Skip to main content
. 2021 Jun 30;21(13):4496. doi: 10.3390/s21134496
Algorithm 1: Complete procedure outline.
Input:
  • -

    (optional) Labelled Dataset (LD) from similar domain as end application;

  • -

    Feature Extractor (FE);

  • -

    Edge device;

Pre-deployment:
  • -

    Place a standard classifier (C) satisfying the edge device constraints on top of the FE;

  • -

    (optional) Further train the FE on LD;

After deployment
  • -
    While (condition to update) not met:
    • -
      Collect data point s
    • -
      Extract features hs=FE(s)
    • -
      (optional) Apply dimensionality reduction hsdr=DR(hs)
    • -
      Use classifier C to get the prediction ys=C(hs), or ys=C(hsdr)
  • -

    Fine-tune C on the new data points collected