LAYA — Layer-wise Attention Aggregator

Minimal reference implementation of LAYA, an interpretable output head that assigns input-conditioned attention weights to hidden layers. This example trains a simple MLP on Fashion-MNIST and visualizes global and class-wise attention profiles.

🚀 Usage

Click the badge above or open: LAYA.ipynb

The notebook:

1. trains LAYA on Fashion-MNIST,
2. evaluates accuracy and macro-F1,
3. extracts layer-wise attention weights,
4. plots global and class-wise attention patterns.

🧠 What is LAYA?

LAYA aggregates all hidden representations ( h_i ) using attention scores ( \alpha_i(x) ), producing:

• depth-aware predictions,
• intrinsic, per-sample interpretability without post-hoc methods.