Author name cluster

Shu Li

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

8 papers

2 author rows

TIST Journal 2026 Journal Article

BUZZ: Beehive-structured Sparse KV Cache with Segmented Heavy Hitters for Efficient LLM Inference

Junqi Zhao
Zhijin Fang
Shu Li
Shaohui Yang
Shichao He

Large language models (LLMs) are critical in natural language processing but face challenges in inference speed and computational efficiency, hindering real-time applications. The key-value (KV) cache mechanism helps reduce computational overhead in transformer models; however, efficient contextual understanding remains problematic. In this paper, we introduce BUZZ, an innovative KV caching algorithm that leverages structured contextual information to optimize cache memory usage while enhancing inference speed. The core concept of BUZZ involves interval sampling of historically significant tokens to maintain sentence structure information, ensuring that KV Cache historical tokens are consistently distributed at nearly equal intervals. Tokens recently removed from the sliding window undergo local-max sampling based on attention values, preserving crucial contextual information. Additionally, we propose BUZZ with $\log n$, an extension that enhances performance under extreme compression and long-context settings. Evaluations on four real-world datasets—CNN/Daily Mail, XSUM, LongBench, Wikitext, and 10-QA—demonstrate that BUZZ (1) achieves a 2.5 $\times$ reduction in cache memory usage for LLM inference while maintaining over 99% accuracy in long-text summarization, and (2) surpasses state-of-the-art multi-document question answering by 7.69% under equivalent memory constraints, avoiding out-of-memory issues faced by full cache approaches. Furthermore, BUZZ achieves substantial inference speed improvements with a $\log{n}$ time complexity. The implementation of BUZZ is available at: https://github.com/JunqiZhao888/buzz-llm.

Details DOI

AAAI Conference 2026 Conference Paper

Unlearning in Cross-Modal Retrieval via Prior-Prototype Guided Partitioned Dampening

Yi Lu
Shu Li
Yurong Qian

Selective deletion of data from deep models, known as unlearning, has become crucial for enforcing the right to be forgotten, while also mitigating the negative impact of flawed training data. Retraining deep models is often impractical due to data access restrictions and computational overhead. Existing retraining-free methods are typically based on the Fisher Information Matrix (FIM), which quantifies the importance of model parameters with respect to forgetting classes, applying equal dampening to these parameters. This approach implicitly assumes a semantically uniform representation space, where all retained classes are equidistant from the forgetting classes. However, this assumption often fails in real-world cross-modal retrieval scenarios characterized by multi-label and non-orthogonal semantics. To overcome this limitation, we propose Prior-Prototype guided Partitioned dampening (PPP), an effective strategy for selective forgetting in cross-modal retrieval. First, PPP defines prior-prototypes, which are semantic centers derived from well-trained models, to identify neighbor classes semantically close to the forgetting set. Then, PPP uses Fisher information to identify parameters sensitive to forgetting and partitions them into buffer and core regions based on their relative importance to the neighbor and retained sets. Finally, PPP applies a hierarchical dampening strategy, where core parameters receive stronger suppression guided by prototype-based semantic disparities. Comprehensive evaluations on four large-scale benchmarks show that PPP performs competitively with retraining-based baselines, highlighting its effectiveness and generalizability in selective unlearning for cross-modal retrieval.

PDF Details DOI

AAAI Conference 2024 Conference Paper

Hidden Follower Detection: How Is the Gaze-Spacing Pattern Embodied in Frequency Domain?

Shu Li
Ruimin Hu
Suhui Li
Liang Liao

Spatiotemporal social behavior analysis is a technique that studies the social behavior patterns of objects and estimates their risks based on their trajectories. In social public scenarios such as train stations, hidden following behavior has become one of the most challenging issues due to its probability of evolving into violent events, which is more than 25%. In recent years, research on hidden following detection (HFD) has focused on differences in time series between hidden followers and normal pedestrians under two temporal characteristics: gaze and spatial distance. However, the time-domain representation for time series is irreversible and usually causes the loss of critical information. In this paper, we deeply study the expression efficiency of time/frequency domain features of time series, by exploring the recovery mechanism of features to source time series, we establish a fidelity estimation method for feature expression and a selection model for frequency-domain features based on the signal-to-distortion ratio (SDR). Experimental results demonstrate the feature fidelity of time series and HFD performance are positively correlated, and the fidelity of frequency-domain features and HFD performance are significantly better than the time-domain features. On both real and simulated datasets, the accuracy of the proposed method is increased by 3%, and the gaze-only module is improved by 10%. Related research has explored new methods for optimal feature selection based on fidelity, new patterns for efficient feature expression of hidden following behavior, and the mechanism of multimodal collaborative identification.

PDF Details DOI

EAAI Journal 2024 Journal Article

PASTS: Progress-aware spatio-temporal transformer speaker for vision-and-language navigation

Liuyi Wang
Chengju Liu
Zongtao He
Shu Li
Qingqing Yan
Huiyi Chen
Qijun Chen

Vision-and-language navigation (VLN) is a crucial but challenging cross-modal navigation task. One powerful technique to enhance the generalization performance in VLN is the use of an independent speaker model to provide pseudo instructions for data augmentation. However, current speaker models based on Long-Short Term Memory (LSTM) lack the ability to attend to features relevant at different locations and time steps. To address this, we propose a novel progress-aware spatio-temporal transformer speaker (PASTS) model that uses the transformer as the core of the network. PASTS uses a spatio-temporal encoder to fuse panoramic representations and encode intermediate connections through steps. Besides, to avoid the misalignment problem that could result in incorrect supervision, a speaker progress monitor (SPM) is proposed to enable the model to estimate the progress of instruction generation and facilitate more fine-grained caption results. Additionally, a multifeature dropout (MFD) strategy is introduced to alleviate overfitting. The proposed PASTS is flexible to be combined with existing VLN models. The experimental results demonstrate that PASTS outperforms previous speaker models and successfully improves the performance of previous VLN models, achieving state-of-the-art performance on the standard Room-to-Room (R2R) dataset.

Details DOI

ICLR Conference 2024 Conference Paper

Polynomial Width is Sufficient for Set Representation with High-dimensional Features

Peihao Wang
Shenghao Yang 0002
Shu Li
Zhangyang Wang
Pan Li 0005

Set representation has become ubiquitous in deep learning for modeling the inductive bias of neural networks that are insensitive to the input order. DeepSets is the most widely used neural network architecture for set representation. It involves embedding each set element into a latent space with dimension $L$, followed by a sum pooling to obtain a whole-set embedding, and finally mapping the whole-set embedding to the output. In this work, we investigate the impact of the dimension $L$ on the expressive power of DeepSets. Previous analyses either oversimplified high-dimensional features to be one-dimensional features or were limited to complex analytic activations, thereby diverging from practical use or resulting in $L$ that grows exponentially with the set size $N$ and feature dimension $D$. To investigate the minimal value of $L$ that achieves sufficient expressive power, we present two set-element embedding layers: (a) linear + power activation (LP) and (b) linear + exponential activations (LE). We demonstrate that $L$ being $\operatorname{poly}(N, D)$ is sufficient for set representation using both embedding layers. We also provide a lower bound of $L$ for the LP embedding layer. Furthermore, we extend our results to permutation-equivariant set functions and the complex field.

Details

AAAI Conference 2021 Conference Paper

Multi-View Representation Learning with Manifold Smoothness

Shu Li
Wei Wang
Wen-Tao Li
Pan Chen

Multi-view representation learning attempts to learn a representation from multiple views and most existing methods are unsupervised. However, representation learned only from unlabeled data may not be discriminative enough for further applications (e. g. , clustering and classification). For this reason, semi-supervised methods which could use unlabeled data along with the labeled data for multi-view representation learning need to be developed. Manifold information plays an important role in semi-supervised learning, but it has not been considered for multi-view representation learning. In this paper, we introduce the manifold smoothness into multiview representation learning and propose MvDGAT which learns the representation and the intrinsic manifold simultaneously with graph attention network. Experiments conducted on real-world datasets reveal that our MvDGAT can achieve better performance than state-of-the-art methods.

PDF Details

AAAI Conference 2020 Conference Paper

Co-GCN for Multi-View Semi-Supervised Learning

Shu Li
Wen-Tao Li
Wei Wang

In many real-world applications, the data have several disjoint sets of features and each set is called as a view. Researchers have developed many multi-view learning methods in the past decade. In this paper, we bring Graph Convolutional Network (GCN) into multi-view learning and propose a novel multi-view semi-supervised learning method Co-GCN by adaptively exploiting the graph information from the multiple views with combined Laplacians. Experimental results on real-world data sets verify that Co-GCN can achieve better performance compared with state-of-the-art multi-view semisupervised methods.

PDF Details

IJCAI Conference 2019 Conference Paper

MiSC: Mixed Strategies Crowdsourcing

Ching Yun Ko
Rui Lin
Shu Li
Ngai Wong

Popular crowdsourcing techniques mostly focus on evaluating workers' labeling quality before adjusting their weights during label aggregation. Recently, another cohort of models regard crowdsourced annotations as incomplete tensors and recover unfilled labels by tensor completion. However, mixed strategies of the two methodologies have never been comprehensively investigated, leaving them as rather independent approaches. In this work, we propose MiSC ( Mixed Strategies Crowdsourcing), a versatile framework integrating arbitrary conventional crowdsourcing and tensor completion techniques. In particular, we propose a novel iterative Tucker label aggregation algorithm that outperforms state-of-the-art methods in extensive experiments.

PDF Details