Men Also Do Laundry: Multi-Attribute Bias Amplification
Authors: Dora Zhao, Jerone Andrews, Alice Xiang
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using our proposed metric, we compare the performance of multi-label classifiers trained on COCO (Lin et al., 2014), im Situ, and Celeb A (Liu et al., 2015), standard benchmarks for bias amplification metrics (Zhao et al., 2017; Wang et al., 2019; Wang & Russakovsky, 2021; Hirota et al., 2022; Ramaswamy et al., 2021). |
| Researcher Affiliation | Industry | 1Sony AI, New York 2Sony AI, Tokyo. |
| Pseudocode | No | The paper provides mathematical equations for its metrics but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/Sony Research/multi_bias_amp. |
| Open Datasets | Yes | Using our proposed metric, we compare the performance of multi-label classifiers trained on COCO (Lin et al., 2014), im Situ, and Celeb A (Liu et al., 2015), standard benchmarks for bias amplification metrics (Zhao et al., 2017; Wang et al., 2019; Wang & Russakovsky, 2021; Hirota et al., 2022; Ramaswamy et al., 2021). |
| Dataset Splits | Yes | The greedy oversampling procedure results in 45,657, 12,351, and 27,499 images respectively in the training, validation, and test sets for COCO. The splits for im Situ are 40,470, 10,668, and 17,036 images. Finally, the splits for Celeb A are 379,661, 65,895, and 52,397 images. |
| Hardware Specification | Yes | All models in this work were developed using Py Torch. The models are trained and evaluated on 1 NVIDIA T4 Tensor Core GPU with 64 GB of GPU memory and 2.5 GHz Cascade Lake 24C processors. The operating system is Linux 64-bit Ubuntu 18.04. |
| Software Dependencies | No | The paper mentions that models were developed using PyTorch and the operating system is Linux 64-bit Ubuntu 18.04, but it does not provide specific version numbers for PyTorch or other key software libraries. |
| Experiment Setup | Yes | The model is trained for 50 epochs using an Adam optimizer (Kingma & Ba, 2015) with L2 weight decay of 10 6, batch size of 32, and a learning rate of 10 5. |