Machine-learning is playing a growing role in data analysis, including in the field of consumer electronics testing and development.
However, there remain doubts about the effectiveness and accuracy of some machine-learning models.
To address these doubts, a team of researchers headed by computer scientist Tamara Broderick, associate professor in MIT’s Department of Electrical Engineering and Computer Science (EECS), have developed a classification system — a “taxonomy of trust” — that defines “where trust might break down in a data analysis and identifies strategies to strengthen trust at each step.”
The researchers began by identifying the points in the data analysis process where trust might break down.
These include: analysts’ choices about what data to collect; the selection of algorithms to fit the model; and the type of use code used to run those algorithms.
In the classification system, some components can be checked for accuracy in a measurable way (for example, ascertaining whether the code has bugs). Other factors are harder to quantify, says Broderick.
“What I think is nice about making this taxonomy, is that it really highlights where people are focusing. I think a lot of research naturally focuses on this level of ‘are my algorithms solving a particular mathematical problem?’ in part because it’s very objective, even if it’s a hard problem,” said Broderick.
“I think it's really hard to answer ‘is it reasonable to mathematize an important applied problem in a certain way?’ because it's somehow getting into a harder space, it's not just a mathematical problem anymore.”
One concrete step covered in the taxonomy is checking the viability of the code being used. One way to catch bugs is to test whether code is reproducible.
However, as things stand this can difficult since sharing code alongside published work is not always a requirement or the norm, and recreate code from scratch is difficult, if not impossible, as models growing in complexity over time.
“Let’s just start with every journal requiring you to release your code,” said Broderick. “Maybe it doesn’t get totally double-checked, and everything isn’t absolutely perfect, but let’s start there.”
The research has been published recently in the journal Science Advances.