You Are the Only Possible Oracle: Effective Test Selection for End Users of Interactive Machine Learning Systems

Abstract

Abstract-How do you test a program when only a single user, with no expertise in software testing, is able to determine if the program is performing correctly? Such programs are common today in the form of machine-learned classifiers. We consider the problem of testing this common kind of machine-generated program when the only oracle is an end user : e.g., only you can determine if your email is properly filed. We present test selection methods that provide very good failure rates even for small test suites, and show that these methods work in both large-scale random experiments using a "gold standard" and in studies with real users. Our methods are inexpensive and largely algorithm-independent. Key to our methods is an exploitation of properties of classifiers that is not possible in traditional software testing. Our results suggest that it is plausible for time-pressured end users to interactively detect failures-even very hard-to-find failures-without wading through a large number of successful (and thus less useful) tests. We additionally show that some methods are able to find the arguably most difficult-to-detect faults of classifiers: cases where machine learning algorithms have high confidence in an incorrect result.

What you see is what you can change: Human-centered machine learning by interactive visualization

“…They also stress the importance of accounting for user behavior, and the potential benefits of collaborative research between ML experts and the human-computer interaction community. Similarly, Groce and colleagues investigate sample selection strategies to test classifiers effectively via systematic feedback requests to end users [23].…”

The Machine Learning Algorithm as Creative Musical Tool

“…Interactive machine learning is the term first used by Fails and Olsen Jr [2003] to describe an approach in which humans can iteratively add training examples in a freeform manner until a model's quality is acceptable; it has since come to encompass a slightly broader set of techniques in which human users are engaged in a tight interaction loop of iteratively modifying data, features, or algorithm, and evaluating the resulting model ( Figure 5). Fails and Olsen originally proposed this approach in the context of computational image analysis, but it has since been applied to a variety of other problems, such as webpage analysis [Amershi et al, 2015], social network group creation [Amershi et al, 2012] and system "debugging" [Groce et al, 2014]. Figure 5: Interactive machine learning involves free-form iteration through different types of changes to the learning algorithm and data, followed by re-training and evaluating the modified model.…”

Crossed Wires

“…For example, in our task, the sensor readings were displayed in the Arduino IDE but they could be incorrect because of miswiring bugs in the sensor or LED connections. Fault localization strategies could also be communicated in such a development environment, possibly drawing from existing troubleshooting checklists (such as [12,42] [8,19]. Approaches in formally verifying physical circuits [13] could also be useful in this respect.…”