Access'17 an_ensemble_deep_learning_method_for_vehicle_type_classification_on_visual_traffic_surveillance_sensors
阅读时间:24.1.12
集成学习分类
- Bagging是bootstrap aggregating的缩写,which was proposed by Breiman [32] to improve the classification by combining prediction results of models trained independently on randomly generated training sets.
- Boosting 迭代训练弱学习器
- Bucket of models 选择适合集成的模型
模型方法
平衡采样用于增加原始数据集中稀有类的样本数量,训练一组初始化的CNN,通过maximum voting policy 将他们的结果组合。
细节
- 数据增强技术(翻转,平移等)。样本数量少的类别通过过度抽样随机增加。为了避免过拟合,数量并没有增加到和大类别的数量一样多,而是增加到一个小数字。
- CNN集合包括ResNet-50、ResNet-101和ResNet-152(都是ResNet种类的,多样性可能不足)
- 模型个数为奇数,避免同票现象
- 使用多个指标来评判,带有消融实验
This line appears after every note.
Notes mentioning this note
Paper Reading
0. Solid Ideas
Yoshua Bengio重新思考ML的投稿
1.Quantum Computing
[[TKDE’16_Relevance Feedback Algorithms Inspired By Quantum Detection]]