Sigmoid Linear Unit

A smooth activation function that scales each value by its own sigmoid score.

The sigmoid linear unit (SiLU) smoothly scales each value by its own sigmoid score, making it a softer alternative to ReLU and an important stepping stone to gated designs such as SwiGLU.

At a glance

Released

February 2017

Authors

Stefan Elfwing, Eiji Uchibe, Kenji Doya

Optimizes

  • Activation Smoothness

What It Is

The sigmoid linear unit (SiLU) is an activation function that multiplies each value x by sigmoid(x). Large positive values pass through strongly, values near zero are softened, and negative values are reduced smoothly instead of being cut off sharply. In transformer feed-forward layers, that gives the model a softer filter than ReLU.

Why It Exists

SiLU optimizes for a smoother activation rule. Instead of a hard cutoff, it lets the negative side fade gradually, which often fits modern wide feed-forward layers better than a sharp zero clamp.

How It Works

SiLU works one value at a time by multiplying that value by its own sigmoid score. Large positive values mostly survive, small values are softened, and negative values shrink rather than snapping to zero. In a transformer feed-forward layer, the model applies that rule after expanding into a wider hidden state and before projecting back down.
Activation Curves
f(x)
x
  • ReLU
  • SiLU

Math Or Compute Schema

The formula below shows the smooth self-gating rule directly. The sigmoid term is what turns one value into its own soft gate.
Sigmoid linear unit activation
SiLU(x)=x σ(x)\mathrm{SiLU}(x) = x\,\sigma(x)
xx
One input value before the activation rule is applied.
σ\sigma
Sigmoid function applied to that same input value.

Compared To Nearby Modules

Compared with ReLU and LeakyReLU, SiLU is the smooth option in this activation family. It still stays inside one dense feed-forward path, but it prepares the reader for the stronger two-branch gate used by SwiGLU.
Comparison dimensionSiLUReLULeakyReLU
Negative branchShrink negative values smoothly with x sigma(x)Clamp every negative hidden value to 0Keep a small slope alpha x for x < 0
Transition shapeSmooth curve through 0Hard corner at 0Hard corner with a weak negative path
Main tradeoffSmoother feature filtering, but more complex than the ReLU family baselineSimple and sparse, but fully drops negative signalKeeps some negative gradient, but still uses a piecewise-linear cutoff

Example Architectures

Many recent decoder-only language models and gated feed-forward blocks use SiLU or a close relative, especially when authors want a smoother dense baseline before moving to SwiGLU.

Limitations And Tradeoffs

SiLU is smoother than ReLU-family activations, but it is still a dense per-token activation, not a routing or expert mechanism. Every token still runs through the full shared feed-forward path.

Why It Still Matters

SiLU sits at an important bridge point between classic activation pages and modern gated feed-forward pages. Readers who understand SiLU can usually decode SwiGLU much faster.

Tags

References

  1. Elfwing, Stefan, Eiji Uchibe, and Kenji Doya. "Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning." arXiv, 2017, https://arxiv.org/abs/1702.03118.
  2. Shazeer, Noam. "GLU Variants Improve Transformer." arXiv, 2020, https://arxiv.org/abs/2002.05202.