Research question
Can a network learn, during training, which examples it finds hard, and use that signal without adding real cost or losing accuracy?
In plain language
Not every training example is equally difficult. EDAT-Lite adds a small auxiliary head that learns to predict how hard each example is. That prediction makes the training process easier to inspect, so you can see which examples the model struggles with and how that changes over time.
Results
On MNIST, the framework preserved accuracy while making training dynamics more interpretable, with only a lightweight addition to the network.