MseLossBackward

Autograd node for mean-squared-error loss.

Computes

with Mean reduction; for Sum the leading $1/N$ factor is dropped, and for None the full per-element squared error $(\hat{y}_i - y_i)^2$ is returned without reduction. Both input and target are saved so that the backward pass can re-emit $\nabla_{\hat y} = \frac{2}{N} (\hat y - y) \cdot g$ (or $2 (\hat y - y) g$ for Sum, or $2 (\hat y - y) \odot g$ for None).

Math