Overview of Fairness and Bias
1. Overview of Fairness and Bias
Welcome to AI fairness & Bias. This module consists of eight lessons. Today, you will learn to: Define types of unfair bias in AI; Discuss why AI fairness is important for machine learning and difficult to achieve; Identify possible causes of biases in AI systems; Recognize AI fairness best practices; And explore tools and techniques to mitigate bias in datasets and models. Let’s start with an overview of fairness and bias. Bias is stereotyping, showing prejudice, or favoritism towards some things, people or groups over others. It is easy to say, but usually challenging to realize the biases we have, because they are hidden inside our own perspectives. Let’s do a quick experimentation. Picture a shoe. What kind of shoe did you picture? Here are some examples of a shoe. Which one is closest to what you visualized? Something like a sneaker on the left? Or a high heel on the right? Although not all biases are harmful, it is important to note that we naturally obtain some tendencies in perceptions, thought processes, and common senses, depending on a group you belong to, or information you get. What if you are asked to visualize “a scientist”? Did you envision a particular gender group or racial group? There are over 100 different types of human biases listed in Wikipedia’s catalog of cognitive biases. When developing products for AI systems, it’s important to be aware of common human biases that can manifest in your data and in your model. This way, you can take proactive steps to mitigate the effects of biases. Let’s look at the five most common biases. Reporting bias occurs when the frequency of events, properties, and/or outcomes captured in a dataset, does not accurately reflect their real-world frequency. This bias can arise because people tend to focus on documenting circumstances that are unusual or especially memorable, assuming that the ordinary is generally accepted. Automation bias is a tendency to favor results generated by automated systems over those generated by non-automated systems, irrespective of the error rates of each. Selection bias occurs if a dataset’s examples are chosen in a way that doesn’t reflect their real-world distribution. Selection bias can take many different forms, such as: Coverage bias. This occurs when data is not selected in a representative fashion. Non-response bias, or participation bias. This happens when data ends up being unrepresentative due to participation gaps in the data collection process. And sampling bias, which occurs when proper randomization is not used during data collection. Group attribution bias is a tendency to generalize what is true of individuals to an entire group to which they belong. Two key manifestations of this bias are in-group bias and out-group homogeneity bias. In-group bias occurs when there is a preference for members of a group to which you also belong, or for characteristics that you also share. Out-group homogeneity bias occurs when there is a tendency to stereotype individual members of a group to which you do not belong, or to see their characteristics as more uniform. An implicit bias occurs when assumptions are made based on one’s own mental models and personal experiences that do not necessarily apply more generally. A common form of implicit bias is confirmation bias, where model builders unconsciously process data in ways that affirm pre-existing beliefs and hypotheses. In some cases, a model builder might actually keep training a model until it produces a result that aligns with their original hypothesis. This is called experimenter’s bias. So, how about AI systems? As you know, they usually have multiple steps, including data collection and labeling, training, evaluation, and deployment. Throughout the lifecycle, bias can enter into the system as a systematic error introduced by sampling or reporting procedures. If the data collection has wrong assumptions or implementation, the dataset itself ends up having a lot of biases, such as selection bias. Engineers can inadvertently train models by feeding the bias dataset, and machine learning models will easily find and repeat the bias patterns and predictions. When building models, bias can also enter in the way model components and logic are defined. Bias can even appear after deployment from feedback loops and model iterations. You can rarely identify a single cause of or a single solution to these bias problems. What happens far more often is the existence of various causes that interact in ML systems to produce problematic outcomes. This results in a range of solutions needed. It’s important to note that not all types of bias are necessarily unfair. The definition of fairness is slightly more restrictive. Decisions made by computers after a machine-learning process might be considered unfair if they were based on variables considered sensitive. Examples of such variables include gender, ethnicity, sexual orientation, or disability. With these variables, fairness also means ensuring everyone has the support they need. It is important to work towards AI systems that are fair and engaging for all, as the impact of AI increases across sectors and societies. Beyond recommending books and television shows, AI systems can be used for more critical tasks, such as matching people to jobs and partners, identifying fraudulent transactions, or identifying a person who is crossing a street on a red light. So what is the opportunity that exists? AI systems have the potential to be fairer and more accessible at a broader scale than decision making processes based on ad-hoc rules or human judgments. And the risk that exists? Unfairness in such systems can also have a negative, widescale impact. In all this, AI fairness is actually a difficult task to achieve. First of all, AI models learn from existing data, and an accurate model might learn or even amplify problematic preexisting biases. Second, even with the most rigorous and cross-functional training and testing, it is a challenge to build systems that will be fair across all situations. Third, there is no standard definition of fairness, whether decisions are made by humans or machines. Identifying appropriate fairness criteria for a system requires accounting for considerations such as user experience, cultural, social, historical, political, legal, and ethical. Several of which may have trade-offs, even for situations that seem simple. People may disagree about what is fair, and it may be unclear what point of view should dictate policy, especially in a global setting. Finally, fairness metrics can be incompatible and impossible to satisfy simultaneously. Hence, it’s impossible to define a universal metric for fairness. Don’t be discouraged, though. It just means that fairness needs to be defined contextually for the given AI problem.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.