Hopfield networks

• Associative memories with

  • Binary units (output is +1 or –1)
  • Continuous updating (not discrete timesteps)
  • Symmetric weights (wij = wji)

• Define an energy function

  • H is “free energy”
  • H exists if wij = wji
  • The memorized patterns are local minima of an energy surface

• The update rule causes ?H to always be negative or zero

Previous slide

Next slide

Back to first slide

View graphic version