Search Method for a Singular Model Taking into AccountSingular Regions in a Search Space

针对多层感知器（MLP），我们提出了一种新的学习方法SSF，该方法很好地利用了奇异区域来稳定且成功地找到与隐藏单元数量相称的优秀解。SSF在我们使用人工和真实数据集的实验中表现良好。我们还提出了另一种利用特征向量下降的MLP学习方法，并证明了该方法在平坦奇异区域中移动以寻找最佳解。此外，我们得到了非常有希望的初步结果，即我们的SSF框架可以应用于复值MLP的学习，以寻找无界或周期解。

We proposed a new learning method called SSF for multilayer perceptron (MLP), which makes good use of singular regions to stably and successfully find excellent solutions commensurate with the number of hidden units. SSF worked well in our experiments using artificial and real data sets. We also proposed another learning method for MLP which utilizes eigen vector descent, and showed that it moved through flat singular regions to find excellent solutions. Moreover, we got very promising preliminary results that our SSF framework can be applied to learning of complex-valued MLP to find unbounded or periodic solutions.