Computer Vision News - October 2020

AI application development platform 18 “The algorithm takes a combination of weak signals for labels for your training data, de-noises them, and combines them in an optimal manner to apply training labels to the data at hand,” Paroma explains. “This is at the core of how we make the process of building an end-to-end machine learning pipeline significantly more efficient. We have helped some of the largest banks and healthcare insurance providers in the country go from spending months labelling their data to being able to do so in a matter of days – less than a day for one of the use cases we worked with.” A key takeaway of the algorithm is that it can learn how accurate the different rules are that users write without needing any ground truth hand- labelled data. Itis able to learn just from the overlaps, conflicts, and labelling patterns of the rules people have written, and then use that information to assign the final training labels to the data. wide variety of use cases and domains. The team have built a horizontal platform , and the core algorithm is agnostic to the domain it is working on, so Snorkel Flow can be used across a wide variety of use cases and domains. This generalizable aspect is one of the platform’s biggest strengths. It makes it a one-stop-shop solution for building machine learning pipelines with different application layers on top – such as marketing, IT, finance, customer service, or healthcare – with no risk of overfitting one over the other. Braden is keen to point out that they do not see Snorkel Flow as just a convenient or time-saving solution; it is a fundamentally different approach to machine learning. Historically this has been more of a two-step process, where you collect the dataset and then try to achieve high performance. However, driving model improvement via changes and updates to the data creates a