, are inefficient and inherently difficult to apply at large scales. In the era of big data, the computational issue becomes critical. To fill this gap, we propose a large-scale MVSC (LMVSC) algorithm with linear order complexity. Inspired by the idea of anchor graph, we first learn a smaller graph for each view. Then, a novel approach is designed to integrate those graphs so that we can implement spectral clustering on a smaller graph. Interestingly, it turns out that our model also applies to single-view scenario. Extensive experiments on various large-scale benchmark data sets validate the effectiveness and efficiency of our approach with respect to state-of-the-art clustering methods.

cs.LG↗

arXiv2017

Preference Modeling by Exploiting Latent Components of Ratings

Understanding user preference is essential to the optimization of recommender systems. As a feedback of user's taste, rating scores can directly reflect the preference of a given user to a given product. Uncovering the latent components of user ratings is thus of significant importance for learning user interests. In this paper, a new recommendation approach, called LCR, was proposed by investigating the latent components of user ratings. The basic idea is to decompose an existing rating into several components via a cost-sensitive learning strategy. Specifically, each rating is assigned to several latent factor models and each model is updated according to its predictive errors. Afterwards, these accumulated predictive errors of models are utilized to decompose a rating into several components, each of which is treated as an independent part to retrain the latent factor models. Finally, all latent factor models are combined linearly to estimate predictive ratings for users. In contrast to existing methods, LCR provides an intuitive preference modeling strategy via multiple component analysis at an individual perspective. Meanwhile, it is verified by the experimental results on several benchmark datasets that the proposed method is superior to the state-of-art methods in terms of recommendation accuracy.

cs.IR↗

arXiv2017

Learning in High-Dimensional Multimedia Data: The State of the Art

During the last decade, the deluge of multimedia data has impacted a wide range of research areas, including multimedia retrieval, 3D tracking, database management, data mining, machine learning, social media analysis, medical imaging, and so on. Machine learning is largely involved in multimedia applications of building models for classification and regression tasks etc., and the learning principle consists in designing the models based on the information contained in the multimedia dataset. While many paradigms exist and are widely used in the context of machine learning, most of them suffer from the `curse of dimensionality', which means that some strange phenomena appears when data are represented in a high-dimensional space. Given the high dimensionality and the high complexity of multimedia data, it is important to investigate new machine learning algorithms to facilitate multimedia data analysis. To deal with the impact of high dimensionality, an intuitive way is to reduce the dimensionality. On the other hand, some researchers devoted themselves to designing some effective learning schemes for high-dimensional data. In this survey, we cover feature transformation, feature selection and feature encoding, three approaches fighting the consequences of the curse of dimensionality. Next, we briefly introduce some recent progress of effective learning algorithms. Finally, promising future trends on multimedia learning are envisaged.

cs.CV↗

arXiv2016

Graph Clustering with Density-Cut

How can we find a good graph clustering of a real-world network, that allows insight into its underlying structure and also potential functions? In this paper, we introduce a new graph clustering algorithm Dcut from a density point of view. The basic idea is to envision the graph clustering as a density-cut problem, such that the vertices in the same cluster are densely connected and the vertices between clusters are sparsely connected. To identify meaningful clusters (communities) in a graph, a density-connected tree is first constructed in a local fashion. Owing to the density-connected tree, Dcut allows partitioning a graph into multiple densely tight-knit clusters directly. We demonstrate that our method has several attractive benefits: (a) Dcut provides an intuitive criterion to evaluate the goodness of a graph clustering in a more natural and precise way; (b) Built upon the density-connected tree, Dcut allows identifying the meaningful graph clusters of densely connected vertices efficiently; (c) The density-connected tree provides a connectivity map of vertices in a graph from a local density perspective. We systematically evaluate our new clustering approach on synthetic as well as real data to demonstrate its good performance.

cs.SI↗

arXiv2015

Online Social Activity Reflects Economic Status

To characterize economic development and diagnose the economic health condition, several popular indices such as gross domestic product (GDP), industrial structure and income growth are widely applied. However, computing these indices based on traditional economic census is usually costly and resources consuming, and more importantly, following a long time delay. In this paper, we analyzed nearly 200 million users' activities for four consecutive years in the largest social network (Sina Microblog) in China, aiming at exploring latent relationships between the online social activities and local economic status. Results indicate that online social activity has a strong correlation with local economic development and industrial structure, and more interestingly, allows revealing the macro-economic structure instantaneously with nearly no cost. Beyond, this work also provides a new venue to identify risky signal in local economic structure.

physics.soc-ph↗

arXiv2014

Community Detection via Local Dynamic Interaction

How can we uncover the natural communities in a real network that allows insight into its underlying structure and also potential functions? In this paper, we introduce a new community detection algorithm, called Attractor, which automatically spots the communities or groups in a network over time via local dynamic interaction. The basic idea is to envision a network as a dynamical system, and each agent interacts with its local partners. Instead of investigating the node dynamics, we actually examine the change of "distances" among linked nodes. As time evolves, these distances will be shrunk or stretched gradually based on their topological structures. Finally all distances among linked nodes will converge into a stable pattern, and communities can be intuitively identified. Thanks to the dynamic viewpoint of community detection, Attractor has several potential attractive properties: (a) Attractor provides an intuitive solution to analyze the community structure of a network, and faithfully captures the natural communities (with high quality). (b) Owing to its time complexity $O(|E|)$, Attractor allows finding communities on large networks. (c) The small communities or anomalies, usually existing in real-world networks, can be well pinpointed. (d) Attractor is easy to parameterize, since there is no need to specify the number of clusters. Extensive experiments on synthetic and real-world networks further demonstrate the effectiveness and efficiency of the proposed approach.

cs.SI↗