Unsupervised learning is a fundamental pillar of machine learning, yet also one of the most challenging. The core challenge lies in how we can gain insight into unlabeled, and possibly unstructured, data in order to make actionable decisions. Businesses use unsupervised learning daily in the form of clustering, for example, but this clustering is only as good as the space in which the data is represented. The Borealis AI unsupervised learning team explores ways to leverage the vast amounts of detailed but unlabeled data that exists in the world. Our primary objective is to find new and creative ways to learn the true latent features and natural distributions of the data. Moreover, while we leverage these methods to find lookalike entities, we also strive to understand the evolution of entities through time and identify new phenomena that are important to the business and to society.
Natural language processing
Natural language processing, a catchall term for the intersection between computer systems and natural language, has existed in some form for the better part of a century. It’s only recently, however, that advances in deep learning have revitalized the field, enabling NLP researchers to tackle previously intractable problems. With open source tools such as Tensorflow’s SLING and Spacy, it’s now possible to quickly create end-to-end models that achieve state-of-the-art performance in natural language understanding tasks, and to apply those models to real-world data.
Our NLP team uses these recent advances to help bring structure to and extract logic from text. These tools allow the team to develop knowledge graphs constructed from a rich dataset of entities and events based on news articles, social media and financial reports. Our models allow us to discover latent information underlying entities and their relations, and to surface the right information at the right time.
Reinforcement learning is one of the most promising sets of technologies for advancing artificial intelligence. Our team uses cutting edge methods in deep and asynchronous reinforcement learning to automate and amplify complex processes that do not lend themselves to traditional modelling as prediction or classification problems. In addition to applying state of the art methods, we look for opportunities to push the state of the art forward, developing new reinforcement learning techniques specifically designed to address the real world challenges that we face.