人工智能快速发展的今天，模式识别和图像分类是其发展最快的领域。深度学习技 术在各项图像分类大赛中力压其他机器学习算法，并且超过了人类的识别精度，被广泛 应用于各个场景。传统的分类方法大多是通过人工设定的特征提取方式并结合分类器实 现的。与之不同，深度学习不需要人工设计而是自动地提取数据中的抽象特征，较人工 特征提取方法，深度学习提取特征训练的分类器表现出更优的性能。 然而，深度模型的结构复杂，设计和训练网络均需要研究人员长时间的经验积累。 目前训练深度网络大多使用基于梯度的算法，在优化网络过程中会遇到例如梯度消失、 梯度爆炸或是陷入局部最小值等问题，同时需要研究人员大量的经验调节适合该网络的 超参数。因此，训练一个可用的深度模型需要大量的时间和人力。 高效快速的优化方法一直是深度学习研究热点。伪逆学习是一种优化多层前馈神经 网络的算法，它不需要迭代更新，不依赖梯度下降，只需要简单线性代数运算即可求得 网络权重。伪逆学习算法相比其他的优化算法训练速度快，同时避免了因梯度出现的一 些问题。 在此背景下，以伪逆学习算法为基础，本文在其工作上进行了拓展，完成了如下工作: 对于图像分类任务，伪逆学习自编码器是将图像拉伸为一维向量输入到网络中，导 致二维空间上的信息丢失。为了解决这个问题，本文提出了联合方向梯度直方图的 伪逆学习自编码器模型，实验结果显示该方法提升了模型的分类性能。 ? 对于训练数据样本数少、样本不均衡而导致训练的网络模型鲁棒性差、出现过拟合 现象的问题，本文提出了伪逆学习层次化混合专家模型的集成方法，并且将该方法 应用到脉冲星候选体筛选中，解决了脉冲星数据样本少，正负样本不均衡的问题， 该集成模型较其他模型相比鲁棒性强，并且一定程度上解决了过拟合问题。 ? 混合误差建模可以拟合数据中复杂的噪声，根据贝叶斯理论，在模型的损失函数中 L1 正则对应拉普拉斯先验，L2 正则对应高斯先验。对于带有 L1 和 L2 的损失函数，由于 L1 正则项在原点处不可导，往往很难求得整个损失函数的解。本文采用了“分 而治之”的策略，提出了混合误差下的伪逆学习集成模型，即使用若干带有不同正 则化项的多层网络作为基学习器，集成基学习器的结果作为新的数据集训练元学习 器，元学习器的输出即为最后结果。结果显示，提出的模型相比基学习器表现出更 好的效果，提升了分类性能。

With the development of artificial intelligence, pattern recognition and image classification are the fastest growing direction. Deep learning techniques, which are widely used in various scenes, force other machine learning algorithms in various image classification contests and ex- ceed human recognition accuracy. Traditional image classification is mostly achieved by manual extraction of features based on feature combination classifier. By contrast, deep learning is a feature learning method with supervised learning, which shows better performance than other ma- chine learning algorithms. However, the structure of the deep model is complex, and both the design and training of the network model need a long-term experience. At present, most deep model training relies on gradient-based algorithm, and the optimization model encounters problems such as gradient disappearance, gradient explosion, or local minimum. At the same time, in the process of training the network, researchers need plenty of experience to set the hyperparameters for the network. Therefore, training a usable deep model requires a lot of time and labor. Efficient and fast optimization methods have always been hot topics of deep learning. The pseudo-inverse learning algorithm (PIL) is an algorithm for optimizing multi-layer feed-forward neural networks, which neither need iterative update and nor rely on gradient descent, and only need linear algebraic operations to obtain network weights. PIL algorithm is faster than other optimization algorithms and avoids most problems caused by the gradient. Based on the PIL algorithm, this paper extends its work and completes the following works: ? For the image classification, the pseudo-inverse learning training auto-encoder accepts a one-dimensional vector, which stretches the image into a one-dimensional vector and send it into the network, to result in the loss of information in the two-dimensional space. In order to solve this problem, this paper proposed the HOG-PILAE model. Experimental results show that this method improves the performance of image classification. Since the small and unbalanced training data samples will cause poor robustness and over- fitting of the trained network model, this paper proposes an ensemble method based on pseudo-inverse learning hierarchical expert model, and applies the method in the pulsar candidate selection, to solve the problem that the pulsar data sample is small and the posi- tive and negative samples are unbalanced. It achieves a stronger robustification than other models and solves the over-fitting problem to some extent. ? Mixed error modeling can fit complex noises in the data. According to Bayesian theory, in the loss function of the model, L1 term corresponds to Laplace prior, and L2 term corre- sponds to gaussian prior. For the loss function with L1 and L2, since the regular term of L1 is not differentiable at the origin, it is often difficult to find the solution of the whole loss function. In order to solve this problem, divided and conquer strategy is adopted, several multi-layer networks with different regularization terms are used as the base learner, and then the results of the base learner are integrated to train a meta-learner as the final result. The results show that the proposed model better performs than that of base learner.