fastExp function

[gessfred]

I am wondering what is the motivation behind the fastExp function used in net_update_wa. I cannot see that it is faster or more numerically stable than the exp function from std in any meaningful way. And I am interested as to where the expression was found:

inline prec fastExp(prec a) {
    a = 1.0 + a / 1024.0;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    a *= a;
    return a;
}

[mgwoo]

I didn't write that part, but in my previous experience (about calculating exponential in deep-learning hardware for sigmoid), that approximate function was much faster than std::exp() function.

Reference: https://codingforspeed.com/using-faster-exponential-approximation/

[gessfred]

Fair enough, thank you for the information.