Debiasing Pretrained Generative Models by Uniformly Sampling Semantic Attributes

Authors: Walter Gerych, Kevin Hickey, Luke Buquicchio, Kavin Chandrasekaran, Abdulaziz Alajaji, Elke A. Rundensteiner, Emmanuel Agu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on debiasing generators trained on popular real-world datasets show that our method outperforms state-of-the-art approaches.
Researcher Affiliation Collaboration Walter Gerych1 , Kevin Hickey1, Luke Buquicchio1, Kavin Chandrasekaran1, Abdulaziz Alajaji2, Elke Rundensteiner1, Emmanuel Agu1 1Worcester Polytechnic Institute, Worcester, MA 2King Saud University, Riyadh, Saudi Arabia
Pseudocode Yes Algorithm 1 Naively collect data from Q ... Algorithm 2 Collecting a dataset of samples from Qλ
Open Source Code Yes Code for our method is available in the supplemental material.
Open Datasets Yes Specifically, the generative model is a DCGAN [32] that we pretrain on a grayscale version of the UTKFace dataset [44]. ... We use the Race of the individual in each image as the semantic attribute, and use the MTCNN classifier from the Deep Face [36] package to classify race. ... Although our focus is primrily on models that map from low dimensional latent space to a higher dimensional feature space, we also evaluate our approach on a latent diffusion model 3 [35] that was trained on the Celeba-HQ dataset [19].
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits with percentages, sample counts, or references to predefined splits for reproduction.
Hardware Specification No The paper mentions 'a high-performance computing system acquired through NSF MRI grant DMS-1337943 to WPI' in the acknowledgments but does not provide specific hardware details such as GPU/CPU models or processor types.
Software Dependencies Yes The distribution mapper used default architecture of SDV s CTGAN 5 version 0.6.0, except for in the Progressive GAN experiment where embedding_dim =512, generator_dim =(512,512) were passed as arguments.
Experiment Setup Yes Details such as hyperparameter choice and architectures are available in the appendix. ... For the networks we trained, we utilized the Adam optimizer with learning rate between 0.002 and 0.0001.