WO2016182422A2 - Associative memory system based on the diagrammatic abstraction of specified contents as attribute-value structures - Google Patents

Abstract

The invention relates to an associative memory system for saving, recognising and returning contents represented as attribute-value structures, in which said structures are also the indices for directly accessing the memory. The aforementioned contents consist of acoustic or visual images or other information represented as discrete functions with finite domains, or families of finite abstractions or functions. The information is stored in a rectangular grid, in a parametric manner, for m attributes with n values. Each state of the grid represents one content or one abstraction and has a unique descriptor and an associated entropy, the definitions of which also form part of the present application. The mechanism is based on the operations of (a) diagrammatic abstraction, in order to record information in the memory, and (b) diagrammatic reduction in order to recognise and/or extract information from the memory, the definitions of which form part of the content of the present application. The invention also includes the logic structure of the memory and its control program. The memory can be simulated in an ordinary digital computer or constructed as an integrated circuit, either as part of a central processing unit or as an independent integrated circuit.

Description

 ASSOCIATIVE MEMORY SYSTEM BASED ON THE DIAGRAMIC ABSTRACTION OF CONTENTS SPECIFIED AS STRUCTURES

ATTRIBUTE-VALUE FIELD OF THE INVENTION

 The present invention is framed in the field of computer systems, in particular in the design of memories for storing, recognizing and distributing large amounts of information, especially of a qualitative nature, such as visual or acoustic scenes, or other information modes, as well as its relationship with standard direct access memories and the central processing unit.

BACKGROUND

The memories in digital computing systems of the ordinary type of direct access (Random Access Memory or RAM) have an associated address which is used to write and read contents thereof. Likewise, to make comparisons or calculations with said contents, it is necessary to extract them from their memory register and take them to the central processing unit where the computation is performed properly.

The use of memories of the ordinary type requires the definition of index system to register and remunerate content, so the storage or registration process and information retribution involves the conversion of content to the Index and vice versa; also, if the contents of two or more memory units are related, said relationship has to be explicitly stored and indexed in an additional memory location.

The standard way to access content directly is through the so-called "hash" keys, schemes or functions. These functions map the coding of the contents, as they are entered into the computer system, to specific memory locations. However, these schemes constitute a very limited way of association since the arguments of the hash functions are syntactic objects and not interpretations.

In associative memories, meanwhile, the reference to a content is the content itself. In these associative memories the access keys are the interpretations or meanings of the descriptions, so that the associative memory systems have an essentially semantic character. The difference between standard and associative type memories is qualitatively illustrated by the corresponding access schemes, as follows:

• RAM Memory: Scene Descriptor => Hash Function => Memory Address • Associative Memory: Descriptor of the scene => Interpretation => Semantic Representation => Access to Memory (Registration, Recognition and recovery). A necessary characteristic of associative memories is that the relationships between stored contents are represented, either implicitly or explicitly, in the memory substrate, which allows them to be directly accessible. For this reason and unlike direct access memories, associative memories do not have individual records with specific addresses and the information is represented in a distributed way, so that a content can be stored in several basic units and a basic unit can contribute to the storage of various contents. This property is what allows "associativity" in said class of memories.

There is very little history of computational associative memories. A proposed scheme within the paradigm of neural networks, for example, is the Hopfield networks (1982). However, the performance of these systems is very limited and to date there are no systems with practical applications for storing large amounts of information. Another antecedent is associative memories of a propositional nature, such as semantic networks and related systems (Anderson and Bower, 1980), but these refer to the associative nature of the contents, which are usually stored in direct access memories of the ordinary type, or are postulated as theoretical models of human memory, but do not refer to the associative nature of the memory substrate itself, as the system that It is presented in this document.

SUMMARY OF XA INVENTION

 The associative motive memory of the present invention consists in a reticular arrangement of processing units in which each cell is a processor capable of performing a minimum set of instructions. This arrangement is illustrated as an RM (Memory Record) memory arrangement in Figure 1. The contents in this arrangement are represented as attribute attribute structures, where each column corresponds to an attribute of the represented object and each line to the value of said attribute. for the corresponding object.

The memory register operation consists in applying the diagrammatic abstraction operation, defined below, to the current content of the memory register with the content to be stored, and the result of said operation is written directly to the same memory register. For this reason the associative memory keeps abstractions of the contents that are recorded incrementally. Each memory state and each content to be stored have a unique identifier associated with it. The abstraction operation is applied also to these descriptors and the resulting abstraction also has a unique descriptor. Consequently, each diagrammatic abstraction has a dual description in the symbolic domain that an external computer system of the ordinary type can be constructed and used.

The recognition operation consists in applying the diagrammatic reduction operation, described below, between the memory register and the content to be recognized. If this operation is successful, the content was previously stored in the memory, and otherwise the content is unknown.

The recovery operation corresponds to the location of the descriptor of the content to be associated in the hierarchy of descriptors that corresponds to the abstraction stored in memory, and in the recovery of the associated descriptors in the hierarchy of abstraction. The scheme is quite open and allows recovering associations of both more general and more specific content of the content to be associated, as well as contents at the same level of abstraction that are subordinated to common general contents.

The memory is associative because the attribute-value structures that are used to recognize or reward content Associated are the same or similar to the content originally stored.

The abstraction operation is destructive, so the abstractions contained in the memory are not completely reversible, in addition to having a degree of indeterminacy. Therefore, in this system an entropy value is defined for all the contents stored in the memory. The entropy value increases as the number of stored contents increases, and there is a maximum entropy value that imposes a practical limit on its capacity, since above that value the memory becomes saturated and begins to produce false positives in recognition and recovery operations.

The central functionality of the memory is supported by a number of registers and auxiliary functions, as shown in Figure 1. In particular, the memory has an RI / O input / output register in which both the contents that are written are written. they are recorded in memory and it is encrypted if the recognition and recovery operations are successful.

The device can be built in two different versions; in the first one a work record RT is included in which the results of the reduction operation are written; in this version the RM memory record is not altered in that operation and memory is conservative; in the second implementation the work register does not exist, and the result of the reduction operation is written in the memory register, so the memory is not conservative, since although the recognition or recovery of a content reinforce that content readable in memory, other associated readable content may be degraded as a result of such operations. The memory also includes an auxiliary register RA to construct abstract contents from the basic descriptors of a scene. For example, a visual scene can be represented by a relatively large set of basic descriptors, and the auxiliary register allows to construct the abstraction of said set of descriptors, which corresponds to the representation of the whole scene. This abstraction is constructed in the auxiliary register and is the content that is registered, recognized or associated in the RM memory register via the RI / O input / output register.

The device also has a unit to calculate and display the entropy value in the RM memory register at the end of a registration operation and the entropy difference in the recognition and association operations and make these values available or deployed to external devices. through an interphase. The device also has a unit to determine the identifiers of the abstractions in the memory and in the entry and exit register in the registration, recognition and association operations to make them available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a unit of process and memory of the ordinary type, which is connected to associative memory through this inferred.

BRIEF DESCRIPTION OF THE FIGURES

Figure 1: Conceptual model of the associative memory system. The system contains four two-dimensional parametric arrays of size m x n of process units, aligned vertically, as follows (the information flow is illustrated with thick arrows while the control one with thin arrows):

REG. AUXILIARY CONV. FUNCTIONS / ABSTRACTIONS (RA)

REG. INPUT / OUTPUT (RI / O) MEMORY RECORD (RM)

WORK RECORD (RT) The system contains the following additional records and control units:

SERIAL INPUT REGISTRATION: This register is fed with functions and / or descriptors of concrete and / or abstract images that are directly abstracted in RA prior to initialization of the latter.

PARALLEL ENT / SAL CONTROL UNIT: This register is an interface to write and / or read concrete or abstract images directly as two-dimensional arrangements in the RI / O register.

PROCESSOR / MEMORY IMAGES (IDs). This record contains a memory with the identifiers of all the abstractions (concrete and abstract images) that have been added to the memory, because they have participated as inputs or outputs in the registration and recognition operations, as well as an associated time stamp to each of these identifiers. The output of this module consists of the identifier of the abstraction stored in a registration operation with its time stamp, or of the recognized image or the associated image in the recognition and recovery operations respectively, also with its time stamps. ENTROPY PROCESSOR: This module calculates and displays the entropy of the images whose identifier is stored in the image memory processor (IDs) and produces the value of the RM entropy at the end of each registration and recognition operation, as well as the difference between RM and RT at the end of a recognition operation (or the difference in RM entropy caused by recognition if RT is dispensed with, as indicated in Proposition 40).

CONTROL UNIT. This module executes the program of control of the operations or steps of registration, recognition and retrieval of information, and sends the corresponding signals to the modules of the architecture involved in each of these operations.

Figure 2: Diagrammatic representation and enumeration of discrete functions. The basic representation format of a discrete function is shown as the attribute value structure, where the columns represent the arguments and the corresponding value lines. It also shows the unique identifier for each function as well as the enumeration of all functions for n arguments with m possible discrete values. Figure 3: Illustration of the basic λ diagrammatic abstraction operation. The arguments of the operation (with the function and argument to the left and right respectively) and the resulting abstraction are shown in the lower and upper parts of the figure respectively.

Figure 4: Illustration of the composite λ diagrammatic abstraction operation. The arguments of the operation (with the function and argument to the left and right respectively) and the resulting abstraction are shown in the lower and upper parts of the figure respectively.

Figure 5: Illustration of the diagrammatic reduction operation β. Two complementary examples of this operation are shown. The arguments of the operation (with the function and argument to the left and right respectively) and the function resulting from said operation are shown in the upper and lower parts of the figure respectively. Figure 6: Illustration of the operation of diagrammatic reduction β between abstractions. The arguments of the operation (with the function and argument to the left and right respectively) and the resulting abstraction are shown in the upper and lower parts of the figure respectively. Figure 7: Illustration of the operations of abstraction λ and diagrammatic reduction β with abstractions that include incomplete knowledge and restrictions. An example of the abstraction operation (on the left) and two example of reduction (center and right) are shown.

DETAILED DESCRIPTION OF THE INVENTION

 Definitions

 For purposes of the present invention, an associative memory should be understood as an abstraction whose content is a set of abstract images and that can be read by a processing architecture.

For purposes of the present invention, registration of an image should be understood as an operation of storing a concrete or abstract image in associative memory.

For purposes of the present invention, recognition should be understood as the verification that a concrete or abstract image is contained in the report.

For purposes of the present invention, recovery should be understood as the retribution of an image associated with a reference (concrete or abstract) image or "track" {tried or clue). It is possible to characterize information of different types, such as visual images, auditory images or other types of perception modalities, in terms a finite set of attributes with their respective values. For example, those of SIFT descriptors in the case of visual images (Lowe, 2004) or the frequency components of an acoustic image in a time segment.

Attribute-value structures are the product of the interpretation of sensory information (e.g., acoustic, auditory, etc.) in relation to a particular mode of interpretation. For example, visual sensations are stored in a "pixel buffer" (not interpreted) which can be considered as a syntactic descriptor of a particular scene. This descriptor can be processed by an interpretive algorithm (e.g. SIFT) and represented as a value attribute structure. Yes, said structure is associated with a concept, said algorithm is an interpreter in relation to a conceptual domain (Proposition 1).

An associative memory is a system to store, recognize or retrieve images from their attribute-value structures directly (Proposition 2). A descriptor corresponds to a mathematical function where the set of attributes is the domain and the set of values is the range (Proposition 3). Two or more discrete functions with the same domain are similar to the extent that the corresponding arguments of both functions have equal or similar values, for all arguments. In extreme cases two or more functions are identical if the corresponding arguments of both functions have the same value, for all arguments, and are significantly different if the values of the corresponding arguments differ significantly, for all arguments (Proposition 4 ). If the range is the set of real numbers, it can be divided and normalized as a discrete set of levels, so that each descriptor is characterized as a function of attributes at levels (Proposition 5). The representation of a set of functions is defined as abstraction, where arguments with common values are represented in a shared way (ie, they use the same substrate as the representation medium) (Proposition 6). The informational content of an abstraction is proportional to the number of arguments with shared values. Consequently, abstractions have an associated entropy, which increases in inverse (logarithmic) relation to the number of arguments with common values. There is a higher level than the entropy of an abstraction for a finite set of functions, which corresponds to the case in which no function in the abstraction shares any of its values (Proposition 7).

A concrete image is defined as the abstraction of a finite set of descriptors of an object or scene (e.g. visual, acoustic, etc.) (Proposition 8).

An abstract image is defined as the abstraction of a set of concrete images of the same object or scene (Proposition 9).

Theory of abstraction and reduction diagrammatic

A discrete function can be represented in a rectangular grid where each column corresponds to an element of the domain and each line to an element of the range, and the functional relationship is characterized by a mark in the cell that intersects the column and corresponding line, for all columns (Proposition 10). The set of functions in a ra * n table can be listed as follows: Let f * be the function k in the enumeration, where k is a numeral of n digits in base consisting of

the concatenation of the index i of each column, such that it is not defined

 so

For example, the table in Figure 2 shows the representation of the total fun function where n = 4 and m = 7.

All total functions can be easily identified in this representation since they do not contain Os included in their indexes. The number of partial and total functions that can be represented in a Notation table

allows to identify the constant functions directly since all the digits of the index are the same, and also the functions one by one, since all the digits of the index are different. Since m, n are enumerable, the notation includes all discrete functions that can be represented in each table. If m ≥ n, the one-to-one functions correspond to the permutations of the m objects in the range taken n (the cardinality of the domain) at the same time and the number of permutations is also the combinations

correspond to the functions whose indexes include the exact same digits, and the number of combinations is

Diagrammatic abstraction is defined as the superposition of two or more discrete functions with the same domain and range on the same grid. This operation is designated as "λ". If there are arguments of two or more functions with the same value, the abstraction captures this relationship directly with the mark in the corresponding cell. An example of the diagrammatic abstraction of two functions that share the same value for one of their arguments (ie, di) (Proposition 12) is shown in Figure 3.

Every abstraction has an index or identifier that consists of a regular expression of n positions (ie, the cardinality of the domain), where each position includes the value of the corresponding argument if it is unique, or the disjunction of the values of the same argument, contributed for the functions that are included in the abstraction. For example, in the index of the abstraction of Figure 3 is "(1 + 2) (2 + 6) 4 (5 + 7)". The index of a function is simply the index of an abstraction in which all arguments have only one value (Proposition 13).

Abstractions produced by the λ operation can be used on a recurring basis in the construction of more complex abstractions. Figure 4 shows a abstraction produced by the abstraction produced in Figure 3 and a new function (Proposition 14).

Indices of higher abstractions are sub-specifications of lower abstractions and vice versa. For example, the index abstraction in Figure 4, is less specific than the index of

the abstraction in the Figure due to

 the disjunctions of the last two positions. The structure of the indices naturally determines a hierarchical structure with the most general abstraction as the upper node and the particular functions in the lower nodes of the structure, with all the abstractions involved in the construction at intermediate levels (Proposition 15).

Let Γφ and Γψ be two arbitrary abstractions. The abstraction operation λ (Γφ, Γψ), where Γφ is the function of abstraction and Γψ the argument of abstraction, is defined as the inclusive disjunction between the corresponding cells in the respective tables, that is:

for everything and everything j, as shown in the

Definition 1. In Table X and B they indicate whether the corresponding cell is "marked" or "blank" respectively (Proposition 16).

DEFINITION 1: DIAGRAMMATIC ABSTRACTION

The informational content of an abstraction is greater while the abstraction is more specific and vice versa, so each abstraction has an associated entropy value. The entropy of an abstraction Γ is defined as the average of the indeterminacy of its arguments, as follows: Let the number of cells in the column BE the number of columns

(ie, arguments) in abstraction Γ. In case the function does not have a defined value for the argument i (ie, v¿ = 0) XÍ = 1. The entropy e (T) or indeterminacy value of the abstraction Γ, is defined as follows:

DEFINITION 2: ENTROPY OF AN ABSTRACTION For example, the entropy value for the case of Figure 3 is Θ (ΓΑΒ «>) = 0.75. In case the abstraction Γ is completely determined (ie, the object is a basic function) the entropy e (T) = 0. There is also a maximum value for the abstraction of functions that do not share any of their arguments; for example, in the case of two functions with 4 arguments this value is 1. Every table has a maximum entropy value for the case in which all the cells are marked; for example, in the case of Figure 3 this is 2.8073. (Proposition 17).

It is defined as reduction (or functional application) diagrammatic, designated as "β". Sean

three abstractions such that or simply

= Γφ ', where Γφ is designated as the function, as Γψ the argument as the value of the reduction. The operation β

it is the inverse operation of diagrammatic abstraction, such that

where Γφ 'is the most specific abstraction contained or subsumed in which in turn

subsume to Γφ. , as specified in definition 3. In case the abstractions involved are completely reversible and the process does not involve loss of information Γφ = Γφ '. (Proposition 18).

DEFINITION 3. REDUCTION AT THE LEVEL OF THE CELLS

Cases a and c in definition 3 show that the reduction operation is not deterministic. In the case the mark in the cell is contributed by the cells

corresponding in the abstractions by

what the brand should be preserved in the reduction; however, in case c, the mark on is provided

by abstraction Γ so it must be extracted in the

reduction. Additionally, for the reduction to be applicable Γψ if contained within (Proposition 19)

With these considerations _/ defined

as the most specific abstraction such that according to

 this "extracts" as many as possible from

in the reduction, as long as no X is removed that was already contained in both Γφ and original in the original construction of this operation (extract the largest number

possible of marcas) marks is defined operationally as the difference between the columns for all

 the columns i. However, if all cells in both columns have the same value

 Then the reduction is the same column:

According to these considerations, the

β reduction operation is defined as follows: Definition 4: Reduction or functional application

1. Application condition: For all cells i

is contained in

2. Reduction Rule: For all columns i if Γφ (ί) = Γψ (ι) then otherwise,

difference between the columns). The reduction operation is illustrated in Figure 5. This example is reversible and the two functions involved in the construction of the abstraction in Figure 3 are fully recovered (Proposition 20).

The abstraction of an abstraction with itself is the abstraction mass, and the reduction of an abstraction with itself is also the same abstraction, and consequently

if only the difference between the

 columns reducing an abstraction with itself would produce empty abstraction,

 (Proposition 21).

Abstraction and reduction operations are not reversible in all cases; This is due to non-determinism in the reduction (ie, cases a and in definition 3); by

 example yes then

however the condition

 It is satisfied. (Proposition 22)

The reduction of an abstraction with another abstraction contained in the first results in a more specific abstraction or directly in a function. Likewise, the abstraction resulting from the reduction has an informative content greater than the reduced abstraction and subsequently a lower entropy value; that is to say: yes then

For example, Figure 6 illustrates an entropy abstraction of 1.1462 applied to an entropy abstraction of 0.75, and the reduction produces an entropy abstraction of 0.5. (Proposition 23)

Up to this point it has been assumed that there is complete knowledge of the abstractions represented in the grid: a mark in a cell is interpreted as having the value associated with the column associated with the corresponding line, and a blank cell is interpreted as the argument associated with the column does not have the value associated with the line. However, the format can be extended to allow the expression of incomplete knowledge. In this extension a blank cell is interpreted as not knowing if the argument associated with the column has the value associated with the corresponding line, that is, the value of blank cells is not completely determined or is sub-specified (a cell with a mark is interpreted as the definitive relationship of the argument with its value, as the base case). This condition does not mean that the value of an argument is not defined since in the latter case the knowledge is complete, nor that all the cells in the column are marked, since this expresses the disjunction of all possible values for the corresponding argument (Proposition 24). In this new scenario it is possible to specify that an argument of a function or an abstraction cannot take a certain value or values; that is, it is possible to specify restrictions. For this purpose we extend the notation with the symbol in such a way that a cell with this symbol is interpreted as the argument associated with the column cannot take the value associated with the corresponding cell line (Proposition 25).

The operation of diagrammatic abstraction with restrictions and incomplete knowledge extends to definition 1 with definition 5, as follows (Proposition 26): Definition 5:

1. Yes then

shared restrictions are

 preserve in abstractions).

2. Yes then

the abstraction of a

 restriction with B or X results in a sub-specification. In this case the corresponding cell abstraction is not marked and its value can be either B or X.

The introduction of restrictions and incomplete information involves several considerations for the definition of diagrammatic reduction; in particular, the introduction of B and by the operation of abstraction, according to definition 5, increases the non-determinism of the reduction. In the event that the abstraction to be reduced has a cell marked with the reduction, it must preserve the restriction; however, a cell with B in an abstraction can be produced by the three types of values (X, B and "-") as a consequence of clause (2) of definition 5, so a cell marked as B in the abstraction to be reduced, is also marked as B in the corresponding reduction. That is, in the reduction with restrictions and incomplete information the less specific abstraction is chosen that guarantees that the reduction is conservative (not is over-specified). With these considerations, definition 4 of the diagrammatic reduction β for

the case in which there is at least one restriction either in column Γφ (i) or column In case the columns

Only contain Xs or Bs, definition 4 applies directly.

Definition 6:

1. Application condition: For all cells i, j, yes

then an X in one

 cell of an abstraction cannot be the abstraction of a constraint) and for all cells i, j, if r

then a restriction can only

 be the abstraction of restrictions). 2. Reduction Rule:

The abstraction and diagrammatic reduction with constraints and incomplete information are illustrated in Figure 7. The case a shows the abstraction of the Γφ and Γψ abstractions that involve both common and different values and constraints, where disjunctive values and common constraints propagate towards Up in abstraction. Case b illustrates the reduction β (λ (Γφ, Γψ), Γφ) that sub-specifies the abstraction Γψ, which is not fully recovered. In this case the constraint that the argument di cannot have as a value of r4 is preserved, but the value m and the constraint of n are sub-specified in the reduction. Case c shows the opposite reduction β (λ (Γφ, Γψ), Γψ) resulting in the sub-specification of Γφ, where the argument di is sub-specified for the value

but the value rs and the constraint r <are preserved.

Figure 7b also illustrates that the reduction operation may produce an abstraction that is less specific than both the "function" abstraction and the "argument" abstraction of a reduction, so that not only the abstraction but also the reduction may involve loss of information.

It would be possible to give a more informative definition of reduction with restrictions and incomplete information, such that reducing an abstraction with itself would result in the same abstraction, even if there were restrictions involved. However, this definition would not be conservative or valid. This can also be seen in Figure 7, where although Γφ ≠ Γψ, abstraction

Γφ) = Γφ, but the alternative reduction

over-specify Γψ (since it would be stipulated that the argument di has the value of rs but also that it has the constraint on r3). Therefore, the most conservative definition of reduction is adopted to ensure that the system is valid (Proposition 27). The introduction of incomplete information and restrictions changes the entropy of abstractions. In this case, cells marked as Bs express indeterminate information with the consequent increase in entropy, together with the increase in entropy due to cells marked with X. On the other hand, the restrictions constitute information completely determined by what does not increase to Entropy Based on these considerations, the entropy for this case is modified as follows: let XI be the number of Xs, bi the number of targets in column i, and n the number of columns in abstraction Γ. The degree of indeterminacy of column i is proportional to a = axi + bi such that 0 <α <1 since although both Xs and Bs increase indeterminacy, Xs are more determined than Bs.

The entropy of abstraction Γ is defined as follows:

DEFINITION 7: ENTROPY OF AN ABSTRACTION WITH INFORMATION

 INCOMPLETE AND RESTRICTIONS

For example, in Figure 7 for a

As you can see, reducing an abstraction with itself can produce a less specific abstraction (Figure 7.b), and abstracting and reducing a function abstraction with the same argument abstraction can also produce a less specific abstraction than the argument abstraction ( Figure 7.c). Likewise, if all the cells in the column are restrictions, the function or abstraction has no value for the argument corresponding to the column, and since this information is completely determined, the entropy of the column is "0". On the other hand, if all the cells in a column are B, so the entire column is undetermined, the entropy is maximum (Proposition 28). The increase in the expressive power of the system with incomplete information and restrictions has the consequence that the reduction operation can produce less specific abstractions and becomes a generator of new abstractions, in the same way as the abstraction operation (Proposition 29). The expression of constraints and incomplete knowledge requires extending the notation of the Indexes or descriptors of abstractions to express that the cells marked with "-" cannot be the value of the argument that corresponds to the column, for all columns of the abstraction. For this purpose the vocabulary of symbols used in the formulation of descriptors with the symbols "res" is increased,

for the formulation of a form term

to specify the restrictions of a column, as well as a term vai (viv ... vvh) to specify the values, where r ± and V denote the restriction or the value of the cell with index i on base m + 1 for each cell a of column j. With this extension the descriptors of abstractions are specified by means of a context-free grammar of form S-> CS | A and

where A is the empty string. For example, him

abstraction descriptor with restrictions on the

Figure 7.a (above) is "res (4) Aval (5v7)" and the descriptor of Γψ in the same figure (bottom right) is "res (3A4) Aval (7)". Likewise and as already mentioned, a partial function is represented by the conjunction of the negation of the identifiers of all the cells in the column, so the digit "0" is not used to designate arguments without value in the descriptor of the column that corresponds to the argument for which a value is not defined.

It is also possible to use this same grammar to specify the descriptor of unrestricted abstractions (instead of the descriptor specified in Proposition 13) since in this case the variable R is rewritten as Λ (ie, the empty string) in production C-> RV. In this case it is necessary to increase the option of writing a "0" in the case that the argument corresponding to the column has no value (i.e., C-> RV | 0). (Proposition 30)

Functional description of the associative minority

 An associative memory is defined as an abstraction whose content is a set of abstract images (in accordance with Proposition 9) and which can be read by a processing architecture. (Proposition 31)

The registration of an image is the operation of storing a concrete or abstract image in associative memory; for this purpose, the operation of diagrammatic abstraction is used where associative memory is the function of abstraction and the image to be stored is the argument of abstraction. The output of a registration operation is the identifier of the corresponding image (Proposition 13). This identifier is You can associate externally with a name or a file whose content is a file with the specific visual or auditory image of the image stored in the memory or with the identifier of another image in another associative memory either of the same or of a modality different (Proposition 32).

Recognition is the verification that a concrete or abstract image is contained in memory; For this purpose the operation of diagrammatic reduction is used, where the associative memory is the function and the image to be recognized is the argument of the reduction respectively. The result of the operation is the image directly recognized, as well as its identifier, which can be associated with the identifier or the specific information associated with the image when it was originally registered (eg, the name of the object in the image or the file with the photo or video of the stored image) (Proposition 33).

Recovery is the retribution of an image associated with a reference image (concrete or abstract) or "track" {tried or clue). In particular, the preferred association is the image that corresponds to the more specific abstraction it contains than the reference image. The recovered image can in turn be used as a track to recover another more general image. For example, the image of a hand can be the clue to recover the image of an arm, which in turn can associate with the image of a person, etc. The output of an association operation is the identification of the recovered image, which can be associated with the identifier or the specific information associated with the image when it was originally registered (eg, the name of the object in the image or the file with the photo or video of the stored image), and can be used as a clue for a new association (Proposition 34). The associative motive memory of the present invention consists in a reticular arrangement of processing units in which each cell is a processor capable of performing a minimum set of instructions. The contents in this arrangement are represented as attribute attribute structures, where each column corresponds to an attribute of the represented object and each line to the value that said attribute has for the corresponding object.

The memory register operation consists in applying the diagrammatic abstraction operation to the current content of the memory register with the content to be stored, and the result of said operation is written directly to the same memory register. The recognition operation consists in applying the diagrammatic reduction operation between the memory register and the readable content to be recognized. The recovery operation corresponds to the location of the content descriptor to be associated in the descriptor hierarchy that corresponds to the abstraction stored in memory, and in the recovery of the associated descriptors in the abstraction hierarchy.

The memory is illustrated in Figure 1. The system contains four arrays of vertically aligned two-dimensional cells; Each cell of each array is directly connected to adjacent cells as indicated by the dashed lines in Figure 1. That is, the cells of the upper array (Auxiliary Register for function conversions to abstractions, in black) are fed directly to the cells corresponding to the adjacent lower arrangement (Entry / Exit Register, in green). In turn, the cells of this last register are connected to the corresponding cells of the Memory Register (in blue), which corresponds directly to the associative memory. Finally, the Work Register cells (in red) to the bottom are fed by the corresponding cells of the Memory Register, as well as the cells of the Entry / Exit Register directly. For notation purposes, the Auxiliary Registry is called RA, the RI / O output input register, the RM Memory Register and the RT Labor Register. The following components are also shown in Figure 1: (1) Serial Input Record: This register is fed with functions or descriptors of concrete and / or abstract images that are directly abstracted in RA prior to initialization of the latter.

(2) Parallel Input / Output Control Unit: This register is an interface for writing and / or reading concrete or abstract images directly as two-dimensional arrangements in the RI / O register.

(3) Abstraction Identification Processing and Memory Unit. This record contains a memory with the identifiers of all the abstractions (concrete and abstract images) that have been added to the memory, because they have participated as inputs or outputs in the registration and recognition operations, as well as an associated time stamp to each of these identifiers. The output of this module consists of the identifier of the abstraction stored in a registration operation with its time stamp, or of the recognized image or the associated image in the recognition and recovery operations respectively, also with its time stamps. (4) Entropy Processing Unit: This module calculates and displays the entropy of the images whose identifier is stores in (3) and produces the value of the RM entropy at the end of each registration and recognition operation, as well as the difference between RM and RT at the end of a recognition operation (or the RM entropy difference caused by the recognition in case RT is dispensed with).

(5) Control Unit. This module executes the program of control of the operations or steps of registration, recognition and retrieval of information, and sends the corresponding signals to the modules of the architecture involved in each of these operations.

The information flow in Figure 1 is indicated by the thick blue arrows, and the control flow by the thin ones (Proposition 35).

The Serial Entry Register in Figure 1 has the function of receiving individual descriptors, abstracting them in the Auxiliary Registry to create the representation of concrete and abstract images through the operation of diagrammatic abstraction, as defined in Definitions 1 and 5 (Propositions 16 and 26). This register reads one function or attribute-value relationship at the same time and abstracts it in the Auxiliary Register, prior initialization in Bs, until a representation of a concrete or abstract image in the Auxiliary Registration Once the input image is completed in the auxiliary register it is transferred to the input / output register to perform the registration operation as follows:

Where

It is interpreted as the assignment of the term on the right to that on the left. (Proposition 36) The Parallel Input and Output Control Unit reads or writes a concrete or abstract image at the same time and abstracts it, prior initialization with Bs, in the Input / Output Register, as follows (Proposition 37 ):

The registration operation consists of abstracting the associative memory (the function of abstraction) with the contents of the Entry and Exit Register (the argument of the abstraction) and writing the result of the operation directly into the Memory Register. This operation is carried out in parallel, as follows:

As a result of the operation, the image identifier or saved abstraction is written as output of the image identifier processor with its corresponding time stamp.

The difference of RM entropies before and after the recording operation is made available as an output of the entropy processor (Proposition 38).

The recognition consists in verifying that an image in RI / 0 (i, j), constructed from its descriptors or directly, as described in Proposition 33, is contained in RM. This operation is performed by means of the operation of reduction of the RM content with the content of RI / O, as the function and the argument of the reduction respectively. If the reduction is applicable (Clause 1, Definition 4, Proposition 20, or Clause 1, Definition 6, Proposition 27): for all

 j columns.

Consequently,

And the recognition is successful. On the other hand, if the reduction is not applicable

So the reference image is not in memory.

The identifier of the image with its respective

Time stamp is registered in the memory of image identifiers.

Finally, the image identifier recognized in the RI / O register is also made available as an output of the image identifier processor, together with the time stamp of the original register, as well as the entropy difference between RM and RT, which is enable as output of the entropy processor (Proposition 39).

Unlike random access memories of the ordinary type that have an associated address in which the contents are stored, in the associative memory system described here there are no particular memory records with associated addresses in advance (Random Access Memory) to store the contents but on the contrary, the contents are stored distributed in a single register (RM) so that each content has an identifier associated that can be subject to registration or manipulation by a processor or processing architecture of the ordinary external type through the interface. It is possible to implement an alternative architecture without the RT job log. In this case the operation the result of the reduction writes directly to the RM memory register. However, since this operation is destructive and not reversible, it is required to complete the operation with memory abstraction operation with the registration again. This operation ensures that the representation of the recognized object is reinforced in memory, regardless of whether the final state of the memory register can be a sub-specification of the original state. For this case, recognition is defined as follows:

for all the

 j columns.

for all the

In case (1) is not applicable, for

all i. j.

Finally, the identifier is registered in the

Image identifier memory.

In case of using the alternative architecture without RT, the difference of RM entropies is calculated and made available before and after the recognition operation (Proposition 40).

The recovery operation is defined as the selection of an image in memory associated with the input image in RI / O. Because the RM register contains all the images registered in the memory in an over-placed way (ie "confused"), these cannot be recovered, so it is necessary to register their IDs individually in the corresponding memory (Proposition 35 , module (3)) at the time of registration.

This memory further guarantees that although it is possible for an image to degrade in memory due to a reduction operation, all the images that were entered into the memory are symbolically registered by their descriptors. Likewise, the time stamp associated with each image allows the most outgoing images to be identified recent, and this property can be used in specific association mechanisms.

The association operation itself is carried out by comparing the identifier of the image "track" {test or cJue) with the identifiers of all stored images, taking into account the time stamp of the same, with a specific algorithm for the type of desired association, which is also defined in the processing and memory module of abstractions. A generic association, for example, is to give back the less specific image that subsumes to the "track" image. This operation is verified directly from the analysis of both identifiers, since the conjunction of the "track" image must be contained within the conjunction of the image that subsumes it, for all columns.

Also, because the process is a symbolic comparison process, the recovery operation is independent of entropy and does not affect the memory status. However, because the association can be remembered, the operation is concluded with the abstraction of the remunerated image (which is rewritten in RI / O) with the FM memory and recorded again in the memory of abstractions, with a stamp of updated time. This operation reinforces the memory of the association and guarantees that it is present in the current state. of RM memory. Finally, the identifier of the abstraction resulting from this last operation is registered in the memory of image identifiers, as well as the difference of entropies in RM before and after this last recording (Proposition 41).

Modalities of the invention

 The associative memory system illustrated in Figure 1 and described in the present application can be simulated on a computer of the ordinary type with respect to the size of the array (ai, n) for specific applications, and is used as a module of a final application that it requires the storage, recognition or recovery of images or abstractions in one or more modalities. For this purpose, the system can be used in three different ways (Proposition 42):

(1) Using a memory to store a set of superimposed contents. (2) Using a parallel array of memories and storing only one content in each memory of the array.

(3) Using combinations of (1) and (2). The memory system can also be implemented in an integrated circuit to be used as a module of a processing architecture; In this mode a parameter is defined to specify the number of memory units, including a standard memory to specify the cardinality of the domain and range of each unit, as well as an additional parameter to indicate if the corresponding memory unit is used to store a content only or a set of contents {Proposition 43).

The memory system can also be implemented as a module of a general purpose process unit encapsulated in the same integrated circuit of said CPU; In this mode a parameter is defined to specify the number of memory units, including a standard memory to specify the cardinality of the domain and range of each unit, as well as an additional parameter to indicate if the corresponding memory unit is used to store a only content or a set of contents (Proposition 44). In any of the modalities listed, the memory can be implemented with or without the work register (Proposition 45). REFERENCES

 1. Anderson J. R., Bower, G. H., Human Associative Memory: A Brief Edition, Lawrence Erlbaum Associates, Punlishers, Hillsdale, New Jersey, 1980.

 2. Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities, Proceedings of the National Academy of Sciences of the USA, vol. 79 no. 8 pp. 2554-2558, April 1982.

 3. Lowe, D. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60 (2): 91-110, 2004.

Claims

 Having described the invention, it is considered as novel and inventive and the content of the following claims is claimed as property:

1. A registration system for storage and recognition of readable content by a processing architecture comprising associative memory in vertically aligned two-dimensional cell arrays characterized in that it comprises:

• An RA auxiliary register where abstract contents are constructed from basic descriptors;

 • An RI / 0 input / output register where the contents or descriptors of concrete and / or abstract images that are recorded in associative memory are written and recognition and recovery are encoded;

 • A memory register (RM);

 • An RT work record where the results of the reduction are written;

 • A serial input register where functions and / or descriptors of specific images that are directly abstracted in RA are registered;

• A parallel ent / salt control unit where concrete or abstract images are written and / or read directly as two-dimensional arrangements in the RI / O register;

 • An image memory processor (IDs) that contains a memory with the identifiers of all the abstractions that have been added to the memory as well as a time stamp associated with each of these identifiers;

 • An entropy processor in the memory register where the entropy value of the images in the RM memory register is calculated and displayed;

• A control unit where the registration, recognition and recovery steps are executed, and sends the corresponding signals to the modules of the architecture involved in each of the operations. 2. The registration system for storage and recognition of readable content by a processing architecture according to claim 1, characterized in that the contents are stored distributed in a single register (RM) such that each content has an associated identifier that can be registered or manipulated by a processor of the ordinary external type through the interface.

3. The registration system for storage and recognition of readable content by a processing architecture according to claim 1, characterized in that the work register does not exist.

Four . The registration system for storage and recognition of a content readable by a processing architecture according to claim 1, characterized in that the abstract contents that are constructed in the auxiliary register are the contents that are registered, recognized or associated in the registration of RM memory via the RI / 0 input / output register.

5. The registration system for storage and recognition of readable content by a processing architecture in accordance with claim 1, characterized in that the entropy value is available or is deployed to external devices through an interface.

6. The registration system for storage and recognition of readable content by a processing architecture according to claim 1, characterized in that the Abstraction identifiers are made available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a process and memory unit of the ordinary type, which is connected to associative memory through said inferphase.

7. The registration system for storage and recognition of readable content by a processing architecture according to claim 1, further characterized in that the registration, recognition and retrieval of abstract contents (memory states) are made available in a device external through an interface.

8. A registration system for storage and recognition of content readable by a processing architecture that when implemented as a module in a process unit performs the following steps:

• Registration of a state of memory, by means of a diagrammatic abstraction, which comprises a content to be stored; • Recognition, by means of a diagrammatic reduction, between the memory register and the content to be recognized;

 • Recovery of a content descriptor that corresponds to the stored abstraction.

9. The registration system for storage and recognition of readable content by a processing architecture of claim 8, further characterized in that each memory state and each content to be stored have a unique identifier or descriptor associated.

10. The registration system for storage and recognition of content readable by a processing architecture of claim 8, further characterized in that the registration, recognition and association are made available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a unit of process and memory of the ordinary type, which is connected to associative memory through said interface. 11. An associative memory system characterized in that it comprises four two-dimensional cell arrays vertically aligned, where the cells of each arrangement are directly connected with the adjacent cells and where the cells of the upper auxiliary register (RA) arrangement are fed directly to the corresponding cells of the adjacent lower arrangement

(RI / 0 input / output register), the latter cells are connected to the corresponding memory register (RM) cells, and where the work register (RT) cells are fed from the corresponding Registry cells of memory, as well as the input / output register cells directly.

12. The system of claim 11 characterized in that the memory register (RM) corresponds to associative memory directly.

13. The system of claim 11 characterized in that the associative memory keeps abstractions of the contents that are recorded.

14. The system of claim 11, characterized in that the array of two-dimensional cells further comprises:

• An RA auxiliary register where abstract contents are constructed from basic descriptors;

· An RI / 0 input / output register where the contents or descriptors of specific images and / or are written ● abstracts that are recorded in associative memory and recognition and recovery are encoded;

● A memory register (RM);

 An RT work record where the results of the reduction are written; ● A serial input register where functions and / or descriptors of specific images that are directly abstracted in RA are registered;

● A parallel ent / salt control unit where concrete or abstract images are written and / or read directly as two-dimensional arrangements in the RI / O register; ● An image memory processor (IDs) that contains a memory with the identifiers of all the abstractions that have been added to the memory as well as a time stamp associated with each of these identifiers; ● An entropy processor in the memory register where the entropy value of the images in the RM memory register is calculated and displayed; ● A control unit where the registration, recognition and recovery steps are executed, and sends the corresponding signals to the modules of the architecture involved in each of the operations.

15. The system of claim 14, characterized in that the work register does not exist.

16. The system of claim 14, characterized in that the abstract contents that are constructed in the auxiliary register are the contents that are registered, recognized or associated in the RM memory register via the RI / 0 input / output register.

17. The system of claim 14, characterized in that the value of the entropy is available or is deployed to external devices through an interface.

8. The system of claim 14, characterized in that the identifiers of the abstractions are made available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a unit of process and memory of the ordinary type, which is connected to associative memory through said interface.

9. The system of claim 14, further characterized in that the registration, recognition and recovery of the abstracts (memory states) are made available in an external device through an interface.

20. An associative memory system that when implemented as a module in a process unit performs the following steps:

 • Registration of a state of memory, by means of a diagrammatic abstraction, which comprises a content to be stored;

 • Recognition, by means of a diagrammatic reduction, between the memory register and the content to be recognized;

 • Recovery of a content descriptor that corresponds to the stored abstraction.

 21. The associative memory system of claim 20, further characterized in that each memory state and each content to be stored have an associated unique identifier or descriptor.

22. The associative memory system of claim 20, further characterized in that the registration, recognition and association are made available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a unit of process and memory of the ordinary type, which is connected to associative memory through said interface.

23. A method for generating an abstraction operation in an associative memory system comprising the steps of:

 • Memory register comprising a content to be stored;

 • Recognition between the memory register and the content to be recognized;

 • Recovery or association of a content descriptor that corresponds to the stored abstraction

4. The method of claim 23, further characterized in that each memory state and each content to be stored has an associated unique identifier or descriptor.

5. The method of claim 23, further characterized in that the registration, recognition and association are made available to an external device, through an interface, and allow the construction of the hierarchy of identifiers, as well as the identification of associations, in a process and memory unit of the ordinary type, which is connected to associative memory through said interface.

26. The use of the registration system for storage and recognition of readable content by a processing architecture comprising an associative memory for the registration, recognition and recovery of memory states or contents.