Vector Symbolic Architectures: A New Building Material For Artificial General Intelligence

Vector Symbolic Architectures: A New
Building Material for Artificial General

Intelligence1
Simon D. LEVY a,2 , and Ross GAYLER b
a
Washington and Lee University, USA
b
Veda Advantage Solutions, Australia
Abstract. We provide an overview of Vector Symbolic Architectures (VSA), a class

of structured associative memory models that offers a number of desirable fea-
tures for artificial general intelligence. By directly encoding structure using famil-
iar, computationally efficient algorithms, VSA bypasses many of the problems that
have consumed unnecessary effort and attention in previous connectionist work.
Example applications from opposite ends of the AI spectrum – visual map-seeking
circuits and structured analogy processing – attest to the generality and power of
the VSA approach in building new solutions for AI.
Keywords. Vector Symbolic Architectures, associative memory, distributed representations,
Holographic Reduced Representation, Binary Spatter Codes, connectionism
Introduction
Perhaps more so than any other sub-field of computer science, artificial intelligence has
relied on the use of specialized data structures and algorithms to solve the broad variety
of problems that fall under the heading of “intelligence”. Although initial enthusiasm
about general problem-solving algorithms [1] was eventually supplanted by a “Society
of Mind” view of specialized agents acting in concert to accomplish different goals [2],
the dominant paradigm has always been one of discrete atomic symbols manipulated by
explicit rules.
The strongest challenge to this view came in the 1980’s, with the emergence of con-
nectionism [3], popularly (and somewhat misleadingly) referred to as neural networks. In
contrast to the rigid, pre-specified solutions offered by “Good Old-Fashioned AI” (GO-
FAI), connectionism offered novel learning algorithms as solutions to a broad variety of
problems. These solutions used mathematical tools like learning internal feature repre-
sentations through supervised error correction, or back-propagation [4], self-organization
of features without supervision [5], and the construction of content-addressable mem-
ories through energy minimization [6]. In its most radical form, the connectionist ap-
1 The authors thank Chris Eliasmith, Pentti Kanerva, and Tony Plate for useful discussion on the topics
presented here, and three anonymous reviewers for helpful suggestions.

2 Corresponding Author: Simon D. Levy, Computer Science Department, Washington and Lee University,
Lexington, Virginia 24450 USA; E-mail: levys@wlu.edu.

proach suggested that the structure assumed by GOFAI and traditional cognitive science
might be dispensed with entirely, to be supplanted by general mechanisms like sequence
learning [7]. In support of the representations derived by such models, proponents cited
the advantages of distributing informational content across a large number of simple pro-
cessing elements [8]: distributed representations are robust to noise, provide realistically
“soft” limits on the number of items that can be represented at a given time, and support
distance metrics. These properties enable fast associative memory and efficient compari-
son of entire structures without breaking down the structures into their component parts.
The most serious criticism of connectionism held that neural networks could not
arrive at or exploit systematic, compositional representations of the sort used in tradi-
tional cognitive science and AI [9]. Criticism that neural networks are in principle un-
able to meet this requirement was met in part by compositional models like RAAM [10];
however, RAAM’s reliance on the back-propagation algorithm left it open to criticism
from those who pointed out that this algorithm did not guarantee a solution, and could
be computationally intractable [11].3
In the remainder of this paper, we describe a new class of connectionist distributed
representations, called Vector Symbolic Architectures (VSA), that addresses all of these
concerns. VSA representations offer all of the desirable features of distributed (vector)
representations (fast associative lookup, robustness to noise) while supporting systematic
compositionality and rule-like behavior, and they do not rely on an inefficient or biolog-
ically implausible algorithm like back-propagation. The combination of these features
makes VSA useful as a general-purpose tool or building material in a wide variety of
AI domains, from vision to language. We conclude with a brief description of two such
applications, and some prospects for future work.
1. Binding and Bundling in Tensor Products
Systematicity and compositionality can be thought of as the outcome of two essen-

tial operations: binding and bundling. Binding associates fillers (John, Mary) with roles
(LOVER, BELOVED). Bundling combines role/filler bindings to produce larger structures.
Crucially, representations produced by binding and bundling must support an operation
to recover the fillers of roles: it must be possible to ask “Who did what to whom?” ques-
tions and get the right answer.
Vector Symbolic Architectures is a term coined by one of us [13] for a general
class of distributed representation models that implement binding and bundling directly.
These models can trace their origin to the Tensor Product model of Smolensky [14].
Tensor-product models represent both fillers and roles as vectors of binary or real-valued
numbers. Binding is implemented by taking the tensor (outer) product of a role vector
and a filler vector, resulting in a mathematical object (matrix) having one more dimension
than the filler. Given vectors of sufficient length, each tensor product will be unique.
Bundling can then be implemented as element-wise addition (Figure 1), and bundled
structures can be used as roles, opening the door to recursion. To recover a filler (role)
from a bundled tensor product representation, the product is simply divided by the role
(filler) vector.
3 Recent results with a linear version of RAAM using principal component analysis instead of back-prop
[12] show promise for overcoming this problem.

