Efficient retrieval from sparse associative memory

摘要

Best-match retrieval of data from memory which is sparse in feature space is a time-consuming process for sequential machines. Previous work on this problem has shown that a connectionist network used as a hashing function can allow faster-than-linear probabilistic retrieval from such memory when presented with probing feature vectors which are noisy or partially specified. This paper introduces two simple modifications to the basic Connectionist-Hashed Associative Memory which together can improve the retrieval efficiency by an order of magnitude or more. Theoretical results are presented for storage/retrieval of memory items represented by feature vectors made up of 1000 randomly selected bivalent components. Experimental results on correlated feature vectors are presented in the context of a spelling correction application.