A weighted voting model of associative memory: theoretical analysis

In this paper we investigate a RAM-based associative memory that uses a weighted voting scheme. We adopt the testing protocols commonly used in the area of face recognition, and propose that the capacity of the system be measured by the results of an identification test (ability to properly recognize known information) and a watch-list test (ability to properly reject inputs that should not be matched with any of the memory set patterns). For the case of binary and random memory sets, we are able to derive theoretical expressions characterizing the performance of the weighted voting memory on both of these tests.