full-grad

**Full-Grad** (Full-Gradient Representation) is an **attribution method that combines input gradients with bias gradients across all layers** — providing a complete, full-gradient saliency map that accounts for both the sensitivity and the bias terms throughout the entire network. **How Full-Grad Works** - **Input Gradient**: Standard gradient $partial f / partial x$ captures input sensitivity. - **Bias Gradients**: For each layer $l$, compute $partial f / partial b_l$ — the sensitivity to each layer's bias. - **Aggregation**: Full saliency = input gradient × input + sum of bias gradients mapped to input space. - **Completeness**: The full-gradient satisfies $f(x) = sum ( ext{input contributions}) + sum ( ext{bias contributions})$. **Why It Matters** - **Complete Attribution**: Unlike vanilla gradients or Grad-CAM, Full-Grad accounts for ALL sources of the prediction. - **Bias Terms**: Standard gradient methods ignore bias terms — Full-Grad includes their contribution. - **High Quality**: Produces cleaner, more faithful saliency maps that better highlight relevant input regions. **Full-Grad** is **the complete gradient picture** — combining input and bias gradients for fully faithful attribution across the entire network.