Table 4 The effect of using SSL compared to supervised counterparts.

From: Self-supervised learning of wrist-worn daily living accelerometer data improves the automated detection of gait in older adults

Model's head

ElderNet

Supervised

FC (without non-linearity)

84.67 (0.44)

79.08 (0.24)

FC (with non-linearity)

84.74 (0.51)

79.21 (0.31)

U-Net

83.02 (0.86)

78.77 (0.73)

  1. Performance is reported as the F1 score (standard deviation between different seeds). In this table, we employed the same architecture (ResNet-V2 + additional model’s head), once using the ElderNet configuration and once by performing fine-tuning from scratch without utilizing the UK Biobank pre-trained model and the MAP data.
  2. FC fully-connected.
Back to article page