Principal Result (AAM): This research presents an Artificial Associative Memory (AAM), which can learn pattern-associations and retrieve them with changeable attention. The attention refers to the fact that the user can specify any subset of the elements in the example query pattern and expect the memory to confine its match within the specified field of attention. Existing AAMs lack such flexibility. All current AAMs and Neural Networks perform match based on unchangeable unary attention over all the elements of the query pattern.
Corollary 1 (robustness): The proposed Artificial Associative Memory (AAM) can retrieve information from much less than 50% of the query frame. Conventional AAMs and Neural Networks require the effective query pattern to be at least 50% of the desired pattern.
Corollary 2 (dynamic attention): It can perform regular Associative Memory like statistically robust matching within a user specifiable field of attention. Unlike the existing AAMs, user can dynamically vary the field of attention during each query.
Corollary 3 (MNC feedback): It also has the unique ability to generate a feedback (called MNC) on the quality of match corresponding to the retrieved pattern. (Thus, makes the memory interactive).
Approach: In contrast to the conventional AAM, the proposed approach is based on, (i) a new holographic representation of information, which includes measurement as well as confidence as components of each element of information, (ii) mapping of measurements on the surface of a hypersphere, instead of on a the linear interval, and (iii) new rule of synaptic efficacy based on trigonometric averaging, rather than a statistical sum. The representations are inner transformations used by this memory corresponds to a multidimensional generalization of holographic operations. Therefore the computational model of this memory is also called Holographic Associative Computing.
Applications: Because of the dynamic attention ability
this memory can be used for searcing image or digital patterns from a small
samples of it. As an immediate application of this new capability an automatic
direct content-based search mechanism has been developed for querying into
image database (IDB). It improves the cost of direct content based IDB
search from O(np) to O(n.logp). It also use a fuzzy logic based query and
inference formalism. This mechanism can help in supplementing existing
content based query mechanisms which often suffer from representation insufficiency
and inaccuracy of intermediate model encoding. Following are some application
areas which can benefit from the result of this work: