[Paper Review] DASH: Visual Analytics for Debiasing Image Classification via User-Driven Synthetic Data Augmentation

Bum Chul Kwon, Jungsoo Lee+4 more|arXiv (Cornell University)|Jan 1, 2022

Cell Image Analysis Techniques8 citations

TL;DR

DASH is a visual analytics system that enables human-in-the-loop debiasing of image classification models by identifying bias factors through interactive visualization, generating synthetic training data via XploreGAN-based image-to-image translation, and iteratively retraining models. It significantly improves model accuracy on biased datasets, with users achieving 90% test accuracy on cartoons and 65% on fruits after iterative refinement.

ABSTRACT

Image classification models often learn to predict a class based on irrelevant co-occurrences between input features and an output class in training data. We call the unwanted correlations ''data biases,'' and the visual features causing data biases ''bias factors.'' It is challenging to identify and mitigate biases automatically without human intervention. Therefore, we conducted a design study to find a human-in-the-loop solution. First, we identified user tasks that capture the bias mitigation process for image classification models with three experts. Then, to support the tasks, we developed a visual analytics system called DASH that allows users to visually identify bias factors, to iteratively generate synthetic images using a state-of-the-art image-toimage translation model, and to supervise the model training process for improving the classification accuracy. Our quantitative evaluation and qualitative study with ten participants demonstrate the usefulness of DASH and provide lessons for future work.

Motivation & Objective

• To address the challenge of data biases in image classification models, where models learn spurious correlations between irrelevant visual features (e.g., swamp background) and class labels (e.g., frogs).
• To support data scientists in identifying bias factors through visual exploration of model predictions and feature importance.
• To enable iterative, user-guided data augmentation by generating synthetic images that remove bias factors using state-of-the-art image translation (XploreGAN).
• To provide a unified visual interface for monitoring model performance across retraining iterations and guiding model refinement.

Proposed method

• DASH integrates multiple coordinated visualization views: Projection View for latent space clustering, Mosaic View for comparing model performance across iterations, and Trace View for tracking prediction changes.
• The Grad-CAM View visualizes attention maps to identify which image regions influence classification decisions, aiding in bias factor detection.
• The Cluster GAN View uses XploreGAN to transfer visual attributes (e.g., background) from source images to target images while preserving the main object, enabling synthetic data generation.
• The Augmented Image View displays newly generated images with labels for retraining, and the Classifiers Board tracks performance metrics across model versions.
• Users iteratively identify bias sources, generate synthetic data, retrain models, and evaluate performance using the system’s visual feedback loop.
• The system supports user-driven model retraining by allowing adjustment of hyperparameters and comparison of model versions through visual and quantitative feedback.

Experimental results

Research questions

• RQ1 How can visual analytics support data scientists in identifying bias factors in image classification models?
• RQ2 What visual interaction patterns enable effective user-driven synthetic data generation for debiasing?
• RQ3 How does user involvement in data augmentation impact model performance and bias mitigation?
• RQ4 What are the key challenges and usability barriers in using such a system for model debiasing?

Key findings

• Ten machine learning experts successfully mitigated data biases in two real-world datasets using DASH, achieving 90% test accuracy on the cartoon dataset.
• Seven out of ten participants reached 65% test accuracy on the fruit dataset, demonstrating the system’s effectiveness on more complex, real-world data.
• Participants required an average of 2.1 retraining iterations on the cartoon dataset and 3.3 on the fruit dataset, indicating iterative refinement is necessary and supported by DASH.
• The Wilcoxon Signed Rank Test confirmed a statistically significant difference in retraining effort between the two datasets (p=0.048), reflecting perceived difficulty differences.
• Users reported that prior knowledge and domain expertise were critical, with some avoiding certain views (e.g., Grad-CAM) due to distrust in the technique.
• Participants requested features like augmentation history tracking and dedicated analysis for frequently misclassified images, highlighting opportunities for future system enhancements.