Development of method to analyze the high-dimensional and small-sample data based on machine learning and its application to cryptanalysis

已经知道，即使在标记数据的数量非常少的情况下，也可以通过使用未标记数据来提高监督学习的估计精度。这种类型的学习被称为半监督学习。最传统的半监督学习需要一些额外的假设来主导监督学习，即使我们有额外的信息。此外，对于模型选择，将常规标准（包括AIC或CV）应用于标记数据。但是，由于这些标准需要大量的标记数据，因此在此设置中无法很好地工作。我们的主要成果是解决了这些问题。

It is already known that the estimation accuracy of supervised learning can be improved by using the unlabeled data even when the number of labeled data is quite small. This type of learning is called semi-supervised learning. The most conventional semi-supervised learning requires some additional assumptions to dominate the supervised learning even though we have additional information. Further, as for model selection, the conventional criteria(including AIC or CV) are applied to the labeled data. However, because such criteria require a large number of labeled data, they do not work well in this setting. Our main result is that we solved these problems.