基于因果干预与不变性的卷积预训练模型优化研究

基于卷积神经网络（CNN）的深度模型在图像识别与分类领域应用广泛，但在全局特征控制、概念层次特征不变性提取和变量之间的因果关系确定方面仍存在不足，使得深度模型缺乏灵活性、适应性及泛化性。基于因果干预和不变性，提出一种基于CNN深度模型的定向修剪和网络结构优化方法。通过对模型输入进行基于不变性的干预调制，根据生成的调制图片序列分析预训练网络卷积子结构的输出分布，筛选和定向修剪噪声敏感子结构。构建基于类间区分度的目标函数，借助经济学领域中的资本资产定价模型构建网络的层间连接，生成在单分类任务下能增大类间区分度的网络拓扑结构，逐层优化构建概念层次的稳定特征。在ImageNet-2012数据集上的实验结果表明，优化后的深度模型相比于ResNet50基线预训练模型的分类准确率约提升了5个百分点，并大幅降低了训练集规模。

Research on Optimization of Convolution Pretraining Model Based on Causal Intervention and Invariance

The deep learning model based on Convolutional Neural Network(CNN) has been widely used in image recognition and classification.However, the model still has some shortcomings in the overall grasp of global features, the effective extraction of feature invariance at concept level, and determining the clear causal relationship between variables.This makes deep models less flexible, adaptable and generalizable.Based on causal intervention and invariance, this study proposes a directional pruning and network structure optimization method based on a CNN deep model.The optimization method performs invariant-based intervention modulation on the model input, then analyzes the output distribution of the pretrained network convolutional substructure according to the generated modulation picture sequence, and filters and directionally prunes the noise-sensitive substructure.On this basis, the objective function based on inter-class differentiation and the inter-layer connection of the network are constructed with the help of the Capital Asset Pricing Model(CAPM) used in the field of economics.The network topology that can increase the inter-class differentiation under a single classification task is generated, and the stable characteristics of the concept level are optimized layer by layer.The experimental results on the ImageNet-2012 dataset show that the optimized deep model improves the classification accuracy of the ResNet50 baseline pre-training model by about 5 percentage points, and greatly reduces the size of the training set.