A5049942254- Multimodal Machine Learning Applications
- Domain Adaptation and Few-Shot Learning
- Advanced Image and Video Retrieval Techniques
- Cancer-related cognitive impairment studies
- Glioma Diagnosis and Treatment
- Dementia and Cognitive Impairment Research
- QR Code Applications and Technologies
- Generative Adversarial Networks and Image Synthesis
- Cancer-related molecular mechanisms research
- Image Retrieval and Classification Techniques
- Web Data Mining and Analysis
- Brain Metastases and Treatment
- Digital Marketing and Social Media
- Functional Brain Connectivity Studies
- Recommender Systems and Techniques
Alibaba Group (China)
2022
Chang Gung University
2020-2021
University of Electronic Science and Technology of China
2020-2021
Peking University
2019-2020
Recently, generative adversarial network (GAN) has shown its strong ability on modeling data distribution via learning. Cross-modal GAN, which attempts to utilize the power of GAN model cross-modal joint and learn compatible features, is becoming research hotspot. However, existing approaches typically 1) require labeled multimodal massive labor cost establish correlation; 2) vanilla that results in unstable training procedure meaningless synthetic features; 3) lack extensibility for...
Recently, a series of deep supervised hashing methods were proposed for binary code learning. However, due to the high computation cost and limited hardware's memory, these will first select subset from training set, then form mini-batch data update network in each iteration. Therefore, remaining labeled cannot be fully utilized model directly obtain codes entire set retrieval. To address problems, this paper proposes an interesting regularized seamlessly integrate advantages efficient...
Due to the inconsistent distributions and representations of different modalities (e.g., images texts), it is very challenging correlate such heterogeneous data. A standard solution construct one common subspace, where are generated bridge heterogeneity gap. Existing methods based on representation learning mostly adopt a less effective two-stage paradigm: first, generating separate for each modality by exploiting modality-specific properties as complementary information, then capturing...
Conventional cross-modal retrieval models mainly assume the same scope of classes for both training set and testing set. This assumption limits their extensibility on zero-shot (ZS-CMR), where consists unseen that are disjoint with seen in The ZS-CMR task is more challenging due to heterogeneous distributions different modalities semantic inconsistency between classes. A few recently proposed approaches inspired by learning estimate distribution underlying multimodal data generative make...
Zero-Shot Cross-Modal Retrieval (ZS-CMR) is an emerging research hotspot that aims to retrieve data of new classes across different modality data. It challenging for not only the heterogeneous distributions modalities, but also inconsistent semantics seen and unseen classes. A handful recently proposed methods typically borrow idea from zero-shot learning, i.e., exploiting word embeddings class labels (i.e., class-embeddings) as common semantic space, using generative adversarial network...
Treatment modalities for breast cancer, the leading cause of cancer-related deaths in women worldwide, include surgery, radiotherapy, adjuvant chemotherapy, targeted therapy, and hormonal therapy. The advancement medical technology has facilitated substantial reduction cancer mortality. However, patients may experience cognitive impairment after chemotherapy. This phenomenon called chemotherapy-induced (i.e., "chemobrain") is common among survivors. function deficits exist before...
The goal of cross-modal retrieval is to search for semantically similar instances in one modality by using a query from another modality. Existing approaches mainly consider the standard scenario that requires source set training and target testing share same scope classes. However, they may not generalize well on zero-shot (ZS-CMR) task, where contains unseen classes are disjoint with seen set. This task more challenging due 1) absence during training, 2) inconsistent semantics across...
In e-commerce, ad creatives play an important role in effectively delivering product information to users. The purpose of online creative selection is learn users' preferences for creatives, and select the most appealing design users maximize Click-Through Rate (CTR). However, existing common practices industry usually place after ranking stage, thus optimal fails reflect influence on stage. To address these issues, we propose a novel Cascade Architecture Creative Selection (CACS), which...
Breast cancer is the most common female worldwide, and breast accounts for 30% of cancers. Of all treatment modalities, survivors who have undergone chemotherapy might complain about cognitive impairment during after treatment. This phenomenon, chemo-brain, used to describe alterations in functions receiving systemic chemotherapy. Few reports detect chemotherapy-induced (CICI) by performing functional MRI (fMRI) a deep learning analysis. In this study, we recruited 55 postchemotherapy (C+...
Our goal was to establish objective 3D deep learning models that differentiate cerebral alterations based on the effect of chemotherapy and visualize pattern recognized by our model. The average performance SE-ResNet-50 accuracy 80%, precision 78%, 70% recall, SE-DenseNet-121 model reached identical results with an 80% accuracy, 86% precision, recall. regions greatest contributions highlighted integrated gradients algorithm for differentiating chemo-brain were default mode dorsal attention...