The Impact of Hyperparameters on Adversarial Training Performance | HackerNoon
Briefly

The NEO-KD algorithm optimally balances knowledge distillation across multiple exits, enhancing adversarial training effectiveness in multi-exit networks, with particular focus on tuning hyperparameters for performance.
In the NEO-KD objective function, careful selection of the hyperparameters α and β is essential; extreme values can either hinder knowledge distillation or compromise adversarial robustness.
Read at Hackernoon
[
|
]