There are some papers indicating the use of analog error correction codes (ECC) by grid cells in the rodent brain. They didn’t show (or explain very well) the algorithm they use for simulation in the paper. When I worked out how to do that on my own it turned out very like a digital error correction code. There are only a few papers on analog error correction, there are billions on digital error correction.
What advantages could applying error correction codes to AI bring?
For one thing the digital ECC are pretty much all or nothing. As you know from you digital TV receiver. Something is either recognized or it is not. You either get an exact pattern or you get noise. This is a cool switching effect, you get clear crispened data or noise=unknow.
A pattern recognizer that has misclassified something can be nudged toward producing either a correct pattern or noise=unknown.
You probably can do something equivalent to deep neural nets by having pattern detectors with ECC at each layer. The training time would be very low.
Sounds like it might help with Hopfield nets too. You learn the code, not the data as a target. At the next time step you decode the output and feed it back as an input to the net. I guess the first time around you would get a quilt patchwork of recognized and unrecognized regions. Second time around more patches would be filled in until the recall completes. It might solve some of the problems with that kind of net such as the need for asynchronous update. There are so many things you might try, but who can try them all? I certainly can’t and I only deal in the most absolute minimal basic things.