A nonlinear mapping produces outputs that cannot be represented as a linear transformation of its inputs. In a logistic model, the sigmoid transforms a weighted sum into a bounded, S-shaped probability between 0 and 1 while remaining differentiable for gradient-based learning. section 8.2.1
Nonlinearity changes what a model can represent, but a single logistic unit still has a linear decision boundary: inputs producing probability 0.5 lie on a hyperplane. Nonlinear feature transformations or hidden layers are needed to capture nonlinear relationships. section 8.2.1
In machine-learning workflows, nonlinear mappings support:
- Transforming features into model inputs and improving a linear model with nonlinear transformations. section 3.0
- Learning complex functions in graph neural networks by applying ReLU after graph-convolution layers; without nonlinearities, stacked linear layers collapse into one linear transformation. section 11.5.2
Sigmoid logistic model
Compute z = sum(x × w) + b, apply y = sigmoid(z), compare y with a target using binary cross-entropy, and backpropagate the loss to obtain gradients for the weights and bias.
Graph neural network branch
Apply ReLU after three graph-convolution layers, pool node representations into graph-level features, then pass them through fully connected layers to produce an embedding.










