HHHu Han

Hu Han

*This is an algorithmically generated author record based on publicly available information.

Hu Han has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Self-Supervised Learning with Swin Transformers

    by Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, Han Hu

    We are witnessing a modeling shift from CNN to Transformers in computer vision. In this work, we present a self-supervised learning approach called MoBY, with Vision Transformers as its backbone architecture. The approach basically has no new inventions, which is combined from MoCo v2 and BYOL and tuned to achieve reasonably high accuracy on ImageNet-1K linear evaluation: 72.8% and 75.0% top-1 accuracy using DeiT-S and Swin-T, respectively, by 300-epoch training. The performance is slightly better than recent works of MoCo v3 and DINO which adopt DeiT as the backbone, but with much lighter tricks. More importantly, the general-purpose Swin Transformer backbone enables us to also evaluate the learnt representations on downstream tasks such as object detection and semantic segmentation, in contrast to a few recent approaches built on ViT/DeiT which only report linear evaluation results on ImageNet-1K due to ViT/DeiT not tamed for these dense prediction tasks. We hope our results can facilitate more comprehensive evaluation of self-supervised learning methods designed for Transformer architectures. Our code and models are available at https://github.com/SwinTransformer/Transformer-SSL, which will be continually enriched.


  • Improved Memory Network for Aspect Sentiment Analysis

    by Hu Han, Xu-Xu Bai, Jin Liu

    In order to explore impact of different memory modules under the framework of deep memory network for aspect level sentiment classification, three kinds of different memory modules are designed. One of them uses CNN to build a memory module, which is capable of capturing local information from original sentence. The other one of them uses BiLSTM to build another memory module, which is capable of capturing sequence information. And the last one of them uses CNN and BiLSTM, which combines both local and sequence information together, to build memory module at the same time. Experiments on laptop and restaurant datasets demonstrate that our three methods achieve better results than MemNet and feature based SVM approach.


  • Open Access

    Attention-Based Memory Network for Text Sentiment Classification

    by Hu Han, Jin Liu, Guoli Liu

    In order to explore the impact of different memory modules under the framework of memory network for aspect level sentiment classification, we use convolutional neural networks (CNN) and bidirectional long short term memory (BiLSTM) to design four kinds of memory network models in this paper. The first model uses CNN for building a memory module, which is capable of capturing local information in documents. The second uses BiLSTM for building another memory module, which captures sequence information in documents. At the same time, the following two models use CNN and BiLSTM for building memory module——one builds a hierarchical neural network by using CNN and BiLSTM for building memory module, which combines both local and sequence information together. The other, respectively, uses CNN and BiLSTM for building two memory modules, respectively, which captures local information and sequence information through different modules. And then, we combine the final representations, which are generated from the different memory networks, for further sentiment classification. Experiments on laptop and restaurant datasets demonstrate that our four methods achieving better results than MemNet. In particular, the latter two models achieve better performance than the first two models and feature-based support vector machine approach.