The three quality pillars are:
-
Functionality: groups analyses that evaluate how ”well“ an AI module performs a given task (i.e. assessing the suitability of an AI module for an application domain).
-
Comprehensibility: groups analyses that try to open the blackbox and enable stakeholders (model producers, users, or regulators) to interpret decisions and the decision-making process.
-
Robustness: groups analyses that assess how the ML component responds to small changes in the input.
This first release performs functionality and robustness analyses for image classification and regression models. The next versions will include comprehensibility analysis and accept text data.
Using the PyPi package:
pip install ml-model-quality-analysis
TensorFlow 2.3 is required:
pip install tensorflow==2.3
For examples on performing quality analysis for ML models, see the Quality Report Notebook.
For an overview of the performance metrics calculated and when to use each one, see the Metrics Notebook.