Convolutional neural networks as a typical deep neural network have shown superior performance in computer vision fields such as image recognition target detection and semantic segmentation However th
Convolutional neural networks (CNNs) have demonstrated exceptional performance in computer vision tasks such as image recognition, target detection, and semantic segmentation. However, the end-to-end learning approach of CNNs makes it challenging to interpret the logical relationships in their hidden layers and the decision-making process, which hinders their applications in industries like healthcare, autonomous driving, and criminal justice. Therefore, the development of interpretable CNNs is crucial to increase their application value.
Currently, the primary approach to interpreting CNNs is post-hoc explanation, which involves a statistical analysis of node activation in pre-trained models with additional information from the model. However, since the post-hoc explanation and the model prediction are independent processes, the results may not align with the model's original logic. Therefore, this thesis proposes an ad-hoc interpretable method for CNNs without additional auxiliary information. The thesis designs self-interpretable CNN models by creating interpretable components and regularization terms that implement the model's internal representation to solve the entanglement problem.
The thesis achieves the following:
-
Proposes a lightweight self-interpretable CNN based on interpretable components and the ad-hoc prototype sample interpretable method. The network uses interpretable components to achieve transparency in the model's inference process and designs a lightweight architecture with performance guarantees by optimizing the convolutional structure to achieve self-interpretation of the network with low deployment cost. The Concept Prototype Interpretation (CPI) component based on high-level features is designed to obtain high-confidence image regions as concept prototype samples automatically, and then use these prototypes to interpret test samples for case inference. The Multi-scale Feature Augmentation Convolution (MFA) module is designed to reduce loss of model classification performance due to increased model interpretability. The Lightweight Concept Prototype Interpretation Network (LCPI-Net) based on an interpretable component is proposed by using MobileNetV3 as the baseline network, designing a lightweight convolutional structure with MFA, and embedding the CPI component. The results of classification and visualization experiments on the CIFAR-10 and ILSVRC2012 datasets show that LCPI-Net has better performance on image classification tasks while maintaining model interpretability and lightweightness.
-
Proposes an interpretable CNN with an adaptive disentangled latent space for the problem that post-hoc concept vector methods rely on an additional classifier independent of the original model. The network uses a graph convolution module to design an adaptive disentangled interpretable classifier and constructs a transparent latent classification space by learning category representative basis concepts and inter-concept dependencies and employing different orthogonal losses to separate the entanglements between basis concepts. The within-class concepts graphs encoder in latent space is designed to encode the within-class basis concepts in the form of graph structure, and then the graph convolution module is used to learn vertex features and directed edge features. The Adaptive Disentangled Interpretable Classifier (ADIC) is proposed to realize the self-separation of the category basis concepts. The results of classification and interpretable image recognition experiments on the Mini-ImageNet and Places365 datasets with ResNet-18 and ResNet-50 as the baseline models and the addition of ADIC show that ADIC can improve the accuracy of the baseline models while ensuring their self-interpretability.
This thesis includes 39 figures, 10 tables, and 107 references. Keywords: convolutional neural networks, interpretability, model renovation, concept prototype sample, category basis concept.
原文地址: https://www.cveoy.top/t/topic/bMBe 著作权归作者所有。请勿转载和采集!