Back-Modality: Leveraging Modal Transformation for Data Augmentation
Authors: Zhi Li, Yifan Liu, Yin Zhang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive evaluations across tasks such as image classification, sentiment classification, and textual entailment demonstrate that our methods consistently enhance performance under data-scarce circumstances. |
| Researcher Affiliation | Academia | Zhi Li Zhejiang University, China zhili@zju.edu.cn Yifan Liu Zhejiang University, China yifan.liu@zju.edu.cn Yin Zhang Zhejiang University, China zhangyin98@zju.edu.cn |
| Pseudocode | No | The paper describes its methods conceptually and with equations but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code used to generate augmented samples and the TNCC dataset are both accessible3. [Footnote 3]: https://github.com/zhilizju/Back-Modality |
| Open Datasets | Yes | Tiny Image Net [Le and Yang, 2015], SST-2 [Socher et al., 2013], and TNCC, based on Crisscrossed Captions [Parekh et al., 2020]. |
| Dataset Splits | Yes | The Stanford Sentiment Treebank-2 (SST-2) [Socher et al., 2013] is a sentiment classification dataset populated with movie reviews gathered from Rotten Tomatoes, paired with their corresponding binary labels. The dataset is partitioned into training, validation, and testing sets, comprising 67,349, 872, and 1,821 instances, respectively. Tiny Image Net [Le and Yang, 2015] Each class is furnished with 500 training images, 50 validation images, and 50 test images. The TNCC dataset is partitioned into training, validation, and testing sets, containing 3,600, 1,200, and 1,560 instances, respectively. |
| Hardware Specification | Yes | All experiments were conducted using Py Torch and executed on RTX 6000 GPUs and the Atlas computing cluster. ... The additional computational overhead of various augmentation methods compared to the base model on RTX A6000. |
| Software Dependencies | No | The paper mentions using PyTorch and models from the Huggingface Transformers library, as well as gpt-3.5-turbo, but does not specify their version numbers. |
| Experiment Setup | Yes | During the training phase, we leverage the Adam optimization algorithm with a learning rate set at 5e 5, the first and second momentum terms, β1 and β2, are respectively set to 0.9 and 0.999. Additionally, we introduce an L2 weight decay of 0.01 to the model. We select a batch size of 2 for all trials. For the image classification task... we employ the SGD optimizer with a learning rate of 0.1, momentum of 0.9, and a weight decay of 0.0001. |