Explainable AI Methods on Vertex AI
How to use Vertex Explainable AI to understand your model's predictions. A guide to the different feature attribution methods available on Vertex AI.
Opening the Black Box
Deep learning models are often referred to as "black boxes" because it can be difficult to understand how they make their predictions. Vertex Explainable AI is a tool that helps you open up the black box and understand what's going on inside your model.
1. Feature Attributions
Explainable AI provides feature attributions, which tell you how much each feature in your model contributed to a particular prediction. This can be useful for:
- Debugging your model: If your model is making incorrect predictions, you can use feature attributions to see which features are causing the problem.
- Understanding your model's behavior: You can use feature attributions to get a better understanding of how your model works and what it's learning.
- Ensuring fairness: You can use feature attributions to make sure that your model is not making predictions based on sensitive features, such as race or gender.
2. Feature Attribution Methods
Vertex Explainable AI supports three different feature attribution methods:
- Sampled Shapley: This method is based on a game theory concept called the Shapley value. It's very accurate, but it can be slow to compute. It's best for tabular data.
- Integrated Gradients: This method is based on the gradients of your model. It's faster than Sampled Shapley, but it's only compatible with differentiable models (i.e., neural networks).
- XRAI: This method is a variation of Integrated Gradients that is specifically designed for image data. It produces a heatmap that shows which regions of the image were most important for a particular prediction.
3. The What-If Tool
The What-If Tool is an interactive visualization tool that allows you to explore the behavior of your model. You can use it to:
- Visualize your data: See the distribution of your data and identify any outliers.
- Analyze your model's performance: See how your model's performance varies across different slices of your data.
- Test counterfactuals: See how your model's prediction would change if you were to change the value of a particular feature.
Knowledge Check
?Knowledge Check
You are training a deep learning model to classify images. You want to understand which parts of the image are most important for a particular prediction. Which feature attribution method should you use?