liveBook · Manning

Neural network

As explained in Grokking Bayes

A neural network is a machine learning model composed of interconnected parameterized units that transform inputs into predictions.

A neural network is a prediction function built from connected layers. Each layer applies a linear transformation followed by a nonlinear activation function, allowing many layers to represent complex mappings from inputs to outputs. section 10.1

Understand a neural network through its parts:

  • Nodes represent dimensions within a layer, while edges connect successive layers. section 10.1
  • Edge weights provide the coefficients of each linear transformation, and biases provide the additional offset in y = mx + b. section 10.1
  • Input and output dimensions follow the learning problem; intermediate hidden-layer sizes are design choices. section 10.1
  • A standard network produces a prediction from parameters, commonly written as f(x; w), but does not inherently quantify prediction uncertainty. section 10.1

Learning a noisy curve

A network can be trained on one-dimensional data generated from a sine function and corrupted by noise, then used to predict across a wider input range. Its architecture and learned weights determine the resulting curve, while a standard network’s predictions do not show uncertainty from observation noise or sparse data. section 10.1

Figure 10.2 The architecture of an example neural network with two hidden layers of size 6 and 4.

Inspect a two-hidden-layer architecture with hidden layers of sizes 6 and 4 to see how nodes, edges, weights, and biases form a network. section 10.1

Figure 10.3 Predictions made by a standard neural network, which don’t include uncertainty quantification despite its importance in indicating both observation noise and data sparsity.

Compare standard neural-network predictions with the uncertainty information needed to indicate observation noise and sparse data. section 10.1

How Grokking Bayes covers it

How deep it goes
Defines layers, activations, nodes, edges, weights, biases, and hidden-layer choices, but stops short of training procedures.
Examples
Uses a noisy sine-function data set, a two-hidden-layer architecture, and adversarial image noise to illustrate predictions and failure.
What you can do after
Evaluate a network’s architecture, parameters, predictions, and vulnerability to confident errors without uncertainty modeling.
How it connects
Connects deep learning with activation functions, linear transformations, uncertainty quantification, and Bayesian neural networks.

Also explained in

Compare how 11 books explain neural network →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.