【AAAI 2021】Few

AAAI 2021】Few"/>

【AAAI 2021】Few

背景知识

One-class classification

learning a binary classifier with data from only one class //从一个类中学习一个二分类器。 【想法】可以用来做异常检测。

The anomaly detection (AD) task

(Chandola, Banerjee, and Kumar 2009; Aggarwal 2015) consists in differentiating between normal and abnormal data samples. AD
AD problems are usually formulated as one-class classification (OCC) problems (Moya, Koch, and Hostetler 1993), where either only a few or no anomalous data samples are available for training the model (Khan and Madden 2014).

内容概要

Our work addresses the few-shot OCC problem and presents a method to modify the episodic data sampling strategy of the model-agnostic meta-learning (MAML) algorithm to learn a model initialization particularly suited for learning few-shot OCC tasks. //我们的工作解决了few-shot OCC问题，并提出了一种方法来修改模型无关元学习(MAML)算法的情景数据采样策略，学习一个特别适合于学习few-shot OCC任务的模型初始化。

We provide a theoretical analysis that explains why our approach works in the few-shot OCC scenario, while other meta-learning al- gorithms fail, including the unmodified MAML. //我们提供了一个理论分析，解释了为什么我们的方法在few-shot OCC场景中有效，而其他元学习算法失败，包括未修改的MAML。

Our exper- iments on eight datasets from the image and time-series do- mains show that our method leads to better results than classi- cal OCC and few-shot classification approaches, and demon- strate the ability to learn unseen tasks from only few nor- mal class samples. //我们在8个图像和时间序列数据集上的实验表明，我们的方法比经典的OCC和few-shot分类方法取得了更好的结果，并证明了仅从少数普通样本中学习不可见任务的能力

we successfully train anomaly detectors for a real-world application on sensor readings recorded during industrial manufacturing of workpieces with a CNC milling machine, by using few normal examples. //通过使用几个普通的例子，我们成功地训练了异常检测器，以便将其应用于CNC铣床工件工业制造过程中记录的传感器读数。

contributions

1. Firstly, we show that classical OCC approaches fail in the few-shot data regime. 我们证明了经典的OCC方法在few-shot数据中失败了。
2. Secondly, we provide a theoretical analysis showing that classical gradient-based meta-learning algorithms do not yield parameter initializations suitable for OCC and that second- order derivatives are needed to optimize for such initializations. 我们提供了一个理论分析，表明经典的基于梯度的元学习算法不产生适合OCC的参数初始化，并且二阶导数需要为这种初始化进行优化。
3. Thirdly, we propose a simple episode generation strat- egy to adapt any meta-learning algorithm that uses a bi-level optimization
   scheme to FS-OCC. Hereby, we first focus on modifying the model-agnostic meta-learning (MAML) al- gorithm (Finn, Abbeel, and Levine 2017) to learn initializa- tions useful for the FS-OCC scenario. The resulting One- Class MAML (OC-MAML) maximizes the inner product of loss gradients computed on one-class and class-balanced minibatches, hence maximizing the cosine similarity be- tween these gradients.
4. Finally, we demonstrate that the pro- posed data sampling technique generalizes beyond MAML to other metalearning algorithms, e.g., MetaOptNet (Lee et al. 2019) and Meta-SGD (Li et al. 2017), by successfully adapting them to the understudied FS-OCC.

主要方法

算法描述