Catalogue of Artificial Intelligence Techniques


Jump to: Top | Entry | References | Comments

View Maths as: Images | MathML

Holographic Associative Memory

Aliases: HAM

Keywords: associative, hologram, holographic, memory, neural network, pattern recognition

Categories: Vision

Author(s): Dario Villanueva

First brought up in 1990 by J. G. Shutherland, Holographic Associative Memory systems are heteroassociative representations of data sets used mainly for pattern recognition purposes. They are a kind of artificial neural network design, where information is represented by phase angle orientations on a complex plane. The reason why they are called Holographic is due to their non-conectionist approach, where stimulus-response patterns are packed into a single neural element. They are very efficient, since they enfold information in different phases onto a single plane, and their learning process is performed in just a single pass by a learning algorithm, as opposed to the multiple pass algorithms of traditional neural networks. Their learning process is based on mapping of stimulus-response associations to a generalized state space within a multidimensional complex domain. Generally, they learn the relationship between a stimulus member, Su in S and a corresponding response member Ru in R such that when given a query pattern Sq it will successfully retrieve a response pattern Rr ≈ Rt such that Rt in R and Sq is closest to St in S according to some criterion D(). There are three stages to this process. The first stage uses a learning algorithm which converts external stimulus {Su, Ru} into an internal representation using state space mappings to a complex plane. On the second phase, these complex numbers are physically stored. On the third phase, a decoding algorithm recollects the stored information according to the criterion D(). The advantages of such a system are manifold. Aside from the quick learning process, they are largely immune to input noise. Retrieval is satisfactory even when a very small piece of information is given as the query. HAMs are very useful for weather prediction systems, storage of record tables in relational databases and images and video frames in multi-media databases. A variety of HAMs, MHACs (multidimensional holographic associative memory) deal for example with representations of attention focus, which are a very problematic area in computer vision. Humans can change effortlessly the focus of attention from one object to another, but a traditional connectionist pattern recognition systems lack this ability. MHACs implement interactive attention and can retrieve information successfully with cues as small as 10% of the query frame.



Add Comment

No comments.