I know that default rule of thumb is 0.5 for sigmoid and tanh can be converted to sigmoid. But, I'm interested in knowing that what is the default rule of thumb for tanh threshold? As it ranges from -1,1, it is the midpoint, i.e. 0?


A good $\tanh$ activation function is

\begin{equation} v_j=1.7159 \left[ \frac{\exp(\frac{2}{3}u_j) - \exp(-\frac{2}{3}u_j)}{\exp(\frac{2}{3}u_j) + \exp(-\frac{2}{3}u_j)} \right], \end{equation}

where $u_j$ is the input to hidden node $j$ and $v_j$ is the output. $1.7159 \tanh(\frac{2}{3}x)$ was suggested by Lecun et al.

Y. LeCun, L. Bottou, G. Orr, K. Muller. Efficient BackProp. In: Orr, G. and Muller K. (Eds), Neural Networks: Tricks of the Trade. Lecture Notes in Computer Sci., 1524:9-50, 1998.

  • 1
    $\begingroup$ thanks, but this still does not answers the threshold question. The function you provide just rescales Tanh. $\endgroup$ – lvdp Feb 7 '17 at 4:53

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.