Back to bookshelf

A Brain-Inspired Algorithm For Memory (Hopfield Networks) | Artem Kirsanov 

science•video•View Source
Appreciation
8
Importance
--
Date Added
1.5.26
TLDR
Easy visual introduction of Hopfield networks, a foundational model of associative memory.
2 Cents
This was a fun, intuitive watch (great introduction). I particularly like the analogy to protein folding that sets the scene for energy minimization.
Tags
  • Motivates the problem of finding associations amongst all possible candidates not as a search problem but one of energy minimization, in the same way protein folding isn’t searching across all possible configurations in 3D space.
  • Key idea is: 1) Sculpt an energy landscape where memories are at the local minima and 2) During inference, drive the system to change the states to minimize energy (roll the ball down the nearest well)
  • More on 1): During learning, imagine we have 1 pattern. We can solve for the lowest energy directly by choosing the weights (interactions) so the desired pattern is the ground state, guaranteeing a minimum there. When there is more than 1 pattern, we superpose the calculated interactions for each pattern