** Next:** Perceptrons
** Up:** 59.771 Research Topics in
** Previous:** Issues in Decision Tree

Read Chapter 4 of Mitchell.

``ANNs are particularly suited to learning in problems where the training data corresponds to noisy, complex sensor data, such as inputs from cameras and microphones.'' (p.83)

Think about using ANNs especially when

- 1.
- Target function is defined over input instances that can be represented as a vector of predefined features.
- 2.
- Target function is a vector of real values (Note that discrete valued target functions are just a special case of this because we can always partition continuous space into discrete intervals).
- 3.
- Training examples may contain noise/errors.
- 4.
- Long training times are acceptable.
- 5.
- Fast evaluation of target function is required.
- 6.
- Human understandability of the learned target function is of low priority, i.e. we don't care how the problem is solved as long as it is solved.

- Perceptrons
- The Perceptron Training Rule
- Gradient Descent/Ascent
- Gradient Descent in ANNs
- Stochastic approximation to gradient descent
- Multilayer Nets, Sigmoid Units
- The Backpropagation Algorithm
- Issues in ANN Learning and Related Topics