Early stopping controls neural-network training by monitoring performance and ending the process when improvement stops. A validation set can indicate when a classifier should stop, while a task-specific metric can guide generative models. section 2.4 section 3.3.3
A practical early-stopping procedure:
- Choose a monitored measure, such as validation performance or generated-sample MSE. section 2.4 section 3.3.3
- Inspect outputs during training when numerical loss alone does not show whether the model is producing useful results. section 3.3.3 section 3.5.3
Stopping a GAN that generates multiples of five
Define distance as the MSE between generated numbers' remainders modulo 5 and zero. Instantiate EarlyStop(800), evaluate distance after each generator-training iteration, and break when the generator has not improved for 800 epochs. Printing every 50 epochs shows the outputs progressing toward multiples of five. section 3.5.3
