Overview of interpretability and transparency
1. Overview of interpretability and transparency
Welcome to AI Interpretability and Transparency. This module consists of eight lessons. Today you will learn to explain what interpretability and transparency mean in machine learning. Identify various interpretability techniques used to understand model behaviors. Discuss the various interpretability techniques that can be applied to models. Describe tools you can use to apply interpretability techniques to your machine learning models describe toolkits you can use to ensure data and model transparency. Let's start with an overview of interpretability and transparency. Interpretability aims to understand machine learning model behaviors in many ways. A machine learning model is a function that converts inputs to outputs. Let's say that in this image classification example, our machine learning model receives an image and output classification result of a husky dog. It seems good, but why did our model say that it was an image of a husky dog? What kind of characteristics are used to judge an image of a husky dog? How can we understand the reasoning behind it? As we progress through this module, we explore multiple techniques to investigate these questions. It should be noted that in some cases, the term explainability is used instead of interpretability. Other times, the two terms have slightly different definitions, with interpretability meaning a deeper and wider understanding of models and holistic systems, and explainability is used for more specific and technical methods. In this course, we define and use interpretability and explainability as having similar definitions. Now, when it comes to transparency, it's very important to obtain trust from stakeholders, including users. Usually, machine learning applications involve multiple modules such as data collection system, data processing pipeline, machine learning, model evaluation system serving systems, and so on. Since each module is dependent on each other, it's always important to make documentation for each system and communicate based on that information. Google offers some useful tools and frameworks for this purpose, which we will discuss in more depth later on. At a high level, there are three stakeholder groups that benefit from interpretability and transparency as it relates to AI systems. For engineers, the focus is more on interpretable ML techniques to model understanding and improve performance. More complex models can be difficult to debug, understand and control interpretability and transparency. Methods help with this. For users, the focus is on trust. For model consumers. Users might not care about internal mode, but are very interested in understanding the impact of model predictions. Explainable systems build trust with these end users that show how model decisions are reliable and equitable. For regulators, the focus is on ensuring that model decisions are in compliance with laws and do not amplify undesirable bias from underlying datasets. Interpretability explanations provide auditable metadata. This allows regulators to trace unexpected predictions back to their inputs to inform corrective actions. Although interpretability is an active area of research, the task is not easy because interpretability issues apply to humans and AI systems. After all, it's not always easy for a person to provide a satisfactory explanation of their own decisions. Understanding complex AI models, such as deep neural networks, can be challenging, even for machine learning experts. There is typically a trade off between complexity and interpretability of models. Understanding and testing AI systems offers new challenges compared to traditional software. Traditional software is essentially a series of if then rules that, through some chasing, can be interpreted and debugged. With AI systems that not only have code but also data and models, it is much harder to pinpoint one specific bug that leads to a faulty decision.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.