.png){kind=icon}
返回
Enhancing Active Learning With Semi-Supervised Loss Prediction Modules
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2024.3449449.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
In recent years, deep learning techniques have substantially improved across numerous applications. However, these advancements come with the requirement for extensive human-labeled data, leading to high labeling costs. Active learning addresses this issue by reducing labeling costs. The primary goal of active learning is to select the most informative data for model training. Among the various active learning strategies, uncertainty-based methods are most prevalent. These methods focus on data points where the model's predictions are uncertain, identifying them as key candidates for additional learning. Recently, uncertainty-based methods have been integrated into deep learning, mainly through learning loss-based methods. These methods involve training a loss prediction module to predict learning loss, using the predicted loss as an indicator of uncertainty. While learning loss-based methods generally perform well, they struggle with more challenging tasks. A notable limitation of previous learning loss-based methods is that they use the same labeled data to train the loss prediction module and the task model. However, since the actual data to be predicted is unlabeled, a loss prediction module trained exclusively on labeled data has inherent limitations. In this study, we propose a new training method for the loss prediction module that is robust to unlabeled data. Our approach employs semi-supervised learning techniques to incorporate unlabeled datasets into the training process of the loss prediction module. Experimental results demonstrate that our proposed method effectively trains a robust loss prediction module, achieving state-of-the-art performance across various datasets.
Keyword:
Training data
Data models
Uncertainty
Labeling
Semisupervised learning
Deep learning
Learning systems
Active learning
data efficient learning
learning loss based method
pool-based sampling
期刊
IF:
3.6
论文数:
9.8W
被引数:
29.4W
机构
引用论文
miRNA-Based Potential Biomarkers and New Molecular Insights in Ulcerative Colitis溃疡性结肠炎中基于miRNA的潜在生物标志物和新分子见解OAAI
OAAI分享收藏Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The LUNA16 challenge计算机断层扫描图像中自动检测肺结节的算法的验证，比较和组合: LUNA16挑战.png)
MEDICAL IMAGE ANALYSIS
OAAI
MEDICAL IMAGE ANALYSISIF11.8
OAAI分享收藏Mechanisms of Disease: local renin–angiotensin–aldosterone systems and the pathogenesis and treatment of cardiovascular disease疾病机制: 局部肾素-血管紧张素-醛固酮系统与心血管疾病的发病机制和治疗
PREAI
分享收藏Foundations for Restraining Bolts: Reinforcement Learning with LTLf/LDLf Restraining SpecificationsOAAI
OAAI分享收藏分享收藏
查看更多内容