Few Shot Learning

A paradigm of Supervised Learning.

1. One-Shot Learning

  • building a performant model given only a single instance
  • mostly used for facial recognition (identifying if another photo is of the same person or is different altogether)
  • One way to tackle this is proceed with Siamese Neural Networks
  • note that the "one shot" refers to the inference time usage of only one example to produce results whereas the training phase might still call for a larger datset.

2. Zero-Shot Learning

  • training a model to assign labels to objects (images usually) but the labels might not be present in the training data.
  • a trick is to represent the labels as well with embeddings so that a notion of familiarity with other already seen labels is captured.
    • moving on from a categorical label to a full blown vector that's later remapped to the necessary labels.
  • for instance : the label pairs (orange, banana) and (orange, red) will be closer in the vector space (pair entity 1 and 2 with each other) than something like (batman, barbie).
  • so assigning a new label can be reduced to forwarding the input down into an embedding and then looking up a label-embedding bidirectional dictionary for the closest match to report as a viable label.
  • for instance, if the model had access to zebras during training, the strip identification filters in a CNN might help classify a tiger seen for the first time.
Tags::ml:ai: