Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities
Authors: Chengkun Sun, Jinqian Pan, Zhuoli Jin, Russell Stevens Terry, Jiang Bian, Jie Xu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on a variety of benchmarks, focusing on both 2D natural and 3D medical imaging applications, including tasks such as classification and segmentation. Our findings highlight Pool Skip s effectiveness in facilitating more robust CNN training and improving model performance. |
| Researcher Affiliation | Academia | 1Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL 32611, USA 2Department of Urology, University of Florida, Gainesville, FL 32611, USA 3Department of Statistics and Applied Probability, University of California Santa Barbara, CA 93106-3110, USA EMAIL, zhuoli EMAIL, EMAIL |
| Pseudocode | No | The paper includes diagrams like Figure 1 'Schematic representation of the computational process of Pool Skip' and mathematical equations, but no explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/sunck1/Pool Skip |
| Open Datasets | Yes | For the classification task, we utilized the CIFAR datasets (Krizhevsky, Hinton et al. 2009). For this task, we utilized Cityscapes (Cordts et al. 2016) and PASCAL Visual Object Classes (VOC) Challenge (Pascal VOC) (Everingham et al. 2010) datasets. We used abdominal multi-organ benchmarks for medical image segmentation, i.e., AMOS (Ji et al. 2022) and Multi-Atlas Labeling Beyond the Cranial Vault (BTCV) (BA et al. 2015) datasets. |
| Dataset Splits | Yes | For the BTCV dataset, 12 scans were assigned to the test set, and 18 to the training and validation set. From AMOS, 360 scans(containing CTs and MRIs) were divided into 240 for training and validation and 120 for testing, with a training-to-validation ratio of 4:1. We performed 5-fold cross-validation on all models, averaging their softmax outputs across folds to determine voxel probabilities. |
| Hardware Specification | Yes | Furthermore, all models were trained using a single NVIDIA A100 GPU with 80G of memory to maintain consistency in computational resources. |
| Software Dependencies | No | The paper mentions using common deep learning frameworks and models (e.g., CNNs, ResNet, VGG16, MobileNet, DeepLabv3+, nnUNet, V-Net, ViT, CCT, CVT) and components like ReLU and Batch Normalization, but it does not specify any software names with version numbers. |
| Experiment Setup | No | All models were equipped with BN and Re LU or Re LU variations following the convolutions, ensuring a standardized architecture for comparison. Furthermore, all models were trained using a single NVIDIA A100 GPU with 80G of memory to maintain consistency in computational resources. Each model was trained with 5 diverse seeds on the official training dataset, and the average and standard deviation of the Top-1 error from the final epoch were calculated on the official test dataset. All the training settings followed Devries and Taylor s work (De Vries and Taylor 2017). |