what “magic number” did miller find to be the capacity of short-term memory?

• Loading metrics

Open Access

Peer-reviewed

Enquiry Article

The Emergence of Miller's Magic Number on a Thin Distributed Memory

• Alexandre Linhares,
• Daniel M. Chada,
• Christian N. Aranha

x

• Published: January five, 2011
• https://doi.org/ten.1371/periodical.pone.0015592

Figures

Abstract

Human memory is limited in the number of items held in one's mind—a limit known equally "Miller'due south magic number". We study the emergence of such limits as a result of the statistics of large bitvectors used to represent items in memory, given two postulates: i) the Sparse Distributed Memory; and ii) chunking through averaging. Potential implications for theoretical neuroscience are discussed.

Commendation: Linhares A, Chada DM, Aranha CN (2011) The Emergence of Miller's Magic Number on a Sparse Distributed Retention. PLoS I half dozen(ane): e15592. https://doi.org/x.1371/journal.pone.0015592

Editor: Colin Allen, Indiana University, United States of America

Received: Baronial xvi, 2010; Accepted: November 13, 2010; Published: Jan 5, 2011

Copyright: © 2011 Linhares et al. This is an open-admission commodity distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: This work has been supported by grant E-26/110.790/2009 from the Federação de Apoio à Pesquisa do Estado exercise Rio de Janeiro (FAPERJ-world wide web.faperj.br), grants 470341/2009-2 and 301207/2009-7 from the Conselho Nacional de Desenvolvimento Científico due east Tecnológico (CNPq-www.cnpq.br) of the Brazilian Ministry of Science & Engineering, the Programa de Apoio à Pesquisa (PROPESQUISA) of the Getulio Vargas Foundation (www.fgv.br), and Cortex Intelligence (www.cortex-intelligence.com), due to the employment of its CTO, Dr. C.N. Aranha. The funders had no role in report design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: Dr. Aranha is Chief Technology Officer of Cortex Intelligence, yet the arrangement had no role or intervention in the research and publication of this report, nor does it intend to pursue interests in consultancy, patents, products in development, or marketed products. Furthermore, the presence of Cortex Intelligence amidst the funders does not alter the authors' adherence to all the PLoS One policies on sharing data and materials.

Introduction

Man short-term memory is severely express. While the existence of such limits is undisputed, there is ample argue concerning their nature. Miller [1] described the ability to increase storage chapters past grouping items, or "chunking". He argued that the span of attention could comprehend somewhere around seven information items. Clamper structure is recursive; every bit chunks may comprise other chunks as items: Paragraphs built out of phrases built out of words built out of messages built out of strokes. This mechanism is used to explain the cognitive capacity to store a seemingly countless flux of incoming, pre-registered, data, while remaining unable to blot and procedure new (non-registered) information in highly parallel fashion.

Miller's 'magic number seven' has been subject of much debate over the decades. Some cognitive scientists have modeled such limits past simply using (figurer-science) "pointers", or "slots" (e.g, [2], [three]—come across [4], [v] for debate). However, such approaches practice not seem plausible given the massively parallel nature of the brain, and we believe memory limits are an emergent belongings of the neural architecture of the human brain. As Hofstadter put it a quarter of a century agone [6] : the "problem with this [slot] arroyo is that it takes something that clearly is a very complex consequence of underlying mechanisms and just plugs it in a complex structure, bypassing the question of what those underlying mechanisms might be."(p. 642)

Our objective in this paper is to study these retention limits every bit emergent effects of underlying mechanisms. We postulate ii mechanisms previously discussed in the literature. The commencement is a mathematical model of human memory brought forth past Kanerva [seven], called Sparse Distributed Retentiveness (SDM). We as well presuppose, following [8], an underlying mechanism of chunking through averaging. Information technology is non within the telescopic of this study to contend for the validity of SDM as a cognitive model; for incursions on this broader topic, we refer readers to [ix]–[11], which discuss the plausibility of this Vector Symbolic Compages family of models (in which SDM is contained).

