[ Wikipedia ]
Fun question: (Scalable) Hopfield networks for the protein-folding problem?
At today’s meeting, I got to see an energy landscape that looked like something out of BIOL 4105 (modeling macromolecules.) It was a simplified (1-D, continuous) landscape showing different states of a hopfield network. I asked if this network was using Monte Carlo, and Dr Anderson said that it was similar, but not the same. I didn’t quite follow, but I think he said that in MC, we only search for minima, but in this network, we were looking for stable states + changing around the weights/nodes(?)
Anyway, the coolest part was getting to kn0w that hopfield nets can model associative memory. That was awesome. I wonder if there’s a physical reason behind this. Maybe I am hoping to see some big molecular structures in the brain (synapses!) that serve as memory (besides DNA?), in the sense that the hopfield weights correspond to physical bond lengths, and the values at the nodes are ATP/ADP/AMP-like molecules (just groups that can conform or chemically change their state), leading to a variety of memories. I don’t know if such a physical basis exists (maybe its already in the books on my reading list!), but I definitely won’t hope for one (that’s a Feynman video! Look for the onion example after 2:14, and beware, “Feynman’s atheism shines through towards the end”, as it was brought up in the Carl Sagan lectures in Howey!)
Audio Sketches 