Validate the model

Explainability view

The Explainability view is used to analyze and validate a trained AI model.

The results shown are based on the test data split of the data set used. By default, 20% of the data set is automatically reserved as test data. The remaining 80% is used for model training.

The test data is not used during training and is used solely to evaluate the model's ability to generalize.

Model statistics

The top left section represents the model's key performance metrics.

• Test accuracy: Describes the proportion of all correctly classified test data. Accuracy provides a general overview of the model's overall performance.
• Test precision: Describes the accuracy of positive predictions. Precision measures how many samples classified as positive are actually correct.
• Test recall: Describes the model's ability to correctly identify relevant samples. A high recall means that only a few relevant samples are missed.
• Test F1: The F1 score combines precision and recall into a single metric. This metric is particularly helpful when dealing with unbalanced data sets.

Confidence

The Confidence section displays statistical measures of model confidence:

• Average value (Mean)
• Minimum value (Min)
• Maximum value (Max)

Confidence describes the model's certainty in making a prediction.

The Confidence Distribution visualizes the distribution of all prediction confidences as a histogram. Each bar represents a confidence interval (bin) and shows the number of predictions within that interval. This makes it possible to assess whether the model makes mostly safe or unsafe decisions.

Interactive filtering

By clicking on a histogram range, you can filter the results to a specific confidence interval. The filtered results are then displayed in the Prediction behavior section. Only samples within the selected confidence range will appear there. The active filter is displayed in the upper section of Prediction Behavior and can also be deleted there.

Confusion matrix

The confusion matrix shows the distribution of correct and incorrect predictions by class.

• Rows correspond to the actual classes (true labels, as labeled by humans).
• The columns correspond to the predicted labels.

The diagonal of the matrix contains correctly classified samples. Values outside the diagonal represent misclassifications.

The size of the matrix depends on the number of classes in the data set.

Interactive filtering

By clicking on individual matrix elements, you can filter for specific combinations of predictions. This makes it possible to analyze misclassified samples or correctly classified samples in a targeted manner. The filtered results are represented in the Prediction behavior section.

Prediction behavior

The "Prediction Behavior" area displays all samples from the test data set, including the prediction results, model confidence, and attention map.

Correct and incorrect predictions can be shown or hidden separately. In addition, the results can be sorted by confidence level.

Attention maps

For each sample displayed, an attention map can optionally be represented. The visualization highlights the image areas that contributed most significantly to the AI model's decision. This makes it clear which characteristics were relevant for the classification. You can disable the Attention Map using the switch in each image.