Mitigating AI Bias, with…Bias
This article was originally posted on my Medium blog as part of my Data Trust series of talks and writing. The purpose of these articles are to break down complex but important socio-technical topics in a manner that is accessible to both practitioners and non-practitioners.
With great power comes complexities
Most tools we use today leverage AI/ML from the moment we wake up and while we sleep.
Humans build Machine Learning, and humans are inherently biased.
Since humans aren’t perfect, we encode our biases into the data we use to train AI.
Encoding our biases into machine learning training often results in unfair treatment or harm against protected classes of people, groups of people, and individuals.
We’ll cover how to get around this but there is one a catch. You cannot guarantee fairness for your AI application, and this is because there is no single definition of fairness. How you define fairness for one solution will be different for another.
Human-centered machine learning practitioners leverage a comprehensive set of fairness metrics and bias mitigation algorithms to solve this catch. Essentially countering bias we don’t want with good bias using these methods.
A lot has been written about fairness, statistical, and algorithmic bias in the last few years, as you probably know, so while I may not dive deep into all the nuances, know that eliminating discrimination in your AI is hard work. I have shared more resources at the end of this post.
We’re only human, and we encode our bias into?AI.
Story time
I’m sharing my lived experience from an incident 18 years ago in Richmond, Virginia, to make a larger point on the complexities of mitigating bias so that you can reduce risk.
It’s 2004; Done with work, I changed into my sweats and white t-shirt, stuffed work clothes into my gym bag, and started the 15-minute walk from my office to the gym.
I went to cross a one-lane road but hopped back onto the sidewalk as two police vehicles, one driving in the opposite direction of traffic, blocked my path.
Officers hopped out, yelling what many black males have heard in many US cities.
“Get down on the ground?now!”
Between the time it took to take off my headphones, I was on the pavement, face to the ground, as my arms stretched painfully behind me while I got detained.
It was a case of mistaken identity, and I fit the description of an apparent criminal.
They let me go. I was traumatized, so I turned around and went home. When thinking about your AI team look for a diverse and representative group because your final solution will be better for it.
Do members of your AI team have similar lived experiences?
…and do they have a voice?
As a digital product delivery practitioner, my lived experience influences how I lead teams to ship software, including AI. Your team should prioritize bringing together a diverse and representative group of people in your organization across functions.
Empowered diverse and representative AI teams bring objective perspectives with metrics to show how software can cause harm in addition to success measures or quantifiable objectives to deliver safe, secure, private, performant, and usable software.
Bias Mitigation Strategies for AI/ML aka adding good?bias
Image adapted from:?Mehrabi et al. Survey on Bias & Fairness in ML?Cornell University ARXIV research?2022
Let’s walk through a thought experiment with how vision-based AI could assist a law enforcement officer without bias in a high-stakes scenario like this.
Assuming you determined an AI is best to solve the problem, defined the need, and you need to map those needs to features. We could look at race, my gender, perhaps my age, the clothes I had on, etc.
Fortunately (and thankfully) race, gender, and age are protected classes in United States labor laws, so that’s a “no go.”
So what do we do…ok let’s look to an AI to use my clothing to help an officer decide to detain or leave me be.
It is worth saying that this is just an example. In a real-world scenario, clothing alone would be a terrible correlation to identifying a potential criminal.
Strategize — Principles and?metrics:
Collect a diverse and representative group of people in your organization across functions to determine if this is a problem an AI is best suited to solve.
Work together to agree on values and fairness objectives and indicators as metrics to address with your team and customers based on your values.
As part of the Strategize phase, you codify one of many definitions of fairness with these baseline metrics.
Amongst other signals, we’re looking at clothing for our thought experiment. We may look at a few examples of fairness metrics. Metrics like:
领英推荐
There are more. AIF360 has quite a few of these if you want to dig in https://aif360.readthedocs.io/
Synthesize — Ask the right questions:
Classic question of how do you verify a sufficient sample size, and quality of data for training?
Be sure to watch out for secondary effects. These are unexpected drawbacks or perverse or opposite results that lead to unintended consequences.