This work, while similar in its mathematical foundations, is different from previous chapters analyses: In [seven], the memory capacity analysis of SDM relates to its long-term retention mechanisms, while nosotros report its short–term memory limits. Our piece of work also differs from that of Plate, in that, regardless of the number of items presented, the retentivity volition only shop (and subsequently retrieve) a psychologically plausible number of items. The deviation becomes salient in Plate'southward own description [12]: "As more than items and bindings are stored in a unmarried HRR the noise on extracted items increases. If besides many associations are stored, the quality will be so low that the extracted items volition be easily confused with similar items or, in extreme cases, completely unrecognizable"(p. 139). Plate is focused on long–term memory; and we will focus on Miller's STM limits.

A number of theoretical observations are fatigued from our computations: i) a range of plausible numbers for the dimensions of the memory, ii) a minimization of a current controversy between dissimilar 'magic number' estimates, and 3) potential empirical tests of the chunking through averaging assumption. Nosotros should first with a brief clarification of our postulates: i) the SDM, and ii) chunking through averaging.

Sparse Distributed Memory

The Sparse Distributed Retentivity (SDM), adult in [vii], defines a memory model in which information is stored in distributed fashion in a vast, sparsely populated, binary accost space. In this model, (a number of) neurons act as address decoders. Consider the space : SDM'due south address infinite is divers allowing possible locations, where defines both the give-and-take length and the number of dimensions of the space: the memory holds binary vectors of length . In SDM, the information is the same as the medium in which it is stored (i.e. the stored items are -bit vectors in -dimensional binary addresses).

SDM uses Hamming altitude as a metric between any ii -bit vectors (hereafter retentiveness items, items, elements, or bitstrings—according to context). Neurons, or difficult locations (see below), in Kanerva's model, hold random bitstrings with equal probability of 0's and 1's—Kanerva [thirteen], [14] has been exploring a variation of this model with a very big number of dimensions (around 10000). (With the purpose of encoding concepts at many levels, the Binary Spatter Code—or BSC—, shares numerous properties with SDM.) By using the Hamming distance every bit a metric, one can readily encounter that the average distance between any two points in the space is given past the binomial distribution, and approximated by a normal curve with mean at with standard deviation . Given the Hamming distance, and large , near of the space lies close to the mean. A low Hamming altitude between whatsoever two items means that these memory items are associated. A distance that is close to the mean means that the memory items are orthogonal to each other. This reflects 2 facts about the organization of human memory: i) orthogonality of random concepts, and ii) close paths betwixt random concepts.

Orthogonality of random concepts: the vast majority of concepts is orthogonal to all others. Consider a non-scientific survey during a cognitive science seminar, where students asked to mention ideas unrelated to the course brought upwardly terms like birthdays, boots, dinosaurs, fever, executive social club, x-rays, and so on. Not only are the items unrelated to cognitive science, the topic of the seminar, merely they are also unrelated to each other.

Close paths between concepts: The organization of concepts seems to present a 'small globe' topology–for an empirical arroyo on words, for example, see [15]. For any two retentiveness items, one tin can readily notice a stream of idea relating two such items ("Darwin gave dinosaurs the boot"; "she ran a fever on her birthday"; "isn't it time for the Supreme Court to ten-ray that executive order?" …and and then forth). Robert French presents an intriguing example in which one all of a sudden creates a representation linking the otherwise unrelated concepts of "coffee cups" and "old elephants" [16]. In sparse distributed retentiveness, any two bitstrings with Hamming distance around would be extremely close, given the aforementioned distribution. And is the expected distance of an average betoken betwixt two random bitstrings.

Of course, for large (such as ), it is impossible to shop all (or fifty-fifty virtually) of the space—the universe is estimated to carry a storage capacity of bits ( bits if one considers quantum gravity) [17]. It is hither that Kanerva's insights concerning sparseness and distributed storage and retrieval come into play: —or a number effectually 1 million—physical memory locations, called hard locations, could enable the representation of a large number of different bitstrings. Items of a large space with, say, locations would be stored in a mere hard locations—the retention is indeed sparse.

