An empirical investigation of sparse distributed memory using discrete speech recognition
Douglas G. Danforth · NASA Technical Reports Server (NASA) · 1990
An experimental investigation of Sparse Disn-ibuted Memory (SDM) (Kanerva, 1988) is presented. SDM is an associative memory which can be thought of as a 3-layer Artificial Neural Network. It uses massive parallelism, associates very large patterns, and is trained rapidly. The theory of SDM was developed for uncorrelated bit _. In this paper the behavior of SDM is examined when the constraint of random input is violated and the memory is presented with highly-correlated dam for classification tasks.Experiments from the domain of discrete-word speech recognition are used. These experiments lead, in a step-by-step manner, to factors which improve the memory's ability to recall and to generalize. It is shown that generaliTAtion can be enhanced with appropriate appLication of: (1) the form of encoding of class labels, (2) the placement of hard locations within the memory, (3) the activation rule of hard locations, and (4) the write rule used to modify the memory. Comparisons are made between SDM, a class-mean model, and the Nearest Neighbor rule.]:or single-talker digit recognition a form of SDM, called the Selected Coordinate Design, attains 99.3 % correct generalization.