Removing protected classes from your training data in the example above may not be enough. Consider what if we are wrong in how weighting the shoes someone is wearing? Or what if this AI-drive solution is used maliciously to filter out people dressed a certain way?
How do we teach a machine to consider the historical trauma, the discrimination, and opportunity denial black and brown people have endured due to historical and societal?issues?
I recently shared a panel with Dr. Nicol Turner Lee of the Brookings Institute and she mentioned something that stuck with me, that socio-technical harm is also experienced through ‘Traumatized data’. Trauma in the sense of collective psychological, emotional and cognitive distress experienced by African Americans historically such as my perception of bias from my incident eighteen years ago.
If you haven’t already, check out the article, Dr. Turner Lee is also an amazing human so give her a?follow.
So back to the story. Was I unjustly harmed? Yes. Physically then and psychologically now.
Eighteen years later, I still think a lot about where this incident could have ended me. I’m traumatized by this incident.
The key takeaway from this and the next section is to understand the payoffs of being right vs. being wrong so that you can fight bad bias with mitigation and interpretable good bias.
Analyze then train — Lead with your fairness indicators
Weave your AI bias mitigation approach throughout your data science lifecycle.
Data Collection. Dataset sourced and inspected for sample diversity. Look into concepts like data-centric AI, principle-centric AI. I have found resources like Google Research PAIR approach to Data Collection and Evaluation truly helpful and I think you will too.
Pre-processing. If you can modify the training data and have fairness metrics for your datasets, then run tests on raw data to detect bias and use pre-processing algorithms to mitigate risk. A pre-processing algorithm such as a re-weighing algorithm where you change the weights applied to your training samples can be used.
During training or in-processing. If you can modify the learning algorithm, then an in-processing algorithm such as adversarial debiasing can be used to mitigate bias.
Post-processing and classifier metrics. AI Bias mitigation can be applied to predicted labels by tuning your labels to address harms in underprivileged and protected classes. However, at times you may find that you can’t modify the training data or the learning algorithm. In this case, a post-processing algorithm can be used.
Reprocess and or retrain.
You should make sure you are running unit tests for accuracy and discrimination along the way, then deploy.
Interpreting machine learning models and Explaining machine learning predictions
Doing all of this work and not being able to explain impacts of a feature towards a model’s prediction to your various stakeholders isn’t quite useful.
NIST has a good “explainer” of Explainable AI (XAI). XAI is a field on ML interpretability techniques whose aims are to understand machine learning model predictions and explain them in human and understandable terms to build trust with stakeholders, both internal teams and for the user.
Interpretability on the other hand focuses on model understanding techniques, while explainability more broadly focuses on model explanations and the interface for translating these explanations in human understandable terms for different stakeholders.
I will be covering XAI and related topics from the angle of Data, Feature, and Model Trust in future blog posts and speaking engagements. Stay tuned.
Improve — Build AI value that avoids?harm
Keeping a human in the loop with external interventions on your raw data is the final lesson.
At this point, you have your definition of fairness and related metrics, you have your bias mitigation algorithms woven into your ML lifecycle, and have your software using your model out in the wild.
Continuous capture of user feedback is important here. There are a few patterns that I can share from a human-centered ML standpoint but I’ll use my 3C (Context, Choice, Control) framework:
If you are calibrating for trust with your solution, you must also provide enough context for how their data or feedback will be used to better the experience, and provide them the ability to opt out.
When there is no human in the loop (a diverse set of stakeholders) to review, refine, build, deploy, evaluate, and monitor unintentional or secondary effects, it leads to these recurring incidents.
My thought experiment fell apart quickly because eliminating bad bias is hard work. By understanding where the bias is, we can introduce good bias through these strategies and approaches.
Additional reading:
To hear more on these topics and if you can make it out, join me in-person or remotely by registering for Refactr Tech in Atlanta and at the Strangeloop Conference in St. Louis this fall.
Noble currently leads Product and the AI practice at Ventera Corporation in Virginia, US.
Reach me on Twitter
Technology, AI, Product and Privacy Counsel|Responsible AI @Accenture
2 年Fantastic article!