VersaTL: Versatile Transfer Learning for IMU-based Activity Recognition using Convolutional Neural Networks

The advent of Deep Learning has, together with massive gains in predictive accuracy, made it possible to reuse knowledge learnt from solving one problem in solving related problems. This is described as Transfer Learning, and has seen wide adoption especially in computer vision problems, where Convolutional Neural Networks have shown great flexibility and performance. On the other hand, transfer learning for sequences or timeseries data is typically made possible through the use of recurrent neural networks, which are difficult to train and prone to overfitting. In this work we present VersaTL, a novel approach to transfer learning for fixed and variable-length activity recognition timeseries data. We train a Convolutional Neural Network and use its convolutional filters as a feature extractor, then subsequently train a feedforward neural network as a classifier over the extracted features for other datasets. Our experiments on five different activity recognition datasets show the promise of this method, yielding results typically within 5% of trained-from-scratch networks while obtaining between a 24-52x reduction in the training time.