- 博客(5)
- 资源 (1)
- 收藏
- 关注
原创 【论文精读】Learn from Neighbour: A Curriculum that Train Low Weighted Samples By Imitation
Learn from Neighbour: A Curriculum that Train Low Weighted Samples By Imitation原文地址:Learn from Neighbour: A Curriculum that Train Low Weighted Samples By Imitation这篇文章提出了一种新的课程学习的策略(见论文第4节),主要思想就是,一...
2018-10-09 16:13:45 349
原创 【论文精读】Select Via Proxy: Efficient Data Selection For Training DeepNetworks
Select Via Proxy: Efficient Data Selection For Training DeepNetworks2019ICLR的文章,介绍了Select Via Proxy(SVP)通过较小规模的模型来确定样本的uncertainty进而决定复杂模型训练使用样本的策略。原文地址:Select Via Proxy: Efficient Data Selection Fo...
2018-10-08 13:18:16 1044
原创 【论文精读】Improving Simple Models with Confidence Profiles
Improving Simple Models with Confidence Profiles原文地址:Improving Simple Models with Confidence ProfilesAbstract用ProWeight方法进行模型迁移,使用linear probes通过flattened intermediate representations生成confidence s...
2018-10-03 14:54:01 485
原创 【论文精读】Learning Bounds for Importance Weighting
Learning Bounds for Importance Weighting原论文地址:Learning Bounds for Importance WeightingAbstract1 Introduction现实世界中机器学习训练数据和测试数据样本的分布会有偏差。一个常见的修正方法叫做importance weighting,它通过给不同的训练样本的带价值赋予权重来平衡这种偏差。一...
2018-10-02 15:01:29 1552 1
原创 【论文精读】Curriculum Learning
Curriculum Learning论文原文:Curriculum Learning课程学习(Curriculum Learning)由Montreal大学的Bengio教授团队在2009年的ICML会议上提出,主要思想是模仿人类学习的特点,由简单到困难来学习课程(在机器学习里就是容易学习的样本和不容易学习的样本),这样容易使模型找到更好的局部最优,同时加快训练的速度。Abstract人...
2018-10-01 08:56:57 30315 1
空空如也
TA创建的收藏夹 TA关注的收藏夹
TA关注的人