CA1232359A - Probabilistic learning element - Google Patents

Probabilistic learning element

Info

Publication number
CA1232359A
CA1232359A CA000472108A CA472108A CA1232359A CA 1232359 A CA1232359 A CA 1232359A CA 000472108 A CA000472108 A CA 000472108A CA 472108 A CA472108 A CA 472108A CA 1232359 A CA1232359 A CA 1232359A
Authority
CA
Canada
Prior art keywords
states
state
objects
sequence
sequences
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired
Application number
CA000472108A
Other languages
French (fr)
Inventor
Thomas B. Slack
Jeffrey N. Denenberg
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
International Standard Electric Corp
Original Assignee
International Standard Electric Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by International Standard Electric Corp filed Critical International Standard Electric Corp
Application granted granted Critical
Publication of CA1232359A publication Critical patent/CA1232359A/en
Expired legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/29Graphical models, e.g. Bayesian networks
    • G06F18/295Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/84Arrangements for image or video recognition or understanding using pattern recognition or machine learning using probabilistic graphical models from image or video features, e.g. Markov models or Bayesian networks
    • G06V10/85Markov-related models; Markov random fields

Abstract

ABSTRACT OF THE DISCLOSURE

A probabilistic learning element particularly adapted for use as a task independent sequential pattern recognition device receives sequences of objects and outputs sequences of recognized states composed of objects and includes a plurality of memories for storing the received sequences of objects and previously learned states as well as predetermined types of knowledge relating to previously learned states. The sequences of received objects are correlated with the information relating to the previously learned states in order to assign probabilities to possible next states in the sequence of recognized states.
Based upon the probabilities of the possible next states the most likely next state is determined and outputted as a recognized next state in the recognized state sequence when the element determines that a state has ended. The element additionally includes means for providing a rating of confidence in the recognized next state. The ratings of confidence for a sequence of recognized stated are accumulated and if the accumulated value exceeds a predetermined threshold level the element will be caused to store the recognized state sequence as a learned state sequence.

Description

lZ~23~9 T. B. SLA~K ET AL 1-6 PROBABILISTIC LEARNING ELEMENT

BACKGROUND OF THE INVENTION

Field of the Invention The present invention relates generally to recognition systems and more particularly to trainable or learning systems, which are capable of modifying their own in~ernal processin,g in response to information descriptive of system performance.

Description of the Prior Art Recognition systems of the type that recognize patterns deal with the problem of relating a set of or sequence of input objects or situations to what is already known by the system.
This operation is a necessary part of any intelliqent ~ystem '' since such a system must relate its current input and input environment to what it has experienced in order to respond appropriately.
A pattern recognitiOD task is typically divided into three steps: data acquisition, feature extraction, and pattern classification. The data acquisition step is performed by a transducer which converts measurements of the pattern into digital signals appropriate for a recognition system. In the feature extraction step these signals are converted into a set of features or attributes which are useful for discriminating the patterns relevant to the purposes of the recognizer. In the final step of pattern classification these features are matched to the features of the classes known by the system to decide which class best explains the input pattern.

~ 3~ . B. SLACK ET AL 1-6 , ~ 2 -The division between the step of Eeature extraction and pattern classification is somewhat arbitrary. A powerful feature extractor would make the classifier's job trivial and conversely, a powerful decision mechanism in the classiEier would perform well even with simple features. However in practice, feature extractors tend to be more task dependent.
For example, data ac~uisition and feature ex~raction for hand-printed character recognition will differ from that needed for speech recognition. Pattern classification on the other hand can be designed to be task independent, although it often is not.
A particular category of pattern recognition tasks is characterized by whether or not the features can be reduced to a linear sequence of input objects for the classificatio~ step.
This category is called sequential pattern recognition.
Examples of tasks which naturally fall into this category are optical character recognition, waveform recognition, and speech recognition. Other tasks such as computer image recognition can be placed within sequential pattern recognition by an appropriate ordering of the features.
Patterns of features must be acquired by the pattern recognizer Por a new c~lass of features before the system can recognize the class. When patterns cannot be learned from examples, acquisition of the patterns is a major problem.
Prior art optical character and speech recogni~ion systems correlate input patterns with a set of templates, in order to determine a "best match". A correlation is performed using a particular algorithm which is sp~cifically derived for the matching operation required for a particular problem such as speech recognition, character recognition, etc..O A change in type font or speaker, for example, would require replacing the templates and changing parameters of the alqorithm in such prior art systems.
Many trainable systems exist in the prior art, of which the following U. S. Patents are descriptive. U. S. Patent No.
3,950,733, an Information Processing System, illustrates an adaptive information processing system in which the learning growth rate is exponential rather than linear. U. S. Patent No. 3,715,730, a Multi-criteria Search Procedure for Trainable Processors illustrates a system having an expanded search capability in which trained responses to input signals are ~9 T. s. SLACK ET AL 1-6 produced in accordance with predetermined criteria. U. S.
Patent No. 3,702,986, a Trainable Entropy System illustrates a series of trainable non-linear processors in cascadP. U. S.
Patent No. 3,700,866, a Synthesized Cascaded Processor System illustrates a system in which a series of trainable processors generate a probabilistic signal for the next processor i-n the cascade which is a best estimate for that processor of a desired response. U. S. Patent No. 3,638,196 and 3,601,811, Learning Machines, illustrate the addition of hysteresis to perceptron-like systems. U. S. Patent No. 3,701,974, Learning Circuit, illustrates a typical learning element of the prior art. UO S.
Patent 3,613,084, Trainable Digital Apparatus illustrates a deterministic synethesized boolean function. U. S. Patent No.
3,623,015, Statistical Pattern Recognition System With Continual Update of Acceptance Zone Limits, illustrates a pattern recognition system capable of detecting similaEities between patterns on a statistical basis. U. SO Patents No.
3,999,161 and 4,066,999 relate to statistical character recognition sys~ems havinq learning capabilities Other patents that deal with learning systems that appear to be adaptive based upon probability or statis~ical experience include U. S. Patents No. 3,725,875; 3,576,976; 3,678,461;
3,440,617 and 3,414,885. Patents showing logic circuits that may be used in the above systems include U. S. Patents No.
3,566,359; 3,562,502; 3,44~,950; 3,1~3,648; 3,646,329;
3,753,243; 3,772,6S8; and 3,934,231.
Adaptive pattern, speech or character recognition systems are shown in the following U. S. Patents No. 4,318,083;
4,189,779; 3,581,281; 3,588,823; 3,196,399~ 4,100,370; and 3,457,552. U. S~ Patent No. 3,988,715 describes a system that develops conditional probabilities character by character with the highest probability being selected as the most probable interpretation of an optically scanned word. ~. S. Patent No.
3,267,431 describes a system that uses a "perceptron", a weighted co~relation network, that is trained on samplepatterns for identification of other patterns.
Articles and publications relating to the subject matter of the invention include the following: Introduction To Artifical ~ g~ , PO C. Jackson Jr., Petrocelli/Charter, .
N. Y. 1974 pages 368 381; "Artifical Intelligence", S. K.
Roberts, Byte, Vol. 6, No. 9, September 1981, pages 164-178;

f~3~3~9 T. B. SLACK ET AL 1-6 "How Artificial Is Intelligence?", W. R. Bennett Jr., American Scientist, Vol. 65, Nov~mber-December 1977, pages 694-702, and "Machine Intelligence and Communications In Future NASA
Missions", T. J. Healy, IEEE Communications Magazine, Vol. l9, No. 6, November 1981, pages 8-15.

SUMMARY OF THE INVENTION

The present invention provides a probabilistic learning system ~PLS) which performs the task independent pattern classification step for sequential pattern recognition systems ~0 and which acquires pattern descriptions of classes by learning from example. Thus, the PLS of the present invention is an adaptive or trainable learning system. Although a PLS could be applied to the problem of selecting good features for the feature extraction step that application will not be described here.
The PLS may comprise a plurality of probabilistic learning elements (PLE's) configured in an array or could be an individual PLE depending upon overall system requirements.
Each PLE is a element operating in accordance with its own set of multi-dimensional databases which are "learned" or altered through feedback from the environment in which it operates.
The array or the PLE has as its input a sequence of objects containing information, such as pixels, characters, speech or digital input from the environment. This information is processed as it passes throuqh the array or the PLE, thereby generating an output which may be either extr~cted knowledge in the form of an output state, such as a recognized pattern, or a set of control signals to be fed back for use as a future input modification, i.e. a process control adaptive equalizer.
The invention includes control mechanisms to provide performance feedback information to the array or the PLE. This information is used locally by each PLE of the array to modify its own databases for more appropriate behavior. Such performance feedback information can be supplied either to the entire array (globally) or to selected positions of the array, l.e one row, column or to the PLE involved in the generation of a particular output.
It is a primary objective of the present invention to utilize, in each individual PLE four interacting, but ~3~3~a9 T. B. SLACK ET AL 1-6 independent processing modules. An input module receives and stores input object sequence information. The input module provides two outputs. Firstly, a set of most probable output states that would end at the present time and their probabilities. Secondly, the probability that some state ends at the present time. A predict module receives and stores information on ou~put state options including state and length information. The predict module provides information on probable state length outputs. A decide module is responsive to the first output of the input modulel the output of the predict module and previous state options to derive a current list of state options. An output module receives the list of state options and the second output of the input module to choose the best state option which is outputted along with a confidence factor signal. ~hen the confidence factor signal exceeds a predetermined threshold value, the databases in both, the input and predict modules are updated with the new valid data.
The data stored concerning the input objects and output states includes several types of knowledge extracted from the actual input objects and output states. Sats of extracted knowledge are stored and correlated in the modules using various method~ of association depending upon the type of knowledge included in the particular set. The membership function of each set is learned using the adaptive process of the PLE.
The types of knowledge extracted and stored include:
frequency of objects and sequences of objects; position and positional frequency of objects and sequence of objects within states; state-lengths and state frequencies.
The PLE uses context driven searching in context organized memories to maintain a high throughput from tha large database.
Efficient searching is facilitated by organizing the inputted ohjects and the various types of extracted intelligence in context.
When a plurality of PLE's are used in an array to form a PLS parallelism may be employed to speed up task execution.
When using an array the size of the individual PLE's may be reduced as opposed to that requixed to do a complete task. The overall task is broken down into subtasks each accomplished by single PLE's or combinations of PLE'sO

r . B. SLAC~ ET AL 1-6 In order to maintain the general purpose nature of the PLS and its use for ~ide applicability the representation step for specific tasks is accomplished in an input preprocessor rather than in the array itself.

- The invention may be summarized according to a firsa broad aspect as a probabilistic learning element, that sequentially receives objects and outputs sequences of recognized states, said learning element comprising: means for sequentially receiving objects; means for storing, said received objects, sequences of received objects, previously learned sequences of states, states contained in said previously learned sequences of states9 and predetermined types of knowledge relating to, said previously learned sequences of states, said states contained in said previously learned sequeDces o~
states, objects contained in said previously learned sequences of states, and sequences of objects contained in said previously learned sequences of states, whereby current object information relating to said received objects and said sequences of received objects is stored as well as statistical information relating to previously learned sequences of states and said states, objects and sequences of objects contained in said previously learned sequences of states; means for correlating said stored current object information with said stored statistical information for assigning probabili~ies to possible next states in the sequence of recognized states;
means, responsive to said probabilities of possible next states, for determining a most likely next state;
means, responsive to the stored current object information and statistical information, for providing a signal corresponding to the probability that a state ,~ ~

~ 3~3 T.B. SLACK ET AL 1-6 - 6a -has ended; and means, responsive to said end of state signal, for outputting said most likely next state as a recognized next state in a recognized state sequence.

According to another broad aspect a method of recognizing a sequence of states from a sequence of inputted objects utilizing a probabilistic learning element, comprising the steps of: correlating said sequence o~ inputted objects and predetermined types of knol~ledge relating to said sequence of inputted objects with stored information relating to p~eviously learned states, objects and sequences of objects and predetermined types of knowledge relating to said previously learned states, objects and sequences of objects including Lhe number of occurrences of each;
providing probabilities for possible next states in a sequence of s~ates based on said correlations;
determining a most likely next state in the se~uence of states each time a new object is received; deriving from stored in~ormation a signal corresponding ~o the probability that a sta~e has ended; and ou~putting the most likely next state as a recognized nex~ state in the sequence of recognized states in response to the end of state probability.

BRI~F DESCRIPTION OF T~E DRA~INGS

Figure 1 is a simplified block diagram of a PLS
including a:n array of PLE's in accordance with the present invention.

Figure 2 is a diagram showing the recognition function o a PLE.

Figure 3 is an example of a context organized memory.

~.

- 6b lZ 3~

~ igure 4 illustrates the probabili~y computating process in a PLE and also illustrates the relationship of the major subroutines of a PLE with respect ~o probability computations.

Figure 5 is a simplified functional block diagram of a PLE in accordance with the present in~ention.

Figure 6 is a block diagram of the input module shown in Figure 5.

Figure 7 is a block diagram of the end time state length function shown in Figure 6.

Figure 8 is a block diagram of the span-length function module shown in Figure 6.

Figure 9 is a block diagram of the length normalizer shown in Figure 6.

Figure 10 is a block diagram of the predict module shown in Figure 5.

Figure 11 is a block diagram of the decide module shown in Figure 5.

Figure 12 is a block diagram of the output module shown in Figure 5.

Figure 13 illustrates the use of a PLS in a character recognition system.

Figure 14 illustrates a script reading system using a plurality of PLS's.

DESCRIPTION OF THE INVENTION

Prior to describing the invention it may be helpful to define certain terms used in the description.

;. ;~

~3~3~9 T. a . SLACR ET AL 1-6 Object: a feature extracted by the device immediately before the PLS and inputted to the PL5. An object may be a picture element ~pixel), a set of pixels, a ch~racter or even a word depending upon the application.
State: a recognized item outputted from the PLS such as a character, a word or a script depending upon the application.
Length: the number of objects in a stateO
State-Length Pair: a state and its length indexed and stored together.
Position: information which identifies that an inputted object is part of a state and where in the sequence oE all objects thak are in that state it occurs. Conversely, this same information identifies that a particular state was formed of a particular set of objects from the sequence sent by the feature extractor. Thus, the position of a state means both where in the input stream the state begins and where it ends.
The position of an object means how far from the beginning of the state and from the end of the state it occurred.
Confidence: a rating related to the probability that a particular state occurred in a particular position and the support coefficient of said probability. Confidence equals Probability* Support Coefficient.
Support Coefficient: a rating related to how much information was available when calculating a given probability.
It i5 possible to have a high probability, based on little information.
R0ferring to Figure l, there is shown a block diagram of a trainable PLS 10 formed of an array 12 of trainable PLE's constructed in accordance with the present invention. The PLS
includes an input 11 for receiving objects and an output 13 for outputting recognized states. The array 12 of PLE's 14a to 14h i5 configured as a two dimensional array of elements, each of which operates according to its own set of multi-dimensional databases. Values for the databases are obtained from the external environment and from outputs which comprise the results of the overall system operation. An output processor 16 also includes a feedback interface portion coupled to a bidirectional bus 18 which is adapted to feedback output information to a maintenance and human interface processor 20. -The interface processor 20 also intercouples data from an input preprocessor 22 and the array 12 on bidirectional busses 2g and 26 to provide ~3~5~ To B. SLACR ET AL 1-6 8 ~

feedback paths not only between the output processor 16 and ~he trainable array 12, but also between the input processor 22 and the trainable array 1~ via the maintenance and human interface processor 20~
Prior to discussing the operation of the PLS 10, which comprises an array 12 of PLE's 14a to 14h, it should be understood that a PLS may consist of only one PLE if it has sufficient capacity for the assigned task.
A PLE inputs a sequence of objects and outputs a sequence of states which it learned from past experience to be the state sequence most closely associated with the input object sequence.
The confidence which the PLE has in the association is indicated by assigning a ra~ing to the output state sequence. This recognition function is shown most clearly in Figure 2. The learning function could be illustrated by reversing the arrow now pointing to the output state sequence and ignoring the confidence rating.
In keeping with the task independent goal of the PLS there is no inherent meaning associated with an input object or an output state~ they are members of finite se~5. The input and output may in fact be the same set, but thia is transparent to the system. The number of unique objects and states appearing in the task does however effect the database size of each PLE.
Although much of PLE's processing power, generality, and speed can be attributed to statistical modeling of its environment and the organization of that model in memory, the basic idea behind the modeling is simple. A sequence of objects as shown in Figure 2 is learned and modeled by counting the n-grams of objects making up the sequence~ where an n-gram is defined simply as a subsequence of n objects. Thus after learning, the element knows how often each object (l-gram) appeared in any sequence for each state, how often each pair of objects (2-gram) appeared in any sequence for each state, and so forth up to a specified limit of n. If D is the number of different objects there can be as many as D to the power of n different n-grams. Howe~er, the number is limited by the realities of a pattern recognition task. The size of D is determined by the feature extractor and the number of unique n-grams is determined by the states being recognized. Typically a finite set of states uses only a small fraction of the pattern space (e.g., this is true in speech recognition and optical character recognition~.

3 5 ~ T. B. SLACK ~T AL 1-6 9 _ The identity and frequency of n-grams are stored in databases in a context organi2ed manner for long term memory.
We call the databases that are organized in thls manner Context Organized Memories or COM's. This type of memory s~orage is a S modified tree structure in which each node represents a particular n-gram and is the parent node of all (n+l)-gram nodes that share the same first n objects. In addition, each node is linked to an ~n-l)-gram node which represents the same object sequence with one less object at the beginning of the sequence.
Figure 3 gives an example of a COM in which the object n-grams are composed of letters for the word "MISSISSIPPI"o For "MISSISSIPPI" there are four objects i.e. S, I, M, P, therefore, D=4 and the highest level n-gram shown is a 3-gram lS for n=3. The objects on the path to a node at level n define the n-gram represented by the node. The number stored at the node is the frequency count of the n-gram. The dotted lines show links to the related (n-l)-grams. For example, the 3-gram "SIS" has occurred in the training once and it is linked to its unique 2-gram "IS".
The COM supports an efficient Context Driven Search. The memory arranges the objects so ~hat the set of objects which statistically occur next in context are directly accessible from the current point or node in the structure. If the next input object does not match any o~ those in the expected set, the next position searched in the structure correspands to the less specific context obtained conceptually by deleting the oldest object and algorithmically following the link to the (n~ gram node. At level n the greatest number of nodes expanded (i.e., searching all sons of a node) before the next object is found will be n. This corresponds to the case when the new object has never been found to follow any subpart of the current n-gram and the sea~ch most "drop all context" to find the object at level 1. An important feature of Context Driven Searching is that the average number of nodes expanded per input object is two. This is obvious if we remember that every failed node expansion (decreasing level by one) must be matched by some successful node expansion (increasing level by one) since the search remains within the finite levels of the tree.

~ 3~31~9 T. B. SLACK ET AL 1- 6 -- 10 ~

The data structure of a PLE consists of four ma~or long term databases and their supporting structures and a number of short term memories which will be discussed subsequently. The long term memories are COM's which may comprise one or more connected or associated tree structures. The COM's of the four major databases are indexed by object, state, length, or state-length.
The object database comprises a plurality of tree structures wherein the nodes of the trees each have an attribute list. One tree is called an alltree while the other trees of the object database are called singletrees. There axe singletrees for each previously learned state, i.eO each state has its o~n object indexed single~ree. Associated with each node of the alltree is an attribute list which acts as a pointer to all singletrees that include the same context as that of the alltree nodeO Thus, for every singletree there is a corresponding place in the alltree and that place in the alltree has an attribute list pointing to the place in the singletree.
Each node in the alltrees provides a record with these components:

(l) The additional object that this node represents~
(2) The frequency of occurrence of this pattern among the learned sequen~es of pattern~. This occurrence is based not only on the object in this node but on ~he pattern of the nodes above it that point down to it.
(3) A calculated value from (2) derived by taking the logarithm value of (2) then multiplying by a constant. This operation maps the integer occurrence values to the integer range of zero to a predefined upper bound.
(4) A calculated value from the nodes under the current node. It is a measure of the usefulness of the information represented by those nodes. This value is called the support coefficient and is initialized as -1 to indicate that no valid data stored here.
Each time a node is updated, its support coefficient is also reset to l to indicate that the support coeficient is not updated yet. This value is calculated when it is requested the first time after T~ B. SLACK ET AL 1-6 last update. The value is then stored in the node.
- And this value is valid till the next update.
(5) The pointer to the node which represents the same corr~lation information except the last object. And its object is greater than the object of this node.
(6) The pointer to the node which represents the same correlation information and with one more objectO
This is called one level deeper. There may be more than one such node. The one that down pointer points to is the one with the smallest object.
~7) The pointer to the node which represents the same pattern except it is one level higher. That is to say it does not have the oldest object.
~8) The pointer to the node which represents the same pattern without the last object. That node is also one level higher.

Singletrees are similar to alltrees in structure and purpose. The only difference in structure is in the attribute lists a~d the only difference in purpose is that an alltree contains pattern information indepeadent of output state recognized, and a singletree contains pattern information concerning a single output state. Thus, an alltree may contain the cumulative of several singletree~ In the described - embodiment we use one alltree to contain and control all the singletrees in the object database.
The entries of sin~letree attribute lists represent detailed correlation information for the state represented by the singletree for the node with which it is associated. It has four components:

(1) The number of objects (distance) in front of the object this node represents. This provides the distance Erom the beginning of the state to the current nodeO
t2) The number of objects (distance) from the object of this node to the end of the state.
(3) The number of times this object in this position has been learned.
(4) The calculated data from (3). The same calculation as done in (3) of alltrees.

3~ 3~ T. B. SLACK ET AL 1-6 The state, length and state-length databases each comprise one singletree structure indexed by state, length and state-length respectively. These singletrees do not have attribute lists as do the singletrees of the object database b~t the typP
of information stored at each node are similar to that stored in the object database tree.
When a COM is used to store the frequency of object n-grams forming parts of states the storage is efficient since only one tree is used for all states. An n-gram which appears in more than one state is stored once and details within the attribute list for the node list the proper states together ~ith separate frequency count~.
Learning the next object in a sequence is simply a matter of creating a new node in the tree whenever the object appears in a new context or incrementing a frequency count when it appears in a previously learned context~ The object is learned in all possible contexts from the (n-l) gram preceding it for some maximum n down to a null context in which the object is recorded by itself as a l-gram.
2~ The databases are arranged to store five different types of knowledge. The five types of knowledge that are modeled by the PLE and stored in COM's are as follows:

Type 1 The frequency of object n-grams forming parts of all possible states; this knowledge is stored at the nodes o the alltree.
Type 2 The position and positional frequency of object n-grams within states; this knowledge is stored in the singletree attribute lists of the object database.
Type 3 The frequency of n-grams composed of states ~i.e. for states T and A a 2-gram of states would be TA); this knowledge i5 stored in the nodes of the singletree of the state database.
Type 4 The frequency of n-grams composed of state lengths ~i.e., the lengths of the underlying object sequence for state lengths of 4, 3 and 5 a 3-gram of state lengths would be 435); this knowledge is stored in the nodes of the singletree of the length database.
Type 5 The frequency of n-grams composed of state-length pairs, which knowledge is stored at ~he nodes of the ~0 state-length database.

-~ 3~ To B. SLACK ET AL 1-6 Con~ider an object 4-gram, YlY2Y3Y4~ sto node j and let fj be the frequency of occurrence of the 4-gram and fi be the frequency of occurrence for its parent node, a 3-gram. Then the conditional probability that object y4 will occur in the context of YlY2Y3 is gi~en by the maximum likelyhood estimate:

P~Y4 I YlY2Y3) = ~
(1) This is the probabilistic basis for pattern matching in the PLE. The following types of conditional probabilistic knowledge maybe retrieved from the COM's using the above knowledge types:

Pl. The probability that object Yt will occur given the previous object context and state Xi, from the nodes of the singletree in the object database.
15 P2. The probability that object Yt will occur with beginning position f, ending position g, given previous object context with consistant positioning and state xi, from the singletree attribute lists in the object database.
P3 The probability that state xi will occur given previous output states, from the nodes of singletree in the state database~ -P4~ The probability of s~ate length Lj given lengths of previous output states, from the nodes of the singletree in the length database~
P5. The probability of state and length xi, Lj given previous sequence of state-length pairs, from the nodes of the singletree in the state length database.

These probabilities will be more formally defined and derived subsequently.
Note that the sequence of state-length pairs is given as much attention by PLE modeliny as the states themselves~ Thi~
was done to permit the PLE to extract all relevant information from its environment so that it could decide what was needed to perform a task. In some pattern recognition tasXs such as Morse code recognition or music classification the length of ~3t~3~ T. ~. SLACK ET AL 1-6 object sequences may be as important as the identity of the objects or the states. The PLE has the ability to use the information which is most h lpful Eor the recognition task being performed.
The databases also include short term memory capability for storing the five types of koowledge that have recen~ly been observed. rrhe recently observed knowledge is correlated with the five types of knowledge that have been experienced, modeled and stored in COMI 5 for long term memory in order to assign probabilities to possible output states. Short term memories build up and maintain the context in which the next object is handled. This saved context includes pointers into the trees of the COM's of long term memory.
Using the conditional probabilities retrieved from the COM's the following two basic probabilities are computed for all states and lengths previously stored in the COM's each time a new input object is received~

1. Input Probability: the probability that an input object sequence beginning at time b will-occur and span a state given that it will end at time t and that the state will occur;
2. Predict Probability: the probability that a state and length will occur given that a previous sequence of states and lengths have occurredO

Figure 4 shows the Input and Predict processes that compute these probabilities~ Since mathematical details will be given subsequently, only an overview about these processes is discussed here. Probabilistic knowledge type P2, in-troduced above as the conditional probability of an object and the object's position, would be sufficient by itself for obtaining the needed input probability if enough training data could be guaranteed. However, in real recognition tasks the knowledge is too specific to stand alone. For example, if n-gram 'bacb' at a position of 5 objects from the beginning and 2 objects from the end in state S is highly likely to occur if state S
occurs then it is likely to occur in other positions when state S occurs given any noise or uncertainty in the input. But if the n-gram occurs in some yet unobserved position, probahilistic knowledge type P2 will give no support for state S occurring J~ r~g T B SLACX ET AL 1-6 based on n-gram 'bacb'. Eor this reason n~gram frequencies for a state are learned independent of position as probabilistic knowledge type Pl. Probabilistic knowledge type P2 is used only to ~stimate the probability of a states be~inning time given an ending time, the intervening object sequence, and the state. Thus~ probabilistic knowledge type P2 s gment~ the input and probabilistic knowledge type Pl identifies it.
5imilarly, in the predict process probabilistic knowledge type P5 containing the probability that a state with a particular length (state-length pair) occurs given that a previous sequence of states and lengths have occurred is very specific and would require a large amount o training and memory to be useful by itself. However it does supply the relationship between states and lengths (e~g., if state S occurs i~ will have length L with probability p, or if length L' occurs it will be state S; with probability p'~. Probabilistic knowledge types P3 and P4 give predictions of state and length respectively based on ~ore context and are combined with probabilities from probabilistic knowledge type P5 to find the best predictor of each state length pair.
The two basic probabilities, Input and Predict, are used in the PLE decision process. Fxo~ the input and predict probabilities at each object input time, t, the decision process computes the probability that a state and a length and the input object sequence spanning the length ending at t will occur gi~en past context. These probabilities are combined over time using the Viterbi Algorithm to compute the k most likely state sequences ending at time t, for some k. The most likely states2quence ending at f inal time T is the recognized state sequence.
The foregoing discussions of the use of probabilities in the PLE will now be axpanded to include another important PLE
concept. In any human decision at least three factors come into play when assigning a confidence to the decision that is finally made:

1. How much do I trust my input information about the current circumstance?;
2. How well do the circumstance match the circumstance for previous decision experience I have had?; and 3. How much experience have I had and do I trust it~

3~ T. B. SLACK ET AL 1-6 The PLE attempts to use the last two factors to make a decision and to compute a rating of confidence in its decisio The PLE assumes ~hat the input object sequence is completely reliable and therefore does not use the first factor. It is understood that this con~traint may not always be trueO The second factor corresponds to the two basic correlation probabilities and the decision processO
The third factor is implemented by computing a 'coefficient of support' for every conditional probability obtained from the COM structures. ~or a particular condition (i.e., context) the coefficient measures how diverse the experience has been and varies between 0 and 1 as the support ranges between no support (experience shows random chance) to complete support (no other possible choice)~ In additionl the support coefficient measures the amoun~ of experience under each condition.
The support coefficients are combined together throughout the computation of the final probability to obtain an overall confidence rating for the probability of the recognized state sequence. The confidence rating is passed on to the in~erface processor 20 shown Figure l for PLS array or to a learning supervision circuit which decides whether or not to learn to associate the output state sequence wi~h the input object sequence. This decision i5 based on a threshold ~est of the confidence rating or external reinforcement. 2xternal reinforcement may be either from another PLE, as in an array, or from a human operator. The reinforcement may also include corrections to some of the state and bounda~y decisions made by the PLE. These corrections are passed on to the PLE databases before the COM's are updated by a command from the learning supervision circui~
This type of correlation of conditional probabilities der~ived from learned experience allows the PLS to be general purpose in nature. To preserve this general purpose nature of the PLS, part of the representation for each specific task will be in the input preprocessor 22 designed for that recognition task and shown in Fiyure 1. This will allow the PLS to be independent of special purpose aspects of application problems since often the representation are problem-specific.

~3~35~ T. B. SLACK ET AL 1-6 The following describeq in detail the computationsperformed by the PLE to assign the most likely sequence of states to a sequence of objects given the probabilistic knowledge stored in the COM's~ Referring to Figure 2 let Yl~Y2~ YT, or more compactly, y(l:T), be an input sequence of objects to the PLE during time units 1 through ~, and x(l:R) be the output sequence of recognized states. Since an output state is represented by one or more input objects, R
is less than or equal to T. Let b(l:R) be the mapping of input objects to output states such that br gives the time unit of the flrst object for state xr. Thus, l~b.<b <T Eor l<i~j~R
(23 The task of the PLE is to find R output states x(l:R) with boundaries in the input sequence of b(l:R) for a given input object sequence y(l:T) such that P(x(l:R),b(l:R)¦y(l:T)) is maximized~
By sayes' rule P(x(l:R),b(l:R)) * P(y(l:T)¦x(l:R),b(l:R)) P(x(loR),b(l:R)¦y(l:T)) = -P(y51:T) ) (3) But since y(l:T) is constant over any one recognition task we can obtain the same solution by maximizing the numerator:

P(y(l:T),x(l:R),b(l:R)) -P(x(l.R),b(l:R)) * P(y(l:T)¦x(l:R),b(l:R)).
(~) It is not computationally practical to compute equation 4 for all possible sets of LR, x(l:R), b(l:R)]. Therefore the restrictions of a more specific model are applied. The model used by the PLE is that the object sequences within a state, the sequences o states~ the sequences of state leng~hs, and the sequences of state-length pairs represent probabilistic functions of Markov processes.
Specifically it is assumed that:

~j3r3~ T. B. SLACK ET AL 1-6 ~ ~8 1. The conditionll probability that object Yt given y(t-c1 t-1) and state xr is independent of t, r, x(l:r~ x(r~l:R), and any other y's or some context level cl determined by the training ofthe PLE, 2. The conditional probability of state xr depends only on x(r-c2:r-1) for some context level c2;
3. The conditional probability of length Lr a br+l ~ br depends only on L(r~c3:r-1) for some context level C3;
and 0 4. The conditional probability of (xr,Lr) depends only on (x(r-c~:r-l),L(r-c4:r-1)) for some context level c~.

We are using what might be called variable order Markov processes since for each Markov STATE (i.e. object, output state, length or state-length pair) of these four Markov processes the context level c varies depending on training of the PLE. The brder of a Markov process is given by the number of previous STATES effecting the current STAT~ We will use "STATE" in bold type to differentiate a Markov STATE from an ouput state of the PLE. Now, an nth ordar Markov process is equivalent to some first order Markov process with an expanded STATE space. In fact, the learning process using the COM of the PLE maintains such a STATE expansion automatically. For example each node on the COM object tree can be viewed as representing a STATE of the ~arkov ~hain encoding a particular n-gram of objects. The transitions to all possible next Markov STATES is given by the links to all sons of the node. These sons encode (n~ gram of objects. New Markov STATES are added as new n-grams are observed and deleted as transitions to them become impxobable.
Given the above Markov assumptions, a simple method for finding the most likely set [R,x(l:R),b(l:R)] for a given y(l:T) is a dynamic programming scheme called the Vitexbi Algorithm.
Let W(t,k) be the kth most likely sequence of states x(l:r) that match input y(l:T) up to time t. Let G(t,k) a P(W(t,k)) = P(y(l:t),xk(l:r),b(l.r),br~l=t+1) (5) 3,r~3r~9 T. B~ SLACK ET AL 1-6 denote the probability of the kth best sequenceO The term br~lis included to make explicit that the last state ~r ends with object Yt~ The goal of the PLE is to determine W(T,l).
W(t,k) can be shown to be an extension of W(t'~k') for some tl<t and k'. Specifically, G(t,k) = kth-max [G(t~k~)~p(y(br t)~xrlbr~br~l=t~llx(l r ),b(l:r ))]
'Xr (6) where r' is the number of states in the klth best sequence ending at time t' and r = r'~l is the number of states in the kth best sequence ending at time t. Computing the best k sequences rather than only the best sequence to time t permits a non-optimal S(t,k) to be part of the final sequence S(T,l) if S~t,k) is suppor~ed by the context of,later states. Thus context learned in a forward direction as P(xrlx(r-c:r-l)) has an effect in a backward direction as P(xr_clx(r~c~l:r)).
Equation 6 is computed in the decision process using P(Y(br t)/Xr~br~br~l=t~l ¦ x~r-c:r-l),b(r-c:r-l)) =
P(y(br:t) ~brlbr+l=t+l~xr) 2~ P(xr,br~l=t~l¦x(r-c:r-l),b(r-c:r-l)).

where r-l has replaced r'. The left and right terms of the product are the Input and Predict probabilities respectively.
These were discussed and appear in Figure 4 as the outpu~ of the Input and Predict processes. HereinaEter we will discuss the computation of the Input and Predict probabilities, but first we will derive the support coefficient used in these computations.
In explaining the support coefficient we want to do four things- describe what a support coefficient is, show how it is used to compute a confidence rating, show how support coefficients are combined, and describe how support coefEicients permit the PLE to weight its knowledg0 according to its experience.
Let p(l:n) be the estimated probability vector for the occurrence of n mutually exclusive events. Therefore the n 3.~3~.a~ T. B. SLACK ET ~L 1-6 probabilities equal 1. The remaining uncertainty about what event will occur in terms of bits of information is given by ~he entropy measure:

~(p(l:n)) = - ~ pjlog pj (~) Let us assume that the probability-~ector gives an accura~e although perhaps incomplete description of reality concerning.the n events.
The fraction of information supplied by the probability vector of the total information needed to correctly predict .
what event will occur is given.by:

H(p~l:n)) S (P) = 1 - ~ (g) log n We call this fraction the support coefficient since it measuras the amount of support given by the probability ~ector towards making a decision. The support coefficient is 1 when the probability of one of the events is 1. Its value is b when the probability of all events are equal.
Let Pi j represent the probability of event ~ obtained by some estimator i and Si be the support coefficient for ~he estimator computed from the probability vector it produced~ We use Pi j *Si.as a measure of confidence in a decision that - chooses event.j as the one that ~ill occur. The PLE uses th.is measure of confidence in four areas:

1. To chose between various conditional probabilities assigned to an object based on different levels of context;
2. To chose between state-based and length-based estimates of the predict probabilities;
3. To chose in the decision processes the kth best state sequence ending at time t in equation 60 Thus equation 6 should be amended with -- 'where the kth max is determined by' kth-max [G(t',k') * P(...) ~ S(G(t',k') * P(...))].
(10) -~ 3' ~. B. 5LACK ET AL 1~6 And to indicate to the learning supervisor the confidence it has in the final decision.

Tha combining o~ support coefficients from different probability vectors to obtain the support coefficient for their joint distribution is quite simple. It can be shown that foY
probability vectors p and q each of length n: -H(p*q) = H(p) -~ H(q)~ (11) From which it follows that:

S(p*q) = (S(p) * S(q)) / 20 Extending this to more than two vectors gives a means by which the PLE can assign a support coefficient to the probability of a sequence of objects (or states) by averaging the support coefficients of the individual conditional probabilities of the objects (or states) making up the sequence~
A weakness of support coefficients as described to this point is that they do not measure the amount of experience behind the estimated probability vectors. ~or e~ample, a probability vector of [0.8,0~2,0,0] for ~our events has a support coefficient of 0.64 according to equation 9, which does not distinguish whether the probabilistic estimate is bas~d on frequency counts of [~,1,0,0] or of [100,25,0,0]. Ideally the support coefficient would be higher in the second case than in the firRt. We will modify equations 8 and 9 to first put them in terms of frequency counts and then to incorporate the 5 concept of experience.
The COM structures store frequency count~s at each node for the object n-gram it represents (or state, length, or, state-length pair n-grams -- depending on the knowledge type). The conditional probability, pj of the node is simply the ratio of the frequency~count, fj, of the node to the frequency count, f'i, of its parent node. Thus, Pi fj/f i fj/ ~ fk (13) 3~ 3r39 T. ~. SLACK ET AL 1-6 wher~ the sum i~ over all sons of the parent node. Substituting equation 13 into equation 8, combining it with equation 9 and simplifying yields:

lo~ f'i ~ fj 1 g S (p) = 1 -log N (14) where N is the number of possible nodes (e~g., equal to the number of unique objects) and n is the number of existing nodes.
We can now incorporate the concept of ~xperience by assuming that all non-existing nodes (objects not yet seen in the current context) occur with some finite frequency, u. The larger u is the greater the frequency counts of the existing nodes must be to achieve the same support coefficient. On the other hand, greater experience raises the support ooefficient even if it does not change the conditional probabilities.
Equation (14) now becomes log f~i ~ f~i ((N-n~ u log u ~ ~ fj log j) 15 S (p) a 1 ~
log N
(15) where the frequency count of the parent node,f'~, has been replaced by:

f i f ' i ~ u* (N~n) .
The value of u doe~ not have to be an integer~ In th~
2a example given above, if u i~ set to 0q5 the ~upport coe~ficient for the probabilities based on low frequency counts of [4,1,0,0] drop~ to 0029. The ~upport coefficient for the frequency counts of [1~,25,0~0] remains almo~t unchanged at 0.63.
~he Input probability i~ given by '~ 3~ D~ T. B. SLACR ET AL 1-6 p~y(br:t)~brlbr+l=t~l~xr) =
P(Y(br:t~ It~xr) ~ P(br¦~(bE:t) ~t,xr) (16) As summarized previously the first term call~d probabilistic knowledge type Pl identifies the input and the second term called probabilistic knowledge type P2 segments the input.
The first term is obtained from (Y(br t) It,Xr) = P~y(br:t) lxr~

= P(Yb lXr~ *b ~l<tP(yily(i-ci i-l),xr) where cb 0 and 0<j<cl 1+l (Yily(~ xr)*s(p)]

(17) Each P(yily(i-ci:i-l),xi) is a conditional probability obtained from the frequ~ncy counts stored at the nodes of the COM tree structure. The