IPA IP 2023 Question 91

It calculates, inside one neuron, the value passed on to the next neuron — the activation function turns a neuron's weighted input into its output.

Each neuron receives values from the neurons of the previous layer, multiplies each by a weight, adds them up with a bias, and applies the activation function to that sum. The result is the neuron's output, which becomes an input of the neurons in the next layer. Common activation functions are the sigmoid function (an output between 0 and 1), ReLU (0 for a negative sum, the sum itself otherwise) and the step function. Because they are non-linear, stacking layers lets the network represent complex relationships.

Answer (d)

Why not others:

• •(a) — a confidence for the network's result is read from its output layer, for example a softmax output; it is not the job of the function applied in every neuron
• •(b) — the number of neurons in the intermediate (hidden) layer is a design choice, set by the designer or tuned by trial; no function calculates it from the input and output layers
• •(c) — likewise, the connections and size of the network are fixed when it is designed, not output by the activation function

Key rule: In every neuron, output = activation function(weighted sum of the inputs + bias); the activation function works on values in one neuron, not on the network's structure.

Basic Theory · AI-written, may contain errors