Why Explainable AI( XAI)?
Explainable AI(XAI) is artificial intelligence in which the results are easily understood by humans. (Wikipedia, 2021).
In the definition of Explainable AI(XAI), Machine Learning (ML) models are often called black box models because they use a set of layers, neurons and weights. All of them are just numbers and even designers cannot explain why the decision is given by AI.
In a simple linear regression model to predict the house prices, the number of parameters can be quite small like 3 such as the number of bedrooms, size in square meter and age of the building, but it can be quite complicated in a neural network such as AlexNet, a Convolutional Neural Network Architecture for image recognition. It has 62,378,344 parameters. It is far beyond a human understanding.
On the other hand, Explainability is not essential for a Machine Learning Model. An ML engineer can build an ML model without considering Explainability, and the model can run and give decisions. Johnson also uses a good analogy for Explainability which he compares ML Model without Explainability with using a car without wearing a seat belt. Most of the time, when things go well, we do not need a seat belt, but it is insurance for life in a crash.
AI is getting more integrated into our lives, therefore they should be trusted because they are making decisions previously entrusted by humans. Deep learning algorithms are quite fancy, they do incredible things in our lives as they clean our homes without us, drive our car autonomously, translate texts, generate our 60 years old versions, take the best shot on our cameras by adjusting the settings best possible way and timing, etc. We love them.
However, the deep neural network can be easily fooled with adversarial attacks. A Deep Learning model can classify a man’s image as a woman as seen in Figure 1 and a traffic sign detection algorithm as a part of the autonomous driving model can recognize the altered stop sign as 45 miles per hour sign as seen in Figure 2.
Would you like to be in an autonomous car that can not detect the stop sign properly?
Explainability is quite important to ensure algorithmic fairness, identify bias problems, increase trustworthiness by summarizing the reasons for neural network behavior, producing insights about the causes of their decisions, responding to questions, and being open for audit.
As a business leader, an ML enthusiast, an ML developer, entrepreneur, volunteer, data scientist, data wrangler, data cruncher, whatever you would like to call, explainability is not nice to have, it is a must-have in your AI systems and initiatives.
References
- Gilpin, L. H. et al., 2019. Explaining Explanations: An Overview of Interpretability of Machine Learning.
- Goel, A., 2020. An Empirical Review of Adversarial Defenses.
https://arxiv.org/pdf/2012.06332.pdf - Johnson, J., 2020. Interpretability vs Explainability: The Black Box of Machine Learning.https://www.bmc.com/blogs/machine-learning-interpretability-vs-explainability/
- O’Sullivan, C., 2020. Interpretable vs Explainable Machine Learning. https://towardsdatascience.com/interperable-vs-explainable-machine-learning-1fa525e12f48
- Wikipedia, 2021. Explainable artificial intelligence.
https://en.wikipedia.org/wiki/Explainable_artificial_intelligence
