OFFSET
1,3
COMMENTS
SELU(x) = scale*(max(0,x)+min(0,alpha*(exp(x)-1))).
SELU is an activation function derived from ELU (Exponential Linear Unit), which uses the identity function for x >= 0 and exp(x)-1 for x <= 0.
It multiplies the negative-region output of ELU by the constant alpha and additionally scales the entire function by the constant lambda so that the outputs of a neural network converge toward zero mean and unit variance.
LINKS
Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter, Self-Normalizing Neural Networks, arXiv:1706.02515 [cs.LG], 2017.
FORMULA
Equals sqrt(2/(1 + (sqrt(2/Pi)/(1 - exp(1/2)*erfc(1/sqrt(2))))^2*(1 - 2*exp(1/2)*erfc(1/sqrt(2)) + exp(2)*erfc(sqrt(2))))).
EXAMPLE
1.0507009873554804934193349852946...
MATHEMATICA
RealDigits[Sqrt[2/(1 + (Sqrt[2/Pi] / (1 - Exp[1/2]*Erfc[1/Sqrt[2]]))^2 * (1 - 2*Exp[1/2] * Erfc[1/Sqrt[2]] + Exp[2]*Erfc[Sqrt[2]]))], 10, 105][[1]] (* Amiram Eldar, May 24 2026 *)
PROG
(PARI) lambda = sqrt(2 / (1 + (sqrt(2/Pi) / (1 - exp(1/2)*erfc(1/sqrt(2.))))^2 * (1 - 2*exp(1/2)*erfc(1/sqrt(2.)) + exp(2)*erfc(sqrt(2.)))))
CROSSREFS
KEYWORD
nonn,cons
AUTHOR
Donghwi Park, May 24 2026
STATUS
approved
