9. Model Explainability on Vertex AI

Question 1

What is explainability?

Answer 1

Explainability is the extent you can explain the internal mechanics of an ML system in human terms.

Question 2

What are the two types of explainability?

Answer 2

Global: Make the overall ML model transparent and comprehensive
Local: Explain the model’s individual predictions

Question 2

Why is explainability important?

Answer 2

It makes customers comfortable with model predictions. It can also help debugging and improvement.

Question 2

What are interpretability and explainability?

Answer 2

Interpretability: how accurately a machine learning model can associate a cause to an effect.
Explainability: Explain the ability of the parameters hidden in dnn to justify the results.

Question 3

What is feature importance?

Answer 3

It indicates how valuable the feature is relative to other features.

Question 4

What are the uses of feature importance?

Answer 4

Variable selection and data leakage check

Question 5

What can Vertex Explainable AI do?

Answer 5

It integrates feature attributions into Vertex AI and helps understand model’s outputs.
It tells how much each feature in the data contributed to the predicted result.
It can be used to identify bias and understand how to improve.

Question 6

What are the models supported by Vertex Explainable AI?

Answer 6

AutoML image models (classification)
AutoML tabular models (classification and regression)
Custom-trained TensorFlow model (tabular and image)

Question 7

What is feature attribution?

Answer 7

Feature attributions indicate how much each feature in your model contributed to the predictions for each given instance.

Question 8

What are the three methods offered by Vertex Explainable AI?

Answer 8

Sampled Shapley: Tabular classification and regression.
Non-differentiable models (ensembles of trees and neural networks with encoding and rounding tasks)
AutoML, Custom-trained (any container)
Integrated gradients: Tabular classification and regression.
Image classification.
Differentiable models (neural networks)
AutoML, Custom-trained TF model (prebuilt container)
XRAI: Image classification
AutoML, Custom-trained TF model (pre-built container)

Question 9

What are differentiable and non-differentiable models?

Answer 9

Differentiable models: You can calculate the derivative of all the operations in your TensorFlow graph.
Non-differentiable models: Includes non-differentiable operations in the TensorFlow graph, e.g., decoding and rounding tasks.

Question 10

What is Vertex AI Example-Based explanation for?

Answer 10

It is used for misclassification analysis and can enable active learning so that data can be selectively labeled.

Question 11

What are data bias and fairness?

Answer 11

Data bias: When certain parts of the data are not collected.
Fairness: Ensure biases in the data do not lead to treating individuals unfavourably.

Question 12

How to detect bias and fairness?

Answer 12

Explainable AI feature attributions
Feature overview functionality through an interactive dashboard with What-If Tool.
Language Interpretability Tool for NLP

Question 13

What are the two concepts of ML solution readiness?

Answer 13

Responsible AI
Model governance

Question 14

What are the responsible AI tools?

Answer 14

Explainable AI
Model cards:
what a model does,
intended audience,
maintenance staff
TensorFlow open source toolkit: provide model transparency.

Question 15

What is model governance?

Answer 15

Provide guidelines and processes to help employees implement the company’s AI principles.

Question 16

What are the ways to achieve model governance?

Answer 16

Human to review model outputs
Responsibility assignment matrix for each model by task
Maintain model cards to track model versioning and data lineage
Evaluate the model on benchmark datasets and validate the model against fairness indicators
Use What-If tools to understand the importance of different data features

Question 17

How to set up explanations in Vertex AI?

Answer 17

Configure explanations for custom-trained models.
No configuration is needed for AutoML tabular regression and classification.

Question 17

What are the explanations available?

Answer 17

Online: Synchronous requests to Vertex AI API
Batch: Asynchronous requests to Vertex AI API
Local kernel: Use User-Managed Vertex AI Workbench notebook to get explanations without deploying the model.

Question 18

How do you get explanations when you use TensorFlow?

Answer 18

Use Explainable AI SDK’s save_model_with_metadata() to infer your model’s inputs and outputs and save this explanation metadata with your model.
Load the model into the Explainable AI SDK using load_model_from_local_path()
Call explain() with instances of data and visualize the feature attributions