To train a one-shot learning model, you need to match training conditions to test time conditions—for example, training on less data but across several batches, just as we did in matching networks. LSTM meta-learner also follows the same concept as matching networks and has been proven to perform really well at task-specific objectives.
To begin understanding the LSTM meta-learner, first, we need to understand certain terms:
- Base learner (M): Main task-specific objective, with parameters,
—for example, a classifier to detect cats - Meta-learner (R): LSTM cell, with parameters,
- Data points (X, Y): Data ...
