java - Neural network activation function -
i've had time building own ann library, i'm having difficulties understanding behaviors.
here activation functions coded:
doublefunction<double> sigmoid = x -> 1 / (1 + math.exp(-x)); doublefunction<double> sigmoidder = x -> { double s = sigmoid.apply(x); return s * (1 - s); }; doublefunction<double> tanh = math::tanh; doublefunction<double> tanhder = x -> 1. - math.pow(tanh.apply(x), 2); doublefunction<double> relu = x -> math.max(0, x); doublefunction<double> reluder = x -> { if (x < 0) return 0; return 1; }; doublefunction<double> softplus = x -> math.log(1 + math.exp(x));
in code i'm selecting pair serve activation function (+ derivative).
none of functions convege solution (i'm trying net learn xor operator) tanh function , derivative.
note: using bias neuron each layer outputs constant 1.
i've checked this out apparently i'm missing fundamental here.
does care enlighten me?
Comments
Post a Comment