Deep convolutional feature transfer across mobile activity recognition domains, sensor modalities and locations

Deep convolutional neural networks are powerful image and signal classifiers. One hypothesis is that kernels in the convolutional layers act as feature extractors, progressively highlighting more domain-specific features in upper layers of the network. Thus lower-level features might be suitable for transfer. We analyse this in wearable activity recognition by reusing kernels learned on a source domain on another target domain. We consider transfer between users, application domains, sensor modalities and sensor locations. We characterise the trade-offs of transferring various convolutional layers along model size, learning speed, recognition performance and training data. Through a novel kernel visualisation technique and comparative evaluations we identify what learned kernels are predominantly sensitive to, amongst sensor characteristics, motion dynamics and on-body placement. We demonstrate a ~17% decrease in training time at equal performance thanks to kernel transfer and we derive recommendations on when transfer is most suitable.