deeplift

**DeepLIFT** (Deep Learning Important FeaTures) is an **attribution method that explains predictions by comparing neuron activations to their reference activations** — decomposing the difference between the output and a reference output into contributions from each input feature. **How DeepLIFT Works** - **Reference**: A reference input $x_0$ (analogous to Integrated Gradients' baseline) with known activations. - **Difference**: For each neuron, compute the difference from reference: $Delta y = y - y_0$. - **Contribution Rule**: Assign contributions $C(Delta x_i)$ to each input such that $sum_i C(Delta x_i) = Delta y$. - **Rules**: Rescale rule (proportional to activation difference) or RevealCancel rule (separates positive and negative contributions). **Why It Matters** - **Summation Property**: Contributions from all features sum exactly to the prediction difference — complete attribution. - **Beyond Gradients**: DeepLIFT handles saturated activations better than raw gradients (which are zero at saturation). - **Efficiency**: Requires only one forward + one backward pass (no iterative interpolation like Integrated Gradients). **DeepLIFT** is **attribution by comparison** — explaining how much each feature contributes to the prediction relative to a reference baseline.