In this model, every single particular is stored in several hard locations, and tin can, likewise, exist retrieved in distributed mode. Storage occurs by distributing the particular in every hard location within a certain threshold 'radius' given by the Hamming distance between the item's accost and the associated difficult locations. Dissimilar threshold values for unlike numbers of dimensions are used (in his examples, Kanerva used 100, 1000 and 10000 dimensions). For , the altitude from a random point of the space to its nearest (out of the i million) hard locations will be approximately 424 bits [7] (p.56). In this scenario, a threshold radius of 451 $.25 will ascertain an admission sphere containing effectually 1000 hard locations. In other words, from any point of the space, approximately grand hard locations lie inside a 451-flake distance. All of these accessible difficult locations will be used in storing and retrieving items from memory. We therefore define the part and a difficult location iff , where defines an access radius around of size (451 if ; is the Hamming distance).

A brief example of a storage and retrieval procedure in SDM is in order: to store an detail at a given (virtual) location (in sparse memory) 1 must activate every hard location within the access sphere of (see below) and store the datum in each ane. Difficult locations conduct adders, one for each dimension. To store a bitstring at a hard location , one must iterate through the adders of : If the -th flake of is 1, increment the -th adder of , if it is 0, decrement it. Repeating this for all hard locations in 's access sphere will distribute the information in throughout these difficult locations.

Retrieval of information in SDM is also massively collective and distributed: to peek the contents of each hard location, one computes its related scrap vector from its adders, assigning the -thursday scrap of as a 1 or 0 if the -th adder is positive or negative, respectively (a coin is flipped if it is 0). Detect, however, that this information in itself is meaningless and may not correspond to any one specific datum previously registered. To read from a location in the address infinite, one must activate the hard locations in the access sphere of and get together each related scrap vector. The stored datum volition be the bulk dominion conclusion of all activated hard locations' related flake vectors. If, for the -th bit, the majority of all bit vectors is 1, the last read datum'south -th flake is set to 1, otherwise to 0. Thus, "SDM is distributed in that many hard locations participate in storing and retrieving each datum, and i hard location can exist involved in the storage and retrieval of many data" [18] (p. 342).

All difficult locations inside an admission radius collectively bespeak to an address. Note too that this process is iterative. The address obtained may not accept information stored on it, but it provides a new access radius to (possibly) converge to the desired original accost. I particularly impressive characteristic of the model is its ability to simulate the "tip-of-tongue" phenomenon, in which ane is certain about some features of the desired retentiveness item, yet has difficulty in retrieving information technology (sometimes existence unable to do so). If the requested address is far enough from the original particular (209 bits if ), iterations of the process will not subtract the distance—and time to convergence goes to infinity.

The model is robust against errors for at least two reasons: i) the contribution of any one difficult location, in isolation, is negligible, and ii) the system can readily bargain with incomplete information and nonetheless converge to a previously registered memory item. The model's thin nature dictates that any signal of the space may be used equally a storage address, whether or not it corresponds to a hard location. Past using about one million hard locations, the retentiveness's distributed nature tin "virtualize" the large address infinite. The distributed attribute of the model allows such a virtualization. Kanerva [7] also discusses the biological plausibility of the model, as the linear threshold function given by the access radius tin can be readily computed by neurons, and he suggests the interpretation of some particular types of neurons equally address decoders. Given these preliminaries concerning the Sparse Distributed Memory, nosotros should now proceed to our 2d premise: chunking through averaging.

Chunking through averaging

To clamper items, the majority rule is applied to each bit: given bitstrings to be chunked, for each of the bits, if the majority is 1, the resulting bitstring'south chunk bit is set to one; otherwise it is 0. In case of perfect ties (no majority), a coin is flipped.

We have chosen the term 'chunking' to depict an averaging operation, and 'clamper' to describe the resulting bitstring, because, through this operation, the original components generate a new ane to be written to memory. The reader should note, in SDM's family unit of high-dimensional vector models, chosen Vector Symbolic Architectures (VSA), the operation that generates blended structures is usually known equally superposition [10]–[12].

Obviously, this new chunked bitstring may exist closer, in terms of Hamming distance, to the original elements, than the mean distance between random elements (500 $.25 if  = 1000), given a relatively small . The chunk may and then be stored in the memory, and it may exist used in future chunking operations, allowing, thus, for recursive beliefs. With these preliminaries, nosotros turn to numerical results in the analysis section.

Assay

Computing the Hamming altitude from a chunk to items

Let exist the ready of bitstrings to be chunked into a new bitstring, . The first chore is to find out how the Hamming distance is distributed between this averaged bitstring and the set of bitstrings being chunked. This is, as discussed, accomplished through majority dominion at each flake position. Imagine that, for each carve up dimension, a supreme court will cast a decision with each approximate choosing yes (1) or no (0). If there is an even number of judges, a fair money will be flipped in the instance of a tie. Given that there are votes bandage, how many of these votes will fall in the minority side? (Each minority-side vote adds to the Hamming distance between an item and the average .)

Note that the minimum possible number of minority votes is i, and that it may occur with either 3 votes cast or ii votes and a coin flip. If there are 2 minority votes, they may stalk from either 5 votes or 4 votes and a coin flip, and then forth. We thus have that, for votes, the maximum minority number is given by (and the ambiguities between an odd number of votes versus an even number of votes plus a coin flip are resolved by because full votes). This leads to independent Bernoulli trials, with success cistron , and the constraint that the minority view differs from the bulk bit vote. Let be a random variable with the number of minority votes. Obviously in this instance, , hence we have, for items, the following cumulative distribution office of minority votes [19]:

While we can now, given votes, compute the distribution of minority votes, the objective is not to understand the beliefs of these minority bits in isolation, i.eastward., per dimension on the chunking process. We want to compute the number of dimensions to (in a psychologically and neurologically plausible way) store and call up around items—Miller'south number of retrievable elements—through an averaging functioning. Hence we demand to compute the post-obit:

1. Given a number of dimensions and a set of items, the probability density function of the Hamming distance from to the chunked elements ,
2. A threshold : a number of dimensions in which, if an element 'due south Hamming distance to is further from that point, then cannot be retrieved,
3. Every bit grows, how many elements remain retrievable?

Given bitstrings with dimension , suppose elements have been chunked, generating a new bitstring . Let be the Hamming altitude from the chunked element to , the -th element of . What is the distance from to elements in ? Here we are led to Bernoulli trials with success factor . Since is large, for can be approximated by a Normal distribution, we may employ and . To model human being short term retentiveness'south limitations, nosotros desire to compute a cutoff threshold which will guarantee retrieval of around items averaged in and "forget" an detail if —where is Miller's limiting number. Hence to guarantee retrieval of around 95% ( ) of items, nosotros have , where is the success factor corresponding to . Note that Cowan [xx] has argued for a "magic number" guess of items—and the exact cognitive limit is nevertheless a affair of debate. The success cistron for iv (or 5) elements is  = .3125; and for 6 (or vii) elements it is  = .34375. By fixing the success factor at plausible values of (at {4,v}, or at an intermediary value betwixt {4,five} and {6,7}, or at {half dozen,7}), different threshold values are obtained for varying , as shown in Table 1. In the remainder of this study, we employ the intermediary success factor for our computations; again without loss of generality betwixt different estimates of .

Nosotros thus take a number of plausible thresholds and dimensions. We tin now proceed to compute the plausibility range: Despite the implicit suggestion in Table 1 that whatever number of dimensions might exist plausible, how does the behavior of these combinations vary every bit a function of the number of presented elements, ?

Varying the number of presented items

Consider the case of information overload, when 1 is presented with a large set up of items. Suppose one were faced with dozens, or hundreds, of distinct items. It is not psychologically plausible that a large number of elements should be retrievable. For an item to be incommunicable to retrieve, the altitude between the averaged item and must exist college than the threshold point of the corresponding . When we have an increasingly large ready of presented items, there will be information loss in the chunking machinery, but it should still be possible to retrieve some elements inside plausible psychological premises.

Figure one(a) shows the behavior of three representative sizes of : 100, 212 and thousand dimensions. (100 and 1000 were chosen considering these are described in Kanerva's original examples of SDM.) has shown to be the virtually plausible number of dimensions, preserving a psychologically plausible number of items afterwards presentations of different set sizes. It is articulate that chop-chop diverges, retaining a high number of items in a chunk (as the number of presented items grows). Conversely, if , the number of preserved memory items rapidly drops to zilch, and the postulated mechanisms are unable to retrieve whatsoever items at all—a psychologically implausible development. Effigy 1(b) zooms in to illustrate behavior over a narrower range of -values and a wider range of presented items. Varying the number of presented items and computing the number of preserved items (for a number of representative dimensions) yields informative results. Based on our premises, experiments show that to accordingly reverberate the storage capacity limits exhibited past humans, certain ranges of must be discarded. With too pocket-size a number of dimensions, the model will remember too many items in a clamper. With also large a number of dimensions, the model volition retrieve at most one or two—maybe no items at all. This is because of the higher number of standard deviations involved in the dimension sizes: for , the whole space has 20 standard deviations, and is less than 2 standard deviations below the hateful—which explains why an ever growing number of items is "retrieved" (i.e., loftier probability of faux positives). For , the space has over 63 standard deviations, and , is around 8.99 standard deviations below the hateful. There is such a minute function of the infinite below that item retrieval is most impossible.

With an intermediary success gene between and established past the cognitive limits 4 and seven, we have computed the number of dimensions of a SDM as lying in the vicinity of 212 dimensions. Variance is minimized when —and retrieval results concur psychologically plausible ranges even when hundreds of items are presented (i.e., the SDM would exist able to retrieve from a chunk no more than nine items and at least one or 2, regardless of how many items are presented simultaneously). Finally, given that this work rests upon the chunking through averaging postulate, in the next section we will debate that the postulated machinery is not only plausible, but besides empirically testable.

Results and Give-and-take

The chunking through averaging postulate

Consider the assumption of chunking through averaging. Nosotros propose that information technology is plausible and worthy of further investigation, for iii reasons.

Starting time, it minimizes the electric current controversy betwixt Miller's estimations and Cowan'due south. The disparity betwixt Miller'south or Cowan's observed limits may be a smaller delta than what is argued by Cowan. Our "chunking-through-averaging" premise may provide a simpler, and perhaps unifying, position to this fence. If chunking 4 items has the same probability as v items, and chunking six items is equivalent to chunking 7 items, ane may detect that the 'magic number' constitutes one cumulative probability degree (say, 4-or-v items) plus or minus one (vi-or-7 items).

A mainstream interpretation of the above phenomenon may be that, as with whatever model, SDM is a simplification; an idealized approximation of a presumed reality. Thus, one may run into it equally insufficiently complete to accurately replicate the details of truthful biological function due to, among other phenomena, inherent noise and spiking neural activity. In this case, one would interpret it every bit a weakness, or an inaccuracy inherent to the model. An alternative view, nonetheless improbable, may be that the model is accurate in this item aspect, in which example, the assumption minimizes the current controversy between Miller's estimations and Cowan's.

The success factors computed in a higher place show that for either 4 or 5 items, nosotros take , while for six or 7 items nosotros have . If we assume an intermediary value of —which is reasonable, due to dissonance or lack of synchronicity in neural processing—the controversy vanishes. We chose to base our experiments on the mean value ( ), and the results herein may be adjusted to other estimates as additional experiments settle the debate.

Moreover, a chunk tends to exist closer to the chunked items than these items are betwixt themselves. For case, with and , the Hamming distance betwixt a clamper and a random item is fatigued from a distribution with and ; in here, from the signal of view of the chunked item , the closest one% of the infinite lies at 53 $.25, while 99% of the space lies at 84 bits. Contrast this with the distances between any two random, orthogonal, items, which are fatigued from and : from the point of view of a random item, the closest 1% of the space lies at 89 bits, while 99% of the infinite lies at 122. This disparity reflects the principles of orthogonality betwixt random concepts and of close paths between concepts (or small worlds [15]): the altitude between 2 items from any five is large, merely the distance to the average of the set is small. Of course, equally grows, the distance to also grows (since ), and items become irretrievable. I thing is clear: with 5 chunked items, the chance of retrieving a faux positive is minute.

Finally, the assumption of chunking through averaging is empirically testable. Psychological experiments concerning the difference in ability to retain items could test this postulate. The assumption predicts that (4, five) items, or more generally that ( ) for integer will exist registered with equal probability. It also predicts how the probability of retained items should drop in relation to if . This is counterintuitive and tin be measured experimentally. Note, yet, 2 qualifications: outset, as chunks are hierarchically organized, these effects may exist hard to perceive in experimental settings. One would have to devise an experimental setting with assurances that merely chunks from the same level are retrievable–neither combinations of such chunks, nor combinations of their constituting parts. The final qualification is that, as grows, the same probability difference tends to zero. Because of the conjunction of these qualifications, this effect would be hard to perceive on normal human behavior.

