Department of Physics

University of New Orleans

New Orleans, Louisiana

Dr. Greene presented a talk on research involving an efficient method for retrieval of data from memory, which is sparse in feature space. Dr. Greene stated that the motivation for this research was to investigate memory-based reasoning, which he defined as the storage of experiences about data in memory that is used to make judgement about new situations. Key characteristic of associative memory include:

• Content addressable,
• Data stored/retrieved using a set of features,
• Auto-associative

Dr. Greene has developed a system called the Connectionist-Hashed Associative Memory (CHAM) to perform efficient retrieval of data from sparse associated memory. This system uses what he calls feature vectors for storage of the memory items, which consist of up to 1000 randomly, selected bivalent components. After giving detail examples of feature vectors, he listed some implementations, which include:

• Connectionist neural network approaches,
• The Hopfield model,
• Kanerva's sparse distributed memory.

Dr. Greene presented examples of feature vectors such as Boolean features and described in detail the use of hash tables. He has proposed to replace the use of hash tables with artificial neural networks in the future. He then discussed the input and output data and presented detailed information on the Fixed Network example.

After describing the current platform for which this system is operational, Dr. Greene described enhancements made to the original CHAM include increasing the number of network outputs, using hash tables, using multiple hashing networks and implementing an active search. He then presented experimental results on correlated feature vectors in a spelling correction application.

Dr. Greene concluded his presentation with a summary of the CHAM system. He stated that it is faster than a linear probabilistic best-match retrieval system, handles noisy data or partially specified data and can be easily expanded. Dr. Greene now needs to find some applications (with lots of data) for his system.