Federated Self-Supervised Learning of Multisensor Representations for Embedded Intelligence

Smartphones, wearables, and Internet of Things (IoT) devices produce a wealth data that cannot be accumulated in centralized repository for learning supervised models due to privacy, bandwidth limitations, the prohibitive cost annotations. Federated provides compelling framework from decentralized data, but conventionally, it assumes availability labeled samples, whereas on-device are generally either unlabeled or annotated readily through user interaction. To address these issues, we propose self-supervised approach termed \textit{scalogram-signal correspondence learning} based on wavelet transform learn useful representations sensor inputs, such as electroencephalography, blood volume pulse, accelerometer, WiFi channel state information. Our auxiliary task requires deep temporal neural network determine if given pair signal its complementary viewpoint (i.e., scalogram generated with transform) align each other not optimizing contrastive objective. We extensively assess quality learned features our multi-view strategy diverse public datasets, achieving strong performance all domains. demonstrate effectiveness an input collection downstream tasks training linear classifier over pretrained network, usefulness low-data regime, transfer learning, cross-validation. methodology achieves competitive fully-supervised networks, outperforms pre-training autoencoders both central federated contexts. Notably, improves generalization semi-supervised setting reduces required leveraging learning.