Juniper learned that designing a neural network architecture for a supervised classification problem wasn’t hard.
Although most of it needed experimentation, one thing Juniper could count on was the design of the output layer of the network.
Which of the following is a correct statement about the neurons in the output layer of a classification network?
The number of neurons in the output layer should always match the number of classes.
The number of neurons in the output layer doesn’t necessarily need to match the number of classes.
The number of neurons in the output layer should always be greater than one.
The number of neurons in the output layer should always be a multiple of 2.
2
When working on a multi-class classification problem, setting the output layer to the same number of classes we are interested in predicting is common. But what happens when we are working on a binary classification problem?
A network to solve a binary classification problem doesn’t need an output with two neurons. Instead, we can use a single neuron to determine the class by deciding on a cutoff threshold. For example, the result could be positive if the output exceeds 0.5 and negative otherwise.
That means we can solve a problem requiring two classes with a single neuron.
We can stretch the same idea to multi-class classification problems: We could interpret the output layer as a binary result, allowing us to represent multiple classes with fewer neurons. For example, we would need only two neurons to classify instances into four different categories (00, 01, 10, 11.) This approach, although not popular, it’s possible.
Therefore, the second choice is the correct answer to this question.
Recommended reading