One of the struggles in the field of data science is striking the necessary balance between human decision making, and automated computer processing. The field of machine learning—which looks to create computing systems that can solve problems on their own—is a perfect example of where human intelligence and mechanical computing power must go hand in hand. Machine learning systems can quickly and effectively assemble a data set, but there can be blindspots in what a machine learning system looks for when collecting data. Humans, on the other hand, are much more competent in determining where there might be holes in a data set.
In his research, Panos Ipeirotis—an Associate Professor at the Center for Data Science and the Stern School of Business—combines the computing power of machines with the problem solving capabilities of humans, as a way of generating more effective machine learning techniques. He recently published a paper titled, “Beat the Machine: Challenging Humans to Find a Predictive Model’s ‘Unknown Unknowns,” which examined the effectiveness of having humans look for the blind spots in machine learning techniques.
Humans can imagine the possible “unknown unknowns” in a data set; machines cannot. When machines are gathering data, they rely on a set of test data to compare their data collection to. This test data is gathered by humans, but is much smaller than the desired data set. During the data collection process, there is no incentive for a computer to try and find the “unknown unknowns” because the machine doesn’t know that these pieces of data are even possible. We spoke with Panos about his research concerning the “unknown unknowns” of machine learning.
How did the project of “unknown unknowns” come to be?
We were thinking of better ways to improve our machine learning systems, and we realized that humans are quite competent at figuring out how to create and look for errors in automatic processes.
Can you give a specific example of one of these “unknown unknowns?”
Right now, we’re building a system for detecting misconduct within financial institutions. We have teams of independent financial analysts—who have significant domain expertise—to come up with “unknown unknown” cases of financial misconduct. In this situation, the data set already includes past cases of misconduct, and so we ask analysts to come up with cases that haven’t been caught yet, or might not be anticipated.
How does the difference between machine intelligence and human intelligence factor into your research?
Machines implicitly operate with the assumption of the “closed world”, i.e., the world of the data at hand. However, the data at hand is not necessarily representative of the world as a whole. Humans are much more capable in coming up with examples that could be missing from a data set. In the case of AlphaGo—Google’s machine learning program that was taught how to play the game Go—the only game that it lost was due to an “unexpected” move from its opponent, which indicates that the machine was not able to cope with “surprises.”
How do you measure a system’s performance when the system might not even know all the variables?
There are multiple metrics that can be used to measure performance in machine learning, and most of these metrics will suffer from that same “closed world” assumptions. So we optimize our models with this “unseen” test data, so that we’re testing against the known world, not just the world of the data set.
What were your findings with this project?
Humans should always be part of machine learning solutions, as they can guide machine learning systems to learn about things that the systems doesn’t yet known—the “unknown unknowns.”