Publication · Published

EDAT-Lite: Predictive Difficulty-Aware Training for Neural Networks

Compared
  1. Backbone
  2. Auxiliary head
  3. Predict difficulty
  4. Adapt training

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.