Interpretable Compositional Convolutional Neural Networks
Authors: Wen Shen, Zhihua Wei, Shikun Huang, Binbin Zhang, Jiaqi Fan, Ping Zhao, Quanshi Zhang
IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments have demonstrated the effectiveness of our method. and We applied our method to CNNs with six types of architectures to demonstrate the broad applicability of our method. We used object images in four different benchmark datasets to learn compositional CNNs for both the binary classification of a single category and the multi-category classification. We designed two metrics to measure the inconsistency of a filter s visual patterns and the diversity of visual patterns. We also visualized feature maps of a filter to qualitatively show the consistency of a filter s visual patterns. We compared the performance of learning interpretable filters in different convolutional layers of a compositional CNN. We also discussed the effects of the group number on the performance of learning interpretable filters. |
| Researcher Affiliation | Academia | 1Tongji University, Shanghai, China 2Shanghai Jiao Tong University, Shanghai, China {wen shen,zhihua wei,hsk,0206zbb,1930795,zhaoping}@tongji.edu.cn,zqs1022@sjtu.edu.cn |
| Pseudocode | No | The paper describes the algorithm and optimization process in text and mathematical equations, but does not include a formally structured pseudocode block or algorithm steps. |
| Open Source Code | No | The code will be released when the paper is accepted. |
| Open Datasets | Yes | All these compositional CNNs were learned based on the CUB200-2011 dataset [Wah et al., 2011], the Large-scale Celeb Faces Attributes (Celeb A) dataset [Liu et al., 2015], the Helen Facial Feature dataset [Smith et al., 2013], and animal categories in the PASCAL-Part dataset [Chen et al., 2014]. |
| Dataset Splits | No | The paper mentions training and testing on datasets and states 'We randomly selected the same number of samples from the PASCAL-Part dataset as negative samples for training and testing. We followed experimental settings in [Zhang et al., 2018] to learn compositional CNNs for binary classification of a single category on the CUB200-2011 dataset and the PASCAL-Part dataset.' However, it does not explicitly provide specific percentages or counts for training/validation/test splits within this paper. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments. |
| Software Dependencies | No | The paper mentions using different CNN architectures and the batch-normalization operation but does not provide specific version numbers for software dependencies like programming languages, libraries, or frameworks. |
| Experiment Setup | Yes | For binary classification of a single category, we set λ = 1.0 for most DNNs except for VGG-16 with λ = 0.1. For multi-category classification, we set λ = 0.1 and β = 0.1... For compositional CNNs learned from the CUB200-2011 dataset, the PASCAL-Part dataset, and the Helen Facial Feature dataset, we set K = 5. For the Celeb A dataset, we set K = 16... We replaced the zero padding with the replication padding for all compositional CNNs. For traditional CNNs based on the Dense Net architectures, we initialized parameters of the fully-connected layers, and loaded parameters of other layers from the same architectures that were pre-trained using the Image Net dataset. |