On the Effects of Artificial Data Modification
Authors: Antonia Marcu, Adam Prugel-Bennett
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Subsequently, through a series of experiments we seek to correct and strengthen the community s perception of how augmenting affects learning of vision models. Based on our empirical results we argue that the impact of the artefacts must be understood and exploited rather than eliminated. |
| Researcher Affiliation | Academia | 1Vision, Learning and Control research group, University of Southampton. Correspondence to: Antonia Marcu <am1g15@soton.ac.uk>. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide the code at https://github.com/Antonia Marcu/Data Modification. |
| Open Datasets | Yes | The main data sets we report results on are CIFAR-10/100 (Krizhevsky, 2009), Tiny Image Net (Stanford, 2015), Fashion MNIST (Xiao et al., 2017), and Image Net (Russakovsky et al., 2015). |
| Dataset Splits | No | The paper mentions using standard datasets like CIFAR-10/100, Tiny Image Net, Fashion MNIST, and Image Net, which often have predefined splits. However, it does not explicitly state the specific training, validation, and test split percentages or sample counts used for its experiments, nor does it cite the exact split methodology. |
| Hardware Specification | Yes | The models were trained on either one of the following: Titan X Pascal, Ge Force GTX 1080 Ti or Tesla V100. For the analyses, a Ge Force GTX 1050 was also used. |
| Software Dependencies | No | This is due to an incompatibility with newer versions of the Py Torch library of the official implementation of Harris et al. (2020), which we use as a starting point for model training. |
| Experiment Setup | Yes | Throughout the paper, we use Pre Act-Res Net18 (He et al., 2016) models, trained for 200 epochs with a batch size of 128. For the MSDA parameters we use the same values as Harris et al. (2020). All models are augmented with random crop and horizontal flip and are averaged across 5 runs. We optimise using SGD with 0.9 momentum, learning rate of 0.1 up until epoch 100 and 0.001 for the rest of the training. |