Developing a robust deep neural network (DNN) for a specific task is not only time-consuming but also requires lots of experienced human experts. In order to make deep neural networks easier to apply or even take the human experts out of the design of network architecture completely, a growing number of researches focus on robust automated machine learning (AutoML). In this paper, we investigated the robustness problem of AutoML systems based on contractive pseudoinverse learners. In our proposed method, deep neural networks were built with stacked contractive pseudoinverse learners (CPILer). Each CPILer has a Jacobian regularized reconstruction loss function and is trained with pseudoinverse learning algorithm. When sigmoid activation function is adopted in the hidden layer, the graph Laplace regularizer is derived from square Frobenius norm of the Jacobian matrix. This learning scheme not only speeds up the training process dramatically but also reduces the effort of hyperparameter tuning. In addition, the graph Laplace regularization can improve the robustness of the learning systems by reducing the sensibility to noise. An ensemble network architecture consisting of several sub-networks was designed to build the AutoML systems. The architecture hyperparameters of the system were determined in an automated way which could be considered as a data-driven way. The proposed method shown good performance in the experiments in terms of efficiency and accuracy, and outperformed the baseline methods on a series of benchmark data sets. The robustness improvement of our proposed method was also demonstrated in the experiments.

为特定任务开发强大的深度神经网络（DNN）不仅耗时，而且还需要大量经验丰富的人类专家。为了使深层神经网络更易于应用，甚至使人类专家完全脱离网络体系结构的设计，越来越多的研究集中在强大的自动化机器学习（AutoML）上。在本文中，我们研究了基于收缩伪逆学习器的AutoML系统的鲁棒性问题。在我们提出的方法中，使用堆叠的收缩伪逆学习器（CPILer）构建了深度神经网络。每个CPILer具有Jacobian正则化重构损失函数，并使用伪逆学习算法进行训练。当在隐藏层中采用S型激活函数时，图Laplace正则化器是从Jacobian矩阵的平方Frobenius范数得出的。这种学习方案不仅大大加快了训练过程，而且减少了超参数调整的工作量。另外，图拉普拉斯正则化图可以通过降低对噪声的敏感度来提高学习系统的鲁棒性。设计了一个由几个子网组成的集成网络体系结构，以构建AutoML系统。系统的架构超参数是通过自动方式确定的，可以将其视为数据驱动方式。提出的方法在效率和准确性方面在实验中显示出良好的性能，并且在一系列基准数据集上优于基线方法。