Figure 1. Tensor product representation of John loves Mary.
Figure 2. Methods for keeping fixed dimensionality in tensor-product representations.
2. Holographic Reduced Representations and Binary Spatter Codes
Because the dimension of the tensor product increases with each binding operation,
the size of the representation grows exponentially as more recursive embedding is per-
formed. The solution is to collapse the N × N role/filler matrix back into a length-N
vector. As shown in Figure 2, there are two ways of doing this. In Binary Spatter Cod-
ing, or BSC [15], only the elements along the main diagonal are kept, and the rest are
discarded. If bit vectors are used, this operation is the same as taking the exclusive or
(XOR) of the two vectors. In Holographic Reduced Representations, or HRR [16], the
sum of each diagonal is taken, with wraparound (circular convolution) keeping the length
of all diagonals equal. Both approaches use very large (N > 1000 elements) vectors of
random values drawn from a fixed set or interval.
Despite the size of the vectors, VSA approaches are computationally efficient, re-
quiring no costly backpropagation or other iterative algorithm, and can be done in paral-
lel. Even in a serial implementation, the BSC approach is O(N ) for a vector of length N ,
and the HRR approach can be implemented using the Fast Fourier Transform, which is
O(N log N ). The price paid is that most of the crucial operations (circular convolution,
vector addition) are a form of lossy compression that introduces noise into the represen-
tations. The introduction of noise requires that the unbinding process employ a “cleanup
memory” to restore the fillers to their original form. The cleanup memory can be imple-
mented using Hebbian auto-association, like a Hopfield Network [6] or Brain-State-in-a-
Box model [17]. In such models the original fillers are attractor basins in the network’s
dynamical state space. These methods can be simulated by using a table that stores the
original vectors and returns the one closest to the noisy version.
3. Applications
As a relatively new technology, VSA is just beginning to be used in the AI and cognitive
science communities. Its support for compositional structure, associative memory, and
efficient learning make it an appealing “raw material” for a number of applications. In
this concluding section we review some of these applications, and outline possibilities
for future work.
3.1. Representing Word Order in a Holographic Lexicon
Jones and Mewhort [18] report using a holographic / convolution approach similar to
HRR, for incorporating both word meaning and sequence information into a model lex-
icon. Their holographic BEAGLE model performed better than (300-dimensional) La-
tent Semantic Analysis [19] on a semantic-distance test, and, unlike LSA, BEAGLE can
predict results from human experiments on word priming.
3.2. Modeling Surface and Structural Properties in Analogy Processing
Experiments on children’s ability to process verbal analogies show the importance of

