Deep learning is fueled by large labeled datasets, but in domains like medicine, each label represents a human life at risk. Avesh Singh and Brandon Ballinger offer an overview of autoencoders, heuristic training, and few-shot learning, with an emphasis on practical tips to create high-performing models utilizing hundreds of thousands of unlabeled data points and only thousands of labeled points.
Unsupervised, semisupervised, and one-shot learning techniques let us supplement scarce labels with vast quantities of unlabeled data. Avesh and Brandon walk you through a large-scale (N=10,000) study with UCSF Cardiology that predicts atrial fibrillation, the most common heart arrhythmia and the number-one predictor of stroke, using heart rate measurements from an Apple Watch, utilizing labeled and unlabeled data to identify episodes with an AUC of 0.88.
Avesh Singh is an engineer at Cardiogram, a startup that applies deep learning to wearable data. Previously, Avesh worked at Nest Labs and Google. He holds a a BS and MS in computer science from Carnegie Mellon University.
Brandon Ballinger is a cofounder at Cardiogram. Previously, Brandon was a cofounder at Sift Science and an engineer at Google on speech recognition and ads quality. He was also one of the engineers called in by the White House to help fix Healthcare.gov. Brandon holds a BS in computer science from the University of Washington.
Comments on this page are now closed.
©2017, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • email@example.com