标签
ProtoX-AD is a prototype-based self-explainable framework for self-supervised time series anomaly detection that provides interpretable explanations for detected anomalies by learning transformation-aware prototypes, achieving performance comparable to black-box methods while offering semantic anomaly characterization.
本文提出了一种基于Transformer的架构,结合原型学习,仅利用行级转录即可从历史文档中进行可扩展的古文字测量,并在仅有少量训练数据的160页手抄本上证明了其有效性。
本文表明,交叉熵和监督对比学习都是超球面上的原型学习形式,并提出了归一化损失函数(NTCE和NONL),这些损失函数通过设计实现Neural Collapse,性能优于标准方法。