Measuring Catastrophic Forgetting in Neural Networks
Authors: Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, Christopher Kanan
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on real-world images and sounds show that the mechanism(s) that are critical for optimal performance vary based on the incremental training paradigm and type of data being used, but they all demonstrate that the catastrophic forgetting problem is not yet solved. |
| Researcher Affiliation | Academia | 1Rochester Institute of Technology 2Swarthmore College {rmk6217, mcm5756}@rit.edu , aabitin1@swarthmore.edu, {tlh6792, kanan}@rit.edu |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper mentions 'Supplemental materials are provided at the end of our ar Xiv submission: https://arxiv.org/abs/1708.02072' but does not explicitly state that source code for the methodology is provided within these materials or at a separate link. |
| Open Datasets | Yes | Table 1: Dataset Specifications (MNIST, CUB-200, Audio Set) ... CUB-200 Caltech-UCSD Birds-200 (CUB-200) is an image classification dataset containing 200 different bird species (Wah et al. 2011). ... Audio Set (Gemmeke et al. 2017) is a hierarchically organized audio classification dataset built from You Tube videos. |
| Dataset Splits | No | Table 1 provides 'Train Samples' and 'Test Samples' for the datasets, but no explicit counts or percentages for a validation set. It mentions 'validation data' in passing, but without details on its size or split. |
| Hardware Specification | Yes | Acknowledgements ... We thank NVIDIA for the generous donation of a Titan X GPU. |
| Software Dependencies | No | The paper mentions models like ResNet-50, but does not provide specific version numbers for software dependencies or libraries used for implementation. |
| Experiment Setup | Yes | The SOM-layer was fixed to 23 23 to have the same number of trainable parameters as the other models. |