The Xavier 2 initializer draws from a uniform distribution [-limit, limit] where limit is squal to (6 / (fanIn + fanOut)) ** 0.25. This initializer is best suited for layers that feed into an activation layer that outputs values between -1 and 1 such as Hyperbolic Tangent and Softsign.
This initializer does not have any parameters.
use Rubix\ML\NeuralNet\Initializers\Xavier2; $initializer = new Xavier2();
- X. Glorot et al. (2010). Understanding the Difficulty of Training Deep Feedforward Neural Networks.