Back in the 1990s, network topologies like Bidirectional Associative Memory and Hopfield Nets were all the rage. Largely, these were billed as having energy states and recurrents that would help them learn/store and recall specific patterns readily.
These topologies are perhaps less pertinent now, but I'm wondering if they are still an optimal way to approach data mining? They seem like they would be pertinent for things like contextual analysis in NLP, with a network trained to observe what it has "seen" before. Will these topologies come back in, or have they been replaced by options with better accuracy rates?