How to make CV models more robust?
Our capstone project is motivated by the problem of model robustness. How do we train, test, and deploy computer vision models with confidence in their reliability and fairness?
The current methods are relatively inefficient, expensive, and time consuming. Most people rely on their test set to diagnose model failures and then subsequently augment their training set to hopefully fix these mistakes. Collecting test and training datasets such that they are perfectly balanced and free of bias is impractical and oftentimes impossible. How can we achieve our goals with less manual effort and make our models more robust in a more targeted manner?
With this high-level motivation in mind, our project provides a pipeline to:
- Model diagnosis: tell you why a model is failing and what kind of data to supplement
- Data augmentation: with fine-grained control, generate targeted training data that improve the model precisely