Photo by Franck V. on Unsplash
Do you remember high school science labs where the teacher would do a demonstration, then you and the rest of the class would reproduce the process to get the desired results? Somehow, I found creative ways to not get the results I needed.
Imagine that same teacher demonstrating an experiment for an artificial intelligence tool to determine creditworthiness. As the teacher is going through the demonstration, you notice him making comments like, “I’m not exactly sure how I got to this part of the experiment,” or “Actually, I can’t tell you about that part of the experiment. It’s proprietary information.” Finally, the teacher tasks you and the rest of the class to reproduce his work. Bonne chance! This sort of thing is actually happening in scientific fields including artificial intelligence. Why is this an important problem and what can be done about it?
For the past several years, there has been quite a bit of conversation about the reproducibility problem facing the artificial intelligence world. Labs conducting experiments to develop various components of AI are running tons of experiments, finding the answer they are looking for, and publishing the results. Other groups come along and fail at their attempts to reproduce the work.
Black Americans Have the Highest Mortality Rates But Lowest Levels of Life Insurance
Are you prioritizing your cable entertainment bill over protecting and investing in your family?
Smart Policies are as low as $30 a month, No Medical Exam Required
Click Here to Get Smart on Protecting Your Family and Loves Ones, No Matter What Happens
Think about all the things that could be wrong with the creditworthiness example I mentioned earlier. What if the code in there poorly rates the creditworthiness of people from a certain zip code for a reason that doesn’t clearly link to their likelihood to repay a loan? What if that neighborhood also happens to be primarily Black?
Listen to GHOGH with Jamarlin Martin | Episode 43: Jewel Burks
Partpic founder and CEO Jewel Burks discusses how she developed an idea to streamline the purchase, repair and maintenance of parts. Partpic was acquired by Amazon in 2018.
As we weave AI into more areas of our lives we’ve got to be sure that the AI is designed well and mitigates potential negative outcomes for folks. There are a number of efforts out there to address these reproducibility issues like the Allen Institute for Artificial Intelligence’s Show Your Work framework and Joelle Pineau’s Machine Learning Reproducibility Checklist.
An opportunity area to push these reproducibility efforts further is to integrate them into the work different organizations are doing to decentralize AI research. To this point, the size of your budget has determined the amount of computation you can use to run your experiments. Budget and ability to use brute force can be the determining factor in a research team’s ability to achieve meaningful results as opposed to a well-designed process. With decentralized AI initiatives like Microsoft’s Decentralized & Collaborative AI on Blockchain, SingularityNET, Ocean, Erasure, and OpenMined more researchers can get access to data and add their own.
The current reproducibility efforts assume research labs are working in isolation. Imagine these labs taking the step to plug into decentralized AI efforts while also taking the steps to show their work. Not only would researchers increase the ability of teams to reproduce their work, they could take advantage of the collaborative nature of these platforms to speed up the rate at which improvements are made to these experiments.
Imagine hip-hop ciphers where MCs are pushing each other to bring their A-game with their lyrics as each person gets her turn. I could see Black AI researchers bringing this flavor to the world of AI research but it would require a different mode of thinking about who gets credit and what is proprietary. The specter of avoiding mistakes such as cutting an entire neighborhood out of a loan due to their zip code seems to make that cost worth it.
Kwame Som-Pimpong leverages relentless research, a knack for connecting dots, human-centered design approach, and effective communications strategy to help organizations realize their strategic objectives. Over a 10-year career, Kwame has supercharged grassroots political organizing efforts, assessed the effectiveness of U.S. federal agencies, managed an international program, founded a digital media startup, and advised government agencies on delighting their end-users. He earned a BA in Political Science from Davidson College and Master of Public Administration from the University of Georgia. He can be reached at firstname.lastname@example.org.