Data Engines Corporation

Data Engines Corporation develops algorithms for unsupervised inference in machine learning settings. Who judges the judges? How common are labels in a data set for which no ground truth is available? These questions are increasingly common in a world awash with data and recognizers to label it. We create tools that help businesses and researchers exploit this modern conjunction of abundant data processed by noisy annotators. Our breakthrough technologies apply across a plethora of fields, including, but not limited to, bioinformatics, national security, medical imaging, and data retrieval.

We have developed algorithms that allow users to “have their cake and eat it too!” In situations where no ground truth for the data is known, it is possible to accurately measure both statistics of the data and the quality of the recognizers that were used to process it. So far, we have developed algorithms for unsupervised inference in regression, labeling, and ranking. Furthermore, we can readily handle correlated recognizers—a crucial functionality for guaranteeing the reliability of our monitoring approach in real-world situations.

The possibilities that open up when you have a collection of good unsupervised inference algorithms are numerous. Here are a small number of the multitude of applications we can envision as being enabled by this technology:

• Robots can begin to perform self-assessment on their sensors and be able to discount bad or failing ones.
• Large datasets could be turned into training data with precise knowledge of the quality of the labels.
• Collections of middling quality recognizers could be system combined into a robust, high-quality recognizer that can provide reliable estimates of its own uncertainty in a variable environment.
• The degree of correlation among decision makers could be monitored in an online setting.
• Heterogeneous recognizers could be treated as black boxes allowing for human and machine decision makers to be combined in a single framework.