EluBackward

Autograd node for the Exponential Linear Unit $y = x$ if $x \ge 0$ else $\alpha\,(e^x - 1)$ (Clevert et al. 2015).

Like Leaky ReLU, the scalar hyper-parameter $\alpha$ is captured on the node via a custom forward() so the backward kernel sees a concrete value rather than chasing it through call sites. The piecewise backward itself is delegated to the backend because the conditional branch is not expressible through the generic storage primitives.

Math