straight

Definition

Almost always part of "straight-through estimator": on the forward pass you apply a non-differentiable operation such as rounding or thresholding, and on the backward pass you pretend it was the identity and pass the gradient through unchanged. It is what makes quantization-aware training and discrete latent variables like those in VQ-VAE trainable at all. The gradient is knowingly wrong — a biased surrogate — yet it works well enough that it remains the default.