The LSTM meta-learner is a type of meta-learning. The LSTM meta-learner has two phases:
- Meta-learner: In this phase, the model focuses on learning general knowledge across various tasks.
- Base learner: In the base learner, the model tries to optimize to learn parameters for a task-specific objective.
The key idea of the LSTM meta-learner is to train an LSTM cell to learn an update rule for our original task. In meta-learning framework terms, an LSTM cell will be used as the meta-learner, whereas task-specific objectives, such as dog breed classification, will be the base learner.
Now, the question arises, why would we use an LSTM cell? The authors of the LSTM meta-learner made a key observation that a cell-state update in LSTM is similar to a gradient-based update in backpropagation, and can be used to learn the update rule of the base learner...