Final remarks

Numerous cerebral scientists model the limits of man short-term memory through explicit "pointers" or "slots". In this paper we accept considered the consequences of a short-term retention limit given the mechanisms of i) Kanerva'southward Thin Distributed Memory, and ii) chunking through averaging. Given an appropriate choice for the number of dimensions of the binary space, nosotros are able to model chunks that limit active memory's storage capacity, while allowing the theoretically endless recursive clan of pre-registered memory items at different levels of abstraction (i.e., chunks may exist chunked with other chunks or items, indiscriminately [1], [21]). This has been pointed out in [22], however, in here we use the short-term retentivity limitations as a bounding factor to compute plausible ranges for .

Some observations are noteworthy. Outset, our work provides plausible premises on the number of dimensions of a SDM—we make no claims concerning Kanerva's recent piece of work (e.1000., [14]). Given our postulates, it seems that 100 dimensions is too low a number, and 1000 dimensions too loftier. In our computations, assuming , variance of the number of items retained (as a function of the number of presented items and at least ane retrievable detail) was minimized at 212 dimensions. This value was chosen as our optimal bespeak of focus for information technology provided stable, psychologically plausible behavior for a wide range of set sizes. We have concentrated on the SDM and chunking through averaging postulates, yet time to come research could also look at alternative neural models; for it is certain that the brain does not use explicit slots or pointers when items are chunked. One tin can reasonably argue: what adept tin can come from replacing one magic number with another? There are 2 potential benefits: first, past fixing parameter , we can restrict the blueprint infinite of SDM simulations and ensure that a psychologically plausible number of items is chunked. Another advantage is theoretical: the number 212 suggests that we should look for neurons that seem to have, or answer majoritarily to, such a number of active inputs in their linear threshold office.

Of form, a single 212 bit vector in SDM does non encode meaningful content at all. The existence of a bitstring can only be meaningful in relation to other bitstrings shut to it. Consider, for instance, an A4 sheet of paper, of size 210mm×297mm (8.3in×eleven.7in). A 1200×1200 dots-per-inch printer holds less than potential dots in an unabridged sheet. While the infinite of possible black and white printed A4 sheets is a very large gear up of possible pages, the vast majority of them, rather like the library of Babel, are equanimous of utter gibberish. Whatever single dot needs only 28 bits to exist described, and because the dots usually cluster into strokes, chunks tin be formed. Moreover, because strokes cluster to course fonts, which cluster to form words, which cluster to form phrases and paragraphs; combinations of large sets of 212 dimensional bitstrings can encode the meaningful content of pages and books—provided those items have been previously chunked in the reader's mind. Without chunks there tin can be no pregnant; this paragraph, translated to Yanomami (assuming that'south possible), would become unreadable to its intended audience and to its authors.

Sparse Distributed Memory holds a number of biologically and psychologically plausible characteristics. It is associative, allowing for authentic retrieval given vague or incomplete information (which is relevant given the potential for asynchronous behavior [23]); it is readily computable by neurons; it seems suitable for storage and retrieval of depression-level sensorimotor data [24], it is a plausible model of the space of human concepts, and it exhibits a phenomenon strikingly similar to the tip-of-the-natural language situation. With the results presented herein, sparse distributed memory also reflects the natural limits of man short-term memory.

Acknowledgments

The authors would similar to thank Eric Nichols for numerous valuable comments.

Writer Contributions

Conceived and designed the experiments: AL DMC CNA. Performed the experiments: AL DMC CNA. Analyzed the data: AL DMC. Wrote the paper: AL DMC CNA.

