Georgia Institute of Technology College of Sciences

In this talk we focus on classification problems where noisy sensor measurements collected over a time window must be classified into one or more categories. For example, mobile phone health and insurance apps take as input time series from the accelerometer, gyroscope and GPS radio of the phone and output predictions as to whether the user is still, walking, running, biking, driving etc. Standard approaches to this problem consist of first engineering features from statistics of the data (or a transform) over a window and then training a discriminative classifier. For two applications we show how these features can instead be learned in an end-to-end modeling framework with the advantages of increased accuracy and decreased modeling and training time. The first application is reconstructing unobserved neural connections from Calcium fluorescence time series and we introduce a novel convolutional neural network architecture with an inverse covariance layer to solve the problem. The second application is driving detection on mobile phones with applications to car telematics and insurance.