Part V: Adaptation
22. Neural Networks and Learning
Read the opening pages · printed page 383
Backpropagation (mlp.c)
A multilayer perceptron with forty hidden units learns to separate two classes of points by gradient descent on its error. The colored field is the network's opinion about every point on the screen; the bright ridge is where it is undecided. The data set changes every minute or so (moons, spirals, rings, XOR blobs, a checkerboard) and the weights are kept, so you watch the boundary bend from the old fit to the new one, and occasionally get stuck in a local minimum. Full screen →
Sections
- Pattern Classification and the Perceptron
- Linear Inseparability
- Multilayer Perceptrons
- Backpropagation
- Function Approximation
- Internal Representations
- Other Applications
- Unifying Themes
- Further Exploration
- Further Reading
