9.5: Conclusion

Machine learning models should be highly accurate and easy to explain. This is more applicable to models that affect the daily lives of human beings. For example, the potential use of a model for approving or rejecting loans. This might have moral, ethical, and legal consequences. Hence, it is essential to provide a suitable explanation for denying a loan.

Similarly, if a computer vision model is deployed for the home monitoring, we will expect it to do it without fail. If it is unable to identify intrusion, it might have legal consequences for the company who licensed the model. The company might be required to explain why it couldn't detect cases of intrusion. In some cases, the company might be liable to pay for the damages. In such a situation, the company might use the model explanation in their defense.

Model explainability is important in making the model predictions reliable. When the model is predicting correct values and is aided by proper explanations, it will instill confidence amongst intended users to use the model. Similarly, when the model is giving a wrong prediction, it will help fix accountability by identifying features and the values that contribute most to the prediction.

There are many methods for explaining the overall model. Similarly, there are many methods for explaining individual model predictions. We should use multiple methods for model explanation, before making final conclusions on the model. Similarly, for explaining a single instance of prediction, we should try different methods and compare against different methods. If more than one method suggests us the same conclusion, we should include this in our finding. If there are contradicting conclusions from different explanation methods, we should choose the conclusions which are closer to domain knowledge.