both surface information (who or what the sentence is about) and structural information
(who did what to whom) [20]. Eliasmith and Thagard [21] have successfully modeled
these properties in DRAMA, an HRR-based model of analogy processing. Because HRR
and other VSA approaches support the combination of surface and structural information
through simple vector addition, DRAMA is able to model both components of analogy
in a single representation.
3.3. Variables and Quantification
GOFAI has excelled at representing and reasoning with universally and existentially
quantified varialbes; e.g., ∀ x Computer(x) → HasBuggyP rogram(x)∨Broken(x).
It has been known for some time, however, that human performance on such reasoning
tasks differs in interesting ways from simple deductive logic [22]. Recent work by Elia-
smith [23] shows that HRR encodings of logical rules yield results similar to those seen
in the experimental literature.
3.4. Future Work: VSA in Visual Map-Seeking Circuits
Arathorn ’s Map Seeking Circuits (MSCs) [24] are recurrent neural networks for rec-
ognizing transformed images, using localist representations. We propose treating the lo-
calist MSC as an input recognition device, with the localist output values subsequently
encoded into VSA representations indicating items in the agent’s environment, and their
spatial relationships to the agent. This would allow the representation and manipulation
of multiple simultaneous items in the agent’s environment.
References
[1] Newell, A., Simon, H.A.: Gps, a program that simulates human thought. Lernende Automaten (1961)
109–124
[2] Minsky, M.L.: The Society of Mind. Simon and Schuster (1988)
[3] Rumelhart, D., McClelland, J., eds.: Parallel Distributed Processing: Explorations in the Microstructure
of Cognition. MIT Press (1986)
[4] Rumelhart, D., Hinton, G., Williams, R.: Learning internal representation by error propagation. In
Rumelhart, D., McClelland, J., eds.: Parallel Distributed Processing: Explorations in the Microstructure
of Cognition. Volume 1. MIT Press (1986)
[5] Kohonen, T.: Self-Organizing Maps. 3 edn. Springer-Verlag, Secaucus, NJ (2001)
[6] Hopfield, J.: Neural networks and physical systems with emergent collective computational abilities.
Proceedings of the National Academy of Sciences 79 (1982) 2554–2558
[7] Elman, J.: Finding structure in time. Cognitive Science 14 (1990) 179–211
[8] McClelland, J., Rumelhart, D., Hinton, G.: The appeal of parallel distributed processing. In Rumelhart,
D., McClelland, J., eds.: Parallel Distributed Processing: Explorations in the Microstructure of Cogni-
tion. Volume 1. MIT Press (1986)
[9] Fodor, J., Pylyshyn, Z.: Connectionism and cognitive architecture: A critical analysis. Cognition 28
(1988) 3–71
[10] Pollack, J.: Recursive distributed representations. Artifical Intelligence 36 (1990) 77–105
[11] Minsky, M.L., Papert, S.A.: Perceptrons: Expanded Edition. MIT Press (1988)
[12] Voegtlin, T., Dominey, P.F.: Linear recursive distributed representations. Neural Netw. 18(7) (2005)
878–895
[13] Gayler, R.: Vector symbolic architectures answer jackendoff’s challenges for cognitive neuroscience.
In Slezak, P., ed.: ICCS/ASCS International Conference on Cognitive Science. CogPrints, Sydney, Aus-
tralia, University of New South Wales (2003) 133–138
[14] Smolensky, P.: Tensor product variable binding and the representation of symbolic structures in connec-
tionist systems. Artificial Intelligence 46 (1990) 159–216
[15] Kanerva, P.: The binary spatter code for encoding concepts at many levels. In Marinaro, M., Morasso,
P., eds.: ICANN ’94: Proceedings of International Conference on Artificial Neural Networks. Volume 1.,
London, Springer-Verlag (1994) 226–229
[16] Plate, T.: Holographic reduced representations. Technical Report CRG-TR-91-1, Department of Com-
puter Science, University of Toronto (1991)
[17] Anderson, J., Silverstein, J., Ritz, S., Jones, R.: Distinctive feathers categorical perception and proba-
bility learning: some applications of a neural model. Psychological Review 84(5) (1977) 413–451
[18] Jones, M.N., Mewhort, D.J.K.: Representing word meaning and order information in a composite holo-
graphic lexicon. Psychological Review 114 (2007)
[19] Foltz, T.L.T., Laham, D.: Introduction to latent semantic analysis. Discourse Processes 25 (1998) 259–
284
[20] Gentner, D., Toupin, C.: Systematicity and surface similarity in the development of analogy. Cognitive
Science 10 (1986) 277–300
[21] Eliasmith, C., Thagard, P.: Integrating structure and meaning: a distributed model of analogical mapping.
Cognitive Science 25(2) (2001) 245–286
[22] Wason, P.: Reasoning. In Foss, B., ed.: New Horizons in Psychology. Penguin, Harmondsworth (1966)
[23] Eliasmith, C.: Cognition with neurons: A large-scale, biologically realistic model of the wason task. In
Bara, G., Barsalou, L., Bucciarelli, M., eds.: Proceedings of the 27 th Annual Meeting of the Cognitive
Science Society. (2005)
[24] Arathorn, D.W.: Map-Seeking Circuits in Visual Cognition. Stanford University Press (2002)

Vector Symbolic Architectures: A New Building Material For Artificial General Intelligence

Uploaded by

Copyright:

Available Formats

Vector Symbolic Architectures: A New Building Material For Artificial General Intelligence

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Vector Symbolic Architectures: A New Building Material For Artificial General Intelligence

Uploaded by

Copyright:

Available Formats

Vector Symbolic Architectures: A New

Building Material for Artificial General

Abstract. We provide an overview of Vector Symbolic Architectures (VSA), a class

presented here, and three anonymous reviewers for helpful suggestions.

Lexington, Virginia 24450 USA; E-mail: levys@wlu.edu.

1. Binding and Bundling in Tensor Products

Systematicity and compositionality can be thought of as the outcome of two essen-

[12] show promise for overcoming this problem.

Figure 2. Methods for keeping fixed dimensionality in tensor-product representations.

2. Holographic Reduced Representations and Binary Spatter Codes

3.1. Representing Word Order in a Holographic Lexicon

3.2. Modeling Surface and Structural Properties in Analogy Processing

Experiments on children’s ability to process verbal analogies show the importance of

3.3. Variables and Quantification

3.4. Future Work: VSA in Visual Map-Seeking Circuits

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.