Examining the Adversarial Test Accuracy of Later Exits in NEO-KD Networks | HackerNoon
Briefly

from Hackernoon1 year ago
The performance degradation observed at later exits in NEO-KD is primarily due to adversarial examples applying greater cumulative losses, thus reducing accuracy in these later stages.
Hackernoonhttps://hackernoon.com/examining-the-adversarial-test-accuracy-of-later-exits-in-neo-kd-networks
Implementing an ensemble strategy can effectively mitigate performance issues observed at later exits, especially when budgeted prediction setups are considered, enhancing overall model effectiveness.
Hackernoonhttps://hackernoon.com/examining-the-adversarial-test-accuracy-of-later-exits-in-neo-kd-networks
Read at Hackernoon
#neo-kd #performance-degradation #adversarial-examples #ensemble-strategy #machine-learning
[
|
]