References

1. 1. Miller GA (1956) The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychol Rev 63: 81–97.
   • View Article
   • Google Scholar
2. ii. Gobet F, Simon HA (2000) Five seconds or lx? Presentation time in skilful retention. Cerebral Sci 24: 651–682.
   • View Commodity
   • Google Scholar
3. iii. Gobet F, Lane PCR, Croker Southward, Cheng PCH, Jones Chiliad, Oliver I, Pine JM (2001) Chunking mechanisms in human learning. Trends Cerebral Sci 5: 236–243.
   • View Commodity
   • Google Scholar
4. 4. Linhares A, Brum P (2007) Understanding our understanding of strategic scenarios: What role do chunks play? Cognitive Sci 31: 989–1007.
   • View Article
   • Google Scholar
5. five. Linhares A, Freitas AETA (2010) Questioning Chase and Simon's (1973) "Perception in chess": The "experience recognition" hypothesis. New Ideas Psychol 28: 64–78.
   • View Article
   • Google Scholar
6. vi. Hofstadter DR (1985) Metamagical Themas. Questing for the Essence of Mind and Pattern, Basic Books, New York, 1985.
7. 7. Kanerva P (1988) Thin Distributed Memory. MIT Press, Cambridge.
8. 8. Kanerva P (1998) Big patterns make great symbols: An example of learning from example.
9. ix. Kanerva P (1993) Sparse Distributed Memory and related models. In: Hassoun MH, editor. Associative Neural Memories: Theory and Implementation. New York: Oxford University Press. pp. 50–76.
10. ten. Stewart TC, Eliasmith M (2009) Compositionality and biologically plausible models. Oxford Handbook of Compositionality.
11. xi. Gayler RW (2003) Vector Symbolic Architectures answer Jackendoff's challenges for cognitive neuroscience. In: Slezak Peter, editor. ICCS/ASCS International Conference on Cognitive Scientific discipline. Sydney, Australia: University of New South Wales. pp. 133–138.
12. 12. Plate TA (2003) Holographic reduced representations. Distributed representations for cognitive structures, CSLI Publications, Stanford, CA.
13. xiii. Kanerva P (1994) The Spatter Lawmaking for encoding concepts at many levels. In: Marinaro M, Morasso PG, 1 vol, editors. ICANN '94, Proceedings of International Conference on Artificial Neural Networks, vol. 1. London: Springer–Verlag. pp. 226–229.
14. fourteen. Kanerva P (2009) Hyperdimensional calculating: An introduction to computing in distributed representation with loftier-dimensional random vectors. Cognitive Comput 1: 139–159.
    • View Article
    • Google Scholar
15. xv. Cancho RF, Sole R (2001) The small-scale earth of man language. Proc Purple Soc B 268: 2261–2265.
    • View Article
    • Google Scholar
16. 16. French RM (2000) 93–100. DOI:https://doi.org/x.1007/978-0-585-29605-0_10.
    • View Article
    • Google Scholar
17. 17. Lloyd S (2002) Computational capacity of the Universe. Phys Rev Lett 88: 237901–1:4.
    • View Article
    • Google Scholar
18. xviii. Anwar A, Franklin S (2003) Thin distributed retentivity for "conscious" software agents. Cognitive Sys Res 4: 339–354.
    • View Commodity
    • Google Scholar
19. 19. Boland PJ (1989) Majority Systems and the Condorcet Jury Theorem. The Statistician 38: 181–189.
    • View Article
    • Google Scholar
20. 20. Cowan N (2001) The magical number four in short-term memory: A reconsideration of mental storage capacity. Behav Brain Sci 24: 87–185.
    • View Article
    • Google Scholar
21. 21. Linhares A (2000) A glimpse at the metaphysics of Bongard problems. Artif Intell 121: 251–270.
    • View Article
    • Google Scholar
22. 22. Kanerva P (1996) Binary spatter-coding of ordered k-tuples. ICANN96, Bogus Neural Networks. Springer. pp. 896–873.
23. 23. Borisyuk R, Denham 1000, Hoppensteadt F, Kazanovich Y, Vinogradova O (2000) An oscillatory neural network model of thin distributed memory and novelty detection. Biosystems 58: 265–272.
    • View Article
    • Google Scholar
24. 24. Assisi C, Stopfer M, Laurent G, Bazhenov Thou (2007) Adaptive regulation of sparseness past feedforward inhibition. Nature Neurosci 10: 1176–1184.
    • View Article
    • Google Scholar

coughlanstindere.blogspot.com

Source: https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0015592

Share this post