Associative Memory for Neural Turing Machines

Neural Turing Machines (NTM) where first described around the same time GANs were invented.  NTM are deep neural networks with an external memory bank attached to provide working memory.

Unfortunately progress seems to have stalled, likely due to lack of insight into how the memory bank should designed.  For example the use of crude spreading mechanisms to provide the soft/fuzzy addressing necessary.   A much better alternative could be this type of locality sensitive hash (LSH) associative memory:

This type of associative memory can be made vast by further using the LSH algorithm to switch in different vector blocks of memory during operation (various options), giving a sort of hybrid between random access memory and soft associative memory.  With fast SSD memory you could potentially have Tbyte sized associative memory that still was sufficiently fast and had sufficient soft/fuzzy/generalization behavior.

More questions in this forum