Tackling bias in machine learning models

• Read original article here

By Anam Mahmood
Published February 14, 2022
Artificial intelligence (AI), machine learning, and deep learning have become buzzwords that we hear often. But, what is the difference between them? Briefly, AI tries to make computers intelligent to mimic the cognitive functions of humans. So, AI is a general field with a broad scope including computer vision, language processing, creativity, and summarization. Machine learning is the branch of AI that covers the statistical part of artificial intelligence. It teaches the computer to solve problems by looking at hundreds or thousands of examples, learning from them, and then using that experience to solve the same problem in new situations. Deep learning is a specialized field of machine learning where computers can learn and make intelligent decisions on their own. Deep learning involves a deeper level of automation in comparison to most machine learning algorithms.
Looking deeper into AI
AI is about teaching machines to learn and how to act and think as humans do. Another dimension is about how you get the machines to impart more of the cognitive and sensory capabilities on the machines. So, it is about analyzing images and videos, about natural language processing and understanding speech, and about pattern recognition and computer vision.
The capabilities of artificial intelligence are not hidden. The intelligence of machines is undergoing a major transformation by continuous self-learning improvements. However, these AI models are still a black box. And, their decisions are often questioned by the clients. The research is moving faster than ever on improving and optimizing the algorithms, but this alone won’t suffice. The conversations around building trust in AI are often a point of interest for developers, sales, and marketing teams who work directly with clients. Therefore, it’s important to look into it.
The application of AI algorithms in domains such as criminal justice, credit scoring, and hiring holds unlimited promise. At the same time, it raises legitimate concerns about algorithmic fairness. AI systems are deciding everything, from which resumes are considered to which insurance claims will be accepted to who gets their loan approved, and even to who receives parole.
Some examples of bias in AI that have surfaced recently include:
A popular credit card company was accused of gender bias. The credit card company ran into major problems when users noticed that it seemed to offer smaller lines of credit to women than to men.
A large, worldwide company created a hiring tool that contained algorithms that systematically discriminated against women applying for technical jobs, such as software engineer positions.
When training machine learning algorithms, it’s imperative to guard against the misapplication of race, gender, religion, or other characteristics in the decisions that AI systems make. It’s often a question of legality, but it’s also a question of basic fairness.
There are different ways you can adopt responsible AI, including:
Have a diverse team. Organizations with diverse teams do better with ensuring diverse representation in their data and AI pipelines, helping to avoid bias.
Have ethical AI practitioners on the AI team to ensure that the five focal points are covered.
Have a framework that works for your team and covers the five focal points of accountability.
Use Design Thinking with the user in mind. Understand user needs and challenges, and design for those.
Use open source and commercial tools to address fairness and explainability.
All of these ideas can help you adopt responsible AI. If responsible AI is not a factor, then serious consequences can occur. For example, with a self-driving car there can be many ethical questions or issues that emerge, such as if the car must decide between running into a sign and hurting passengers in the vehicle, or running into pedestrians on the side of the road, but potentially saving the passengers. How do you make those decisions? It’s an interesting area with lots of open questions, and it’s unclear how you can properly define regulations so that different companies that are manufacturing the vehicles all behave in a consistent, ethical manner that meets expectations.
Now, imagine owning a computer vision company that builds AI classification models for people in the healthcare industry to diagnose cancer using MRIs, CT scans, and X-rays. It can be difficult for a doctor to rely on the diagnosis suggested by an AI model when a person’s life is involved. Therefore, building trusted AI pipelines has become increasingly important within AI applications.
Considerations for building trusted AI pipelines
How do you build trusted AI pipelines? You must consider robustness, fairness, and explainability.
Robustness: Robustness measures the stability of the algorithms’ performance when a model that is deployed in the real world is attacked and noise is introduced in the training data. It characterizes how effective your algorithm is while being tested on the new independent (but similar) data set. This ensures that the model’s algorithm can handle the unseen, perturbed data. It addresses the questions of estimating uncertainties in its predictions and whether the model is robust.
The Adversarial Robustness 360 Toolbox is a library that is dedicated to adversarial machine learning. Its purpose is to allow rapid crafting and analysis of the attack, defense, and detection methods for machine learning models. Applicable domains include finance and self-driving vehicles. ART provides an implementation for many state-of-the-art methods for attacking and defending classifiers.
Fairness: Unfair biases can exist in the data that is used to train the model and in the model’s decision-making algorithm. Fairness emphasizes the identification and tackling of the biases that are introduced in the data. This ensures that a model’s predictions are fair and do not unethically discriminate.
The AI Fairness 360 is an open source library to help detect and remove bias in machine learning models. AIF360 converts algorithmic research from the lab into practice. Applicable domains include finance, human capital management, healthcare, and education. The AI Fairness 360 Python package includes a comprehensive set of metrics for data sets and models to test for biases, explanations for these metrics, and algorithms to mitigate bias in data sets and models. It contains over 70 fairness metrics and over 10 bias mitigation algorithms.
Explainability: Explainability shows how a machine learning model makes its predictions. It gives an improved understanding of the model by clarifying how the model works. It is essential to data scientists for detecting, avoiding, and removing its failure modes; to SMEs and customers for earning public trust in the algorithm; and for introducing effective policies to regulate the technology.
The AI Explainability 360 toolkit is built on an open source library and helps to explain AI and machine learning models and their predictions. This includes three classes of algorithms: local post-hoc, global post-hoc, and directly interpretable explainers for models that use image, text, and structured or tabular data. The AI Explainability 360 Python package includes a comprehensive set of explainers, both at a global level and a local level.
A comprehensive bias pipeline
Artificial intelligence is becoming a crucial component of enterprises’ operations and strategies, and there’s a growing demand for fairness, accountability, and transparency from machine learning systems. You must keep in mind that training data isn’t the only source of possible bias. It can also be introduced through inappropriate data handling, inappropriate model selection, or incorrect algorithm design. Bias can also affect the usage data.
What you need is a comprehensive bias pipeline that fully integrates into the AI lifecycle. Such a pipeline requires a robust set of checkers, de-biasing algorithms, and bias explanations.
To give your clients the confidence that they need to responsibly take advantage of AI today, you must find ways to instill transparency, explainability, fairness, and robustness into AI. The following list provides five ethical focus areas.
Accountability: Human judgment plays a role throughout a seemingly objective system of logical decisions. It is humans who write algorithms, who define success or failure, who make decisions about the uses of systems, and who might be affected by a system’s outcomes. Every person who is involved in the creation of AI at any step is accountable for considering the system’s impact in the world, as are the companies that are invested in its development.
Value alignment: AI works alongside diverse, human interests. People make decisions based on any number of contextual factors, including their experiences, memories, upbringing, and cultural norms. These factors provide a fundamental understanding of “right and wrong” in a wide range of contexts at home, in the office, or elsewhere. This is second nature for humans because we have a wealth of experiences to draw upon. Today’s AI systems do not have these types of experiences to draw upon, so it is the job of designers and developers to collaborate to ensure consideration of existing values. Care is required to ensure sensitivity to a wide range of cultural norms and values.
Explainability: In general, you don’t blindly trust those who can’t explain their reasoning. The same goes for AI, perhaps even more so. As AI increases in capabilities and achieves a greater range of impact, its decision-making process should be explainable in terms that people can understand. Explainability is key for users interacting with AI to understand the AI’s conclusions and recommendations. Your users should always be aware that they are interacting with an AI. Good design does not sacrifice transparency in creating a seamless experience.
Fairness: AI provides deeper insight into our personal lives when interacting with our sensitive data. Because humans are inherently vulnerable to biases and are responsible for building AI, there are chances for human bias to be embedded in the systems that are created. It is the role of a responsible team to minimize algorithmic bias through ongoing research and data collection that is representative of a diverse population.
User data rights: AI must be designed to protect user data and preserve the user’s power over access and uses. It is your team’s responsibility to keep users empowered with control over their interactions. The percentage of people who feel that having control of their information is important and that it is unacceptable for companies to use user information without permission will rise as AI is further used to either amplify your privacy or undermine it. Your company should be fully compliant with the applicable portions of the European Union’s General Data Protection Regulation and any comparable regulations in other countries to ensure that users understand that AI is working in their best interests.
Designers and developers of AI can help mitigate bias and disenfranchisement by practicing within these five areas of ethical considerations.
AI pipelines
Now, take a look at how the AI pipeline works. For example:
Your portfolio has four key pillars that can be mixed and matched with other solutions in the market depending on your need.
You have Watson Knowledge Catalog on the organizing and governing side.
The build phase wants to look into your data, prep it, and build some models or dashboards that can give you predictive insight into your data. This is where Watson Studio comes in.
After you have built your models, the next step is to deploy those models into production, and to use the output predictions to influence business decisions. It can be easy to move your models from development to production with Watson Machine Learning.
So, you have governance, you’ve built your model, you’ve deployed it, and now you’re getting the predictions of whom you should give a loan to or who you should market to. The natural next questions are “Do you trust the output of the model? Do you trust its decision? How do you make sure that the model isn’t biased by the person who built it?” And that’s exactly where Watson OpenScale comes in.
Stepping back, you can see that these solutions fulfill the whole end-to-end AI journey. It is easy to deploy, manage, and monitor your models, which goes back to the team sport strategy. And all of these things work together seamlessly. You can pick specific pieces or you can bring them together to complete your data science and AI journey.
Mitigation methods
Now you know the pillars of trustworthy AI. But, you must test them too. The pillars of trustworthy AI have mitigation approaches that are grouped into the same three categories.
The first category contains preprocessing methods that improve the statistics of the training data set. The second category constrains the training of the AI in favorable ways. The third category post-processes the predictions that are produced by the AI.
Mitigation methods are full-fledged AI algorithms themselves. The main goal of these algorithms is to adapt the data to better match the wanted world and to make the AI model perform as best as it can in the worst-case scenario.
Summary
This article has given you an overview of some examples of how bias can be in your machine learning models as well as mitigation ideas to try to remove as much of that bias as possible.
Share our content