Explicit State Vector Representation for Heteroassociative Memories
NAVAL RESEARCH LAB WASHINGTON DC
Pagination or Media Count:
In summary, it has been shown that the commutative correlation operator used in most linear associative memory AM models enforces symmetries that preclude the learning of several classes of important mapping functions. It has also been shown, however, that these symmetries can be eliminated simply by using a different information representation scheme. The explicit state ES representation has been proposed as an alternative to the commonly used bipolar form and has been demonstrated to permit standard AM architectures to learn nonlinear mappings. In particular, it has been shown that the ES representation permits first-order AM models to learn nonlinear transformations of the form y xT b. This characterization is important because its properties are directly amenable to analysis by using the known properties of affine transformation. For example, it has been noted that this transformation renders the ES formulation immune to the effects of constant additive noise. It has also been shown that the ES representation can be easily generalized for the use of higher order correlation information.
- Statistics and Probability
- Operations Research