Skip to content

Latest commit

 

History

History
20 lines (15 loc) · 1.65 KB

study_description.md

File metadata and controls

20 lines (15 loc) · 1.65 KB

Interactively Correcting AI Model Errors

Aim of the Study

The primary goal of this study is to understand the limitations and inaccuracies of a specific AI model in the domain of image classification. Your insights will help us identify common failure modes, thereby aiding in model improvement.

Participant Tasks

  • Review Images: You will be shown a series of images that our AI model has classified. Some classifications are correct, while others are not. Your task is to identify what the incorrect examples have in common.
  • Propose Descriptions: Once you have an understanding of the model's inaccuracies, you will suggest short, descriptive phrases (1-3 words) that encapsulate these failure modes. For instance, if the model often misclassifies images containing water, you might write "pond" or "water."
  • Inspect Results: The study interface will then display new images based on your proposed attributes. You'll need to verify if these new images align with your initial observations. You will repeat steps 2 and 3 at least 3 times for each task.

Sensitive Information

No sensitive personal information is required for participation.

Reward and Bonus Payments

Participants will be rewarded promptly upon successful submission. Additionally, a bonus of 50 cents will be given for accurately identifying true failure modes, up to a total of $1.

Tips for Success

  • Keep your descriptive phrases short; usually 1-2 words will suffice.
  • Take your time to verify if the model's filtered images align with your initial observations to ensure you qualify for the bonus. A score above our suggested goal (which varies by task) will qualify for the bonus.