Example-based explanations – Explainable AI
In the case of explanations based on examples, Vertex AI makes use of the closest neighbor search to produce a list of instances (usually taken from the training set) that are most comparable to the input. These examples allow users to investigate and clarify the behavior of the model since users can reasonably anticipate that comparable inputs would result in similar predictions.
Consider the following scenario: users have a model that analyzes photos to determine whether they depict a bird or an aircraft; however, the model incorrectly identifies certain birds as planes. To figure out what is going on, we may extract other photos from the training set that is comparable to the one we are looking at and utilize example-based explanations to explain what is occurring. When we look at those instances, we see that many of the incorrectly identified birds and the training examples that are comparable to them are dark silhouettes and that most of the dark silhouettes that were aircraft were found in the training set. This suggests that users might potentially increase the quality of the model by including more silhouetted birds in the training set.
Explanations that are based on examples may also help identify confusing inputs that might be improved with human labelling. Models that provide embedding or latent representation for input variables are supported. Tree based models which do not provide embeddings for the inputs are not supported in examples-based explanations.
Feature-based explanations is another way of explaining model output based on the features. The amount of contribution that each feature in the model made to the predictions that were made for a particular instance is shown by the feature attributions. When users make a request for predictions, they will get anticipated values that are suitable for the model you are using. Feature attribution info will be provided when users request for the explanations.
Feature attributions work on image and tabular data. They are supported for AutoML and custom trained models. (Classification models only for image data and classification/regression models for tabular data).