Author name cluster

Jun Kong

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.

4 papers

1 author row

AAAI Conference 2025 Conference Paper

Sample-aware Adaptive Structured Pruning for Large Language Models

Jun Kong
Xinge Ma
Jin Wang
Xuejie Zhang

Large language models (LLMs) have achieved outstanding performance in natural language processing, but enormous model sizes and high computational costs limit their practical deployment. Structured pruning can effectively reduce the resource demands for deployment by removing redundant model parameters. However, the randomly selected calibration data and fixed single importance estimation metrics in existing structured pruning methods lead to degraded performance of pruned models. This study introduces AdaPruner, a sample-aware adaptive structured pruning framework for LLMs, aiming to optimize the calibration data and importance estimation metrics in the structured pruning process. Specifically, AdaPruner effectively removes redundant parameters from LLMs by constructing a structured pruning solution space and then employing Bayesian optimization to adaptively search for the optimal calibration data and importance estimation metrics. Experimental results show that the AdaPruner outperforms existing structured pruning methods on a family of LLMs with varying pruning ratios, demonstrating its applicability and robustness. Remarkably, at a 20% pruning ratio, the model pruned with AdaPruner maintains 97% of the performance of the unpruned model.

PDF Details DOI

AAAI Conference 2025 Conference Paper

Temporal Action Localization with Cross Layer Task Decoupling and Refinement

Qiang Li
Di Liu
Jun Kong
Sen Li
Hui Xu
Jianzhong Wang

Temporal action localization (TAL) involves dual tasks to classify and localize actions within untrimmed videos. However, the two tasks often have conflicting requirements for features. Existing methods typically employ separate heads for classification and localization tasks but share the same input feature, leading to suboptimal performance. To address this issue, we propose a novel TAL method with Cross Layer Task Decoupling and Refinement (CLTDR). Based on the feature pyramid of video, CLTDR strategy integrates semantically strong features from higher pyramid layers and detailed boundary-aware boundary features from lower pyramid layers to effectively disentangle the action classification and localization tasks. Moreover, the multiple features from cross layers are also employed to refine and align the disentangled classification and regression results. At last, a lightweight Gated Multi-Granularity (GMG) module is proposed to comprehensively extract and aggregate video features at instant, local, and global temporal granularities. Benefiting from the CLTDR and GMG modules, our method achieves state-of-the-art performance on five challenging benchmarks: THUMOS14, MultiTHUMOS, EPIC-KITCHENS-100, ActivityNet-1.3, and HACS. Code：https://github.com/LiQiang0307/CLTDR-GMG

PDF Details DOI

EAAI Journal 2024 Journal Article

Self-ensembling with mask-boundary domain adaptation for optic disc and cup segmentation

Yanlin He
Jun Kong
Di Liu
Juan Li
Caixia Zheng

Due to different retinal fundus image acquisition devices having various imaging principles, domain shift often occurs between different datasets. Hence, a segmentation network well-trained on one dataset (i. e. , source domain) usually obtains very poor performance on another dataset (i. e. , target domain), which results in us having to annotate the new dataset (target domain) to train the segmentation network again. However, annotating a new dataset is usually time-consuming and laborious. To address this problem, we proposed a novel unsupervised domain adaptation method for optic disc and cup segmentation. To be specific, we first utilized a domain adaptation method based on self-ensembling to effectively align the features of the source domain and target domain. Then, we designed a novel backbone network (MBU-Net) to make full use of the mask and boundary information to improve the segmentation performance of self-ensembling. Finally, we proposed an output-level adversarial domain adaptation (OADA) to address the domain shift problem of the structured output space in self-ensembling. In experiments, we test our proposed method on three different target domain datasets including Target Domain 1 (RIM-ONE_r3 dataset), Target Domain 2 (Drishti-GS dataset) and Target Domain 3 (REFUGE dataset). The experimental results demonstrate that our proposed method outperforms the compared state-of-the-art methods in the optic disc and cup segmentation tasks.

Details DOI

AIIM Journal 2021 Journal Article

A novel deep learning model DDU-net using edge features to enhance brain tumor segmentation on MR images

Min Jiang
Fuhao Zhai
Jun Kong

Glioma is a relatively common brain tumor disease with high mortality rate. Humans have been seeking a more effective therapy. In the course of treatment, the specific location of the tumor needs to be determined first in any case. Therefore, how to segment tumors from brain tissue accurately and quickly is a persistent problem. In this paper, a new dual-stream decoding CNN architecture combined with U-net for automatic segmentation of brain tumor on MR images namely DDU-net is proposed. Two edge-based optimization strategies are used to enhance the performance of brain tumor segmentation. First, we design a separate branch to process edge stream information. Here, high level edge features are reduced in dimension of channel and integrated into the conventional semantic stream in the way of residual. Second, a regularization loss function is used to encourage the predicted segmentation mask to align with ground truth around the edge mainly by penalizing pixels where the predicted segmentation masks and labels do not match around the edge. In training, we employ a novel edge extraction algorithm for providing edge labels with higher quality. Moreover, we add a self-adaptive balancing class weight coefficient into the cross entropy loss function for solving the serious class imbalance problem in the backpropagation of edge extraction. Our experiments show that this leads to a very efficient architecture which can produce clearer prediction at the edge of the tumor. Our method achieves ideal performance on BraTS2017 and BraTS2018 in terms of Dice coefficient.

Details DOI