Activity Recognition Using Deep Convolutional Neural Networks
WHOOP is a performance optimization system that combines a wrist-worn hardware device with an innovative cloud-based analytics system. Together, the system offers athletes the ability to harness physiological data to inform sleep, training, and recovery decision making. WHOOP is able to gather more data and provide a better user experience by automating the collection and classification of workout data, a task which is achieved with the use of convolutional neural networks.
An earlier version of the WHOOP product provided an interface for the manual collection and classification of athletic activities. In exchange for providing this data, athletes received actionable feedback about how their bodies responded to the endured load. Using tens of thousands of these user-reported and user-labeled activities, a series of algorithms were trained to pick out a workout from a stream of WHOOP-generated biometric data; then, using a deep convolutional neural network, correctly classify the sport. This network is capable of correctly associating relevant movement patterns with a wide range of potential activities, allowing the model to distinguish even very similar types of activities, where there are a considerable number of shared traits.
Brian Todd is a data scientist for WHOOP, a wearable tech startup in Boston aimed towards optimizing performance in elite athletes. He works primarily with applications of machine learning and computational statistics on real time biometric sensor data received from our wrist worn device. Graduated from the University of Michigan in 2013, with a degree in Mathematics.