Advanced Deep Learning with Python

Layers as operations

The next level in the NN organizational structure is the layers of units, where we combine the scalar outputs of multiple units in a single output vector. The units in a layer are not connected to each other. This organizational structure makes sense for the following reasons:

• We can generalize multivariate regression to a layer, as opposed to only linear or logistic regression for a single unit. In other words, we can approximate multiple values with a layer as opposed to a single value with a unit. This happens in the case of classification output, where each output unit represents the probability the input belongs to a certain class.
• A unit can convey limited information because its output is a scalar. By combining ...