L. A. Jaeckel. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28, NASA Ames Research Center, Moffett Field, USA, July 1989.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....of active registers into the summation register which in turn is matched against an external threshold. Results are stored into the output register. 3 Modifications to the original SDM are the selected coordinate design, the hyperplane design and several intermediate designs, as described in [4, 5]. In principle they differ in the methods of selecting sets of active registers. The differences between Kanerva s base model and those variants can be briefly described as follows: In the Kanerva SDM , the address matrix is a uniform randomly initialized, constant and binary matrix of dimension m ....

....such that all of them can be derived by an appropriate choice of parameters [16, 15] Addressing operations serve to activate a precisely predictable number of registers in a reproducible way. Close addresses should join a larger intersection of active registers than distant ones. Though in [4, 5] the measure has been based on the logical and, we will here use measures 5 based on the logical nand, the Zero distance d (n) nand (x; y) n Gamma1 X i=0 (x i fi y i ) with a fi b = 0 if a = b = 1 1 else or based on the logical xor, the Hamming distance, as used in the original ....

[Article contains additional citation context not shown here]

L. A. Jaeckel. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28, NASA Ames Research Center, Moffett Field, USA, July 1989.

.... Introduction 1 2 Different activation probabilities for writing and reading 1 3 Scaling up the memory 5 1 Real World Computing Partnership 2 Swedish Institute of Computer Science 1 Introduction We consider a Sparse Distributed Memory of Jaeckel s selected coordinates design, see Jaeckel [1]. Let us give a short explanation of the concept: An address is a binary string (vector) of 1s and 0s. A datum is a binary string (vector) of 1s and Gamma1s. Let N be the length (dimension) of the addresses and U the length (dimension) of the data. A hard location is a place in the memory where ....

....fi) 2 fi 2 2(1 Gamma fi)fi Delta ( Gamma1) 1 Gamma 2fi) 2 . 7) comes from the independence of B t;k and B s;j . We have E(L t;k;h ) Pr(L t;k;h = 1) Pr(h is activated by both Y and X t;k ) If t = 0, this equals [ 1 2 (1 Gamma ffl) 1 Gamma j) 1 2 fflj] L Delta [ 1 2 ] K GammaL = 1 2 ] K Delta r fl = pr fl , and we get (8) If t 1 it equals [ 1 2 Delta 1 2 ] L Delta [ 1 2 ] K GammaL = pr, and we get (9) For k 6= j we have E(L t;k;h L t;j;h ) Pr(h is activated by Y , X t;k and X t;j ) If t = 0, this equals [ 1 2 ....

[Article contains additional citation context not shown here]

L. A. Jaeckel. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28, Research Institute for Advanced Computer Science, NASA Ames Research Center, 1989.

.... u 2 3 An estimate of T 3 4 Karlsson s design 3 1 Real World Computing Partnership 2 Swedish Institute of Computer Science 1 Introduction We consider a Sparse Distributed Memory, either of Kanerva s original design, see Kanerva [2] or of Jaeckel s selected coordinates design, see Jaeckel [1]. Let us give a short explanation of the concept: An address is a binary string (vector) of 1s and 0s. A datum is a binary string (vector) of 1s and Gamma1s. Let N be the length (dimension) of the addresses, U the length (dimension) of the data. A hard location is a place in the memory where to ....

L. A. Jaeckel. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28, Research Institute for Advanced Computer Science, NASA Ames Research Center, 1989.

....locations) from the address can be considered as noise that will more or less cancel out. To implement a good threshold mechanism is somewhat involved but details can be found in [3] Several activation mechanisms have been used, e.g. Hamming distance or the Jaeckel Selected Coordinate Design [1, 2]. The activation mechanisms have all been based on matching the input address with some matching pattern for each hard location. This makes the implementations impractical without some massive parallel hardware support (except for small problems) This paper introduces an efficient mechanism for ....

Jaeckel L. A. An Alternative Design for a Sparse Distributed Memory. Technical report, Research Institute for Advanced Computer Science, NASA Ames Research Center, 1989.

....original design of Sparse Distributed Memory, Jaeckel s design, and a modification of Jaeckel s design invented by Roland Karlsson at SICS, having great implementation advantages for simulations on a sequential computer. For a general discussion of these different designs see Kanerva [1] Jaeckel [2], and Karlsson [3] The performance of the different designs will be studied assuming that 1. the hard locations in the memory, the data and the addresses for storing data are randomly chosen, 2. the reading procedure is the standard one (used in [1] and [2] better performance can be achieved ....

....designs see Kanerva [1] Jaeckel [2] and Karlsson [3] The performance of the different designs will be studied assuming that 1. the hard locations in the memory, the data and the addresses for storing data are randomly chosen, 2. the reading procedure is the standard one (used in [1] and [2]; better performance can be achieved with other reading procedures, being investigated in Sj din [4] 3. the reading address may be disturbed (i.e. it may contain noise) In particular, we will consider the question of determining the optimal probability p of activation. For his original design ....

[Article contains additional citation context not shown here]

Jaeckel, Louis A. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28. NASA Ames Research Center, 1989.

....key idea is that data written sufficiently far away (i.e. having few common activated hard locations) from the address can be considered as noise that will more or less cancel out. Several activation mechanisms have been used, e.g. Hamming distance or the Jaeckel Selected Coordinates Design [1, 2]. Two papers [6, 7] introduced and tested an efficient mechanism for finding the hard locations that are close to the input address. The proposed mechanism may be viewed as a fast implementation of a restricted Jaeckel design. The plain Kanerva SDM memory performs best if address and data are ....

Jaeckel L. A. An Alternative Design for a Sparse Distributed Memory. Technical report, Research Institute for Advanced Computer Science, NASA Ames Research Center, 1989.

....(x i Phi y i ) with a Phi b = ae 1 if a 6= b 0 if a = b for the convenience of easy conversion. Arguments to these functions are memory items located in the space f0; 1g n . A further modification has been applied: Since the combinatorial deduction tends to become rather complicated [Jaeckel, 1989a; 1989b] we preferred a statistical foundation. All probabilities are derived by application of the Binomial distribution or the Gaussian distribution as its approximation. The new generalized design requires various parameters: n is the dimension of the address space, k the number of selected ....

....by Kanerva [1988] ch. 8) His solution assembles separate SDMs into a folded SDM of a definite depth performing an accurate indexing on those sequences. A k folded SDM employs k inner SDMs, each one acting on individually delayed addresses. Sec. 5 gives an example. 2 Compare [Kanerva, 1992; Jaeckel, 1989a; Chou, 1988] for a deduction. 3 Algorithms and Data Types The CM 2 hardware architecture is almost optimally well suited to the requirements of SDM calculations. A large number of processors can be connected efficiently in the form of a virtual grid of arbitrary dimensions. Each processor is ....

Louis A. Jaeckel. An Alternative Design for a Sparse Distributed Memory. RIACS Technical Report 89.28, NASA Ames Research Center, Moffett Field, USA, July 1989.

....The memory is addressed within a binary address space of dimension N, where M 2 N . A picture of the general structure is given in figure 1. In 1.2 a list of commonly used notations is given. It should serve as a general reference when reading this paper. For relevant literature on SDM cf. [4], 5] 6] 7] 8] 10] When a U dimensional binary vector W is stored at an address X a mechanism activates a set of memory locations and W is added to the contents of these after first transforming 0:s to Gamma1:s. In the original SDM model every location was given a binary address and ....

L. A. Jaeckel. An alternative Design for a Sparse Distributed Memory. Technical Report TR-89.28, RIACS, 1989.

....Distributed Memory (SDM) is a memory with a set of M locations of U dimensional vectors of integers. The memory is addressed within a binary address space of dimension N, where M 2 N . A picture of the general structure is given in figure 1. For relevant literature on SDM cf. 3] 4] 5] [2], 7] 1] When a U dimensional binary vector W is stored at an address X an activation mechanism activates a set of the locations and W is added to the contents of these after first transforming 0:s to Gamma1:s. In the original SDM model every location was given a binary address and those ....

L. A. Jaeckel. An alternative Design for a Sparse Distributed Memory. Technical Report TR-89.28, RIACS, 1989.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC