FairBatch: Batch Selection for Model Fairness

Authors: Yuji Roh, Kangwook Lee, Steven Euijong Whang, Changho Suh

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments conducted both on synthetic and benchmark real data demonstrate that Fair Batch can provide such functionalities while achieving comparable (or even greater) performances against the state of the arts.
Researcher Affiliation Academia Yuji Roh1, Kangwook Lee2, Steven Euijong Whang 1, Changho Suh1 1KAIST, {yuji.roh,swhang,chsuh}@kaist.ac.kr 2University of Wisconsin-Madison, kangwook.lee@wisc.edu
Pseudocode Yes Algorithm 1: Bilevel optimization with Minibatch SGD
Open Source Code No The paper demonstrates the ease of implementing Fair Batch with a single-line change in PyTorch code (Fig. 1b) but does not explicitly state that its own source code is publicly available or provide a link to a repository.
Open Datasets Yes We use the real benchmark datasets: Pro Publica COMPAS (Angwin et al., 2016) and Adult Census (Kohavi, 1996) datasets with 5,278 and 43,131 examples, respectively. We also employ the UTKFace dataset (Zhang et al., 2017) with 23,708 images...
Dataset Splits Yes We perform cross-validation on the training sets to find the best hyperparameters for each algorithm. We evaluate models on separate test sets, and the ratios of the train versus test data for the synthetic and real datasets are 2:1 and 4:1, respectively.
Hardware Specification Yes We use Py Torch, and our experiments are performed on a server with Intel i7-6850 CPUs and NVIDIA TITAN Xp GPUs.
Software Dependencies No The paper mentions using 'Py Torch' but does not specify its version or any other software dependencies with their respective version numbers.
Experiment Setup Yes We use logistic regression in all experiments except for Sec. 4.2 where we fine-tune Res Net18 (He et al., 2016) and Goog Le Net (Szegedy et al., 2015) in order to demonstrate Fair Batch s ability to improve fairness of pre-trained models. We use the Adam optimizer for all trainings. The default batch sizes are: 100 (synthetic); 200 (COMPAS), 1,000 (Adult Census); and 32 (UTKFace).