Curriculum Learning is a powerful training method that allows for faster and better training in some settings. This method, however, requires having a notion of which examples are difficult and which are easy, which is not always trivial to provide. A recent metric called C-Score acts as a proxy for example difficulty by relating it to learning consistency. Unfortunately, this method is quite compute intensive which limits its applicability for alternative datasets. In this work, we train models through different methods to predict C-Score for CIFAR-100 and CIFAR-10. We find, however, that these models generalize poorly both within the same distribution as well as out of distribution. This suggests that C-Score is not defined by the individual characteristics of each sample but rather by other factors. We hypothesize that a sample's relation to its neighbours, in particular, how many of them share the same labels, can help in explaining C-Scores. We plan to explore this in future work.

翻译：课程学习是一种强大的培训方法,可以在某些环境里更快地进行更好的培训。但是,这种方法需要有一个概念,说明哪些例子困难,哪些容易,但并非总能提供。最近的一个称为C-STR的衡量标准作为替代标准,例如难以将其与学习的一致性联系起来。不幸的是,这种方法计算得相当密集,限制了其对替代数据集的适用性。在这项工作中,我们通过不同的方法为C-STRA-100和CIFAR-10培训模型。然而,我们发现,这些模型在分布和分布上都不太普遍。这表明C-STRA并不是由每个样本的个别特征来界定,而是由其他因素来界定。我们假设,抽样与其邻居的关系,特别是其中有多少人拥有相同的标签,可以帮助解释C-STRA。我们计划在未来的工作中探索这一点。