18. Natural and Analog Computation

Read the opening pages · printed page 307

Associative Memory (assoc.c (and hopfield.c))

A network of neurons where every neuron is wired to every other, trained by Hebb's rule: neurons that fire together strengthen their connection. Show it a handful of patterns and each becomes a valley in an energy landscape. Start the network from a badly corrupted pattern and let each neuron update in turn; it rolls downhill into the nearest memory. The page dissolves each recovered pattern into noise and lets the next memory crystallize out of it. Full screen →

Sections

1. Artificial Neural Networks
2. Associative Memory and Hebbian Learning
3. Recalling Letters
4. Hopfield Networks and Cost Optimization
5. Unifying Themes
6. Further Exploration
7. Further Reading

Figures

Figure 18.1Figure 18.2Figure 18.3Figure 18.4Figure 18.5Figure 18.6Figure 18.7Figure 18.8