Reproduce “Debiasing Classifiers by Amplifying Bias with Latent Diffusion and Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Donggeun Ko, Dongjun Lee, Namjun Park, Wonkyeong Shim, Jaekwang Kim
DiffuBias is a framework that mitigates dataset bias in image classifiers by generating bias-conflict samples using pre-trained latent diffusion models and large language models. The method effectively identifies and amplifies biases to create diverse, challenging samples, leading to improved classifier accuracy on benchmark datasets while significantly reducing computational costs and carbon emissions compared to prior methods.
Key insights
- Leveraging pre-trained large language models (LLMs) for image captioning and latent diffusion models (LDMs) for image generation allows for efficient creation of bias-conflict samples without training generative models from scratch.
- A biased classifier trained with Generalized Cross Entropy (GCE) loss can effectively identify samples that challenge existing dataset biases, serving as a basis for targeted data augmentation.
- The approach significantly reduces computational cost and carbon emissions compared to traditional GAN-based debiasing methods by avoiding the need for extensive generative model training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.