Straight-through estimator

Summary

The straight-through estimator (STE, also called straight-through trick) uses both a hard/discrete value and a soft value in the forward pass, thereby allowing gradients to flow through its continuous input during backpropagation: $y=\mathrm{hard}(x)-\mathrm{stopgrad}(x)+x$, where $y=\mathrm{hard}(x)$ is the forward pass, and $\partial y/\partial x\approx 1$ in the backward pass. This allows rounding, argmax, quantization, etc., and is practically applied to do discrete sequence optimization in protein design by hallucination and codebook lookups in VQ-VAEs.

See also

• Inversion of protein folding neural networks
• Variational autoencoders
• Protein sequence and structure compression using VQ-VAE tokens is prone to codebook collapse
• Categorical Jacobian method
• Loss functions