[Solved]Backpropagation Allows Weight Adjustment Neural Networks Many Layers Answer Following Ques Q37145578
Backpropagation allows weight adjustment for neural networks ofmany layers. Answer the following questions related to itspractical implementation (you might need to do some researchonline, but please use your own words). Can the activation functionbe linear? i.e. f(g(x)) = k*g(x)? What problem does it lead to? Ifthe activation is sigmoid (special case of logistic functions) suchas f(g(x)) = S(g(x)) = 1/(1+exp(-g(x))), it is clear that thegradient is very small if |g(x)| is large (the function saturatesin both directions). Due to the chain rule, the products of manysmall values will get exponentially smaller, and this ultimatelyprevents weights from being updated. This is known as the vanishinggradient problem. Is there a better activation function? Why? Hint:For the first question, make sure you study slides #24-26. Thenetwork should be one-layer with the sigmoid function (c=1) as theactivation function. Using an augmented input vector is equivalentto adding x3=-1 and w3=θ. The math for calculating the gradient islisted on slide #26. Notice for a single-layer net, oi is xi, asthere is no previous layer. So the part about inner layer is notused. One epoch is to process every pattern of the four provided(N=4). You could choose to use iterative (incremental) update, i.e.Δw is generated after each input.
Expert Answer
Answer to Backpropagation allows weight adjustment for neural networks of many layers. Answer the following questions related to i… . . .
OR
