This work propose DivideMix, a novel framework for learning with noisy labels by leveraging semi-supervised learning techniques. In particular, DivideMix models the per-sample loss distribution with a mixture model to dynamically divide the training data into a labeled set with clean samples and an ...
Learning with noisy labels (LNL) aims to traina high-performing model using a noisy dataset. We observe thatnoise for a given class often comes from a limited set of categories,yet many LNL methods overlook this. For example, an image mislabeledas a cheetah is more likely a leopard than ...
尽管LNL和SSL各自取得了进步,但它们之间的联系尚未得到充分的探索。在这项工作中,我们提出了DivideMix,它以半监督的方式处理标签噪声的学习。与大多数现有的LNL方法不同,DivideMix丢弃了极有可能是噪声的样本标签,利用噪声样本作为未标记数据,使模型从过拟合中得到正则化,提高泛化性能。这项工作的主要贡献是: 我们提出c...
Learning with Noisy Labels (LNL) The main running file ismain.pywith arguments as follows: noise_type: symmetric | asymmetric noise_rate: noise rate loss: AGCE | AUL | AEL | CE (Cross Entropy) | FL (Focal Loss) | MAE | GCE | SCE | NFL | NCE | ... ...
第二种 Real-World Noisy Datasets 是需要特定的数据集的，它对于数据集制作者来说可能是成本不高的，...
Model predictions for noisy labels 许多方法在训练过程中使用模型本身的预测来生成伪标签或识别不正确的示例(或两者都有)。对于深度学习,这通常利用了正确的标签往往比不正确的标签更早逼近的现象【26】. Reed et al[33]首先提出了bootstrapping loss,每个例子的回归目标都采用带注释的标签和当前预测的固定线性组合...
相比于无监督学习,learning with noisy label 更贴近深度学习在工业界的落地。典型的状态如下: 初始阶段有一定量的标注质量未知的数据。 一般会有持续的人工投入,不断提升标注质量。人工投入的形式,可能是付费众包,可能是借助用户反馈。 对...
几篇论文实现代码:《Learning with Noisy Labels via Sparse Regularization》(ICCV 2021) GitHub:https:// github.com/hitcszx/lnl_sr 《De-rendering Stylized Texts》(ICCV 2021) GitHub:https:// github.co...
2.1 LEARNING WITH NOISY LABELS 现有的训练带噪声标签的dnn的方法大都是为了修正loss函数。修正方法可以分为两类。第一种方法对所有样本一视同仁,通过重新标记噪声样本来显式或隐式地纠正损失。对于重标记方法,对噪声样本的建模采用有向图模型(Xiao et al., 2015)、条件随机场(Vahdat, 2017)、知识图(Li et al...
First, we provide a simple unbiased estimator of any loss, and obtain performance bounds for empirical risk minimization in the presence of iid data with noisy labels. If the loss function satisfies a simple symmetry condition, we show that the method leads to an efficient algorithm for ...