Early Stopping Mechanism for qiskit ML Algorithms

[Ashad001]

What should we add?

Implement an early stopping callback for qiskit-machine-learning algorithms, allowing us to halt training when the objective function shows no improvement for a set number of iterations.

Early stopping is a common technique i.e. keras to save computation and prevent overfitting. Adding support for this across qiskit algorithms (like VQC, QSVC) would enhance the training process without needing custom implementations.

[edoaltamura]

This can already be achieved by setting max_iter to a desired number of iterations in the optimizers. A dedicated class could be created to translate the early stopping strategy for the optimizers in the style of Keras, at the cost of introducing an extra overhead.

In my experience, the most popular early stopping mechanisms can be implemented simply by tweaking the optimizer options as indicated above (see the scipy.minimize documentation for reference). Anything more complex than this is likely to be bespoke for specific projects and may not benefit enough users yet for the core team to commit to developing it at this time.
Of course, we'd welcome external contributions that aim at implementing such a feature.