Experimental Observations of the Topology of Convolutional Neural Network Activations
Authors: Emilie Purvine, Davis Brown, Brett Jefferson, Cliff Joslyn, Brenda Praggastis, Archit Rathore, Madelyn Shapiro, Bei Wang, Youjia Zhou
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide experimental results that show how a topological viewpoint of hidden-layer activations can summarize and compare the complex structures within them and how the conclusions align with our human understanding of the image classification task. We then show our experiments, which use two tools from TDA: persistent homology and mapper. |
| Researcher Affiliation | Collaboration | 1 Pacific Northwest National Laboratory 2 Scientific Computing and Imaging (SCI) Institute and School of Computing, University of Utah |
| Pseudocode | No | The paper describes the Persistent Homology and Mapper algorithms conceptually but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Code for the models and additional details regarding the dataset, as well as the parameters and computing infrastructure specific to each set of experiments, are provided in the ar Xiv technical appendix. The paper also mentions 'we plan to implement our methods into user-friendly tools,' indicating future release, not current. |
| Open Datasets | Yes | We use the standard benchmark dataset CIFAR-10 (Krizhevsky and Hinton 2009) on a Res Net-18 architecture (He et al. 2016). a subset of 10 classes from the Image Net dataset (Deng et al. 2009). |
| Dataset Splits | No | The paper mentions using CIFAR-10 training images for some experiments and test set images for others, but does not explicitly state training/validation/test splits, percentages, or a methodology for creating a validation set in the main text. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used (e.g., GPU models, CPU types, memory) in its main text. It mentions that 'parameters and computing infrastructure specific to each set of experiments, are provided in the ar Xiv technical appendix', but these details are not present in the provided text. |
| Software Dependencies | No | The paper mentions software like 'Open CV library' and 'Mapper Interactive' but does not specify their version numbers or other key software dependencies with specific versions. |
| Experiment Setup | Yes | We perform standard preprocessing to normalize the images by the mean and variance from the full training set. The mapper graph filter function is the l2-norm of each spatial activation. We trained 100 Res Net-18 models with different initialization seeds on CIFAR-10. To explore the stability of mapper graphs to noise in the input data, we injected pixel-wise Gaussian noise to all 50k images with different standard deviations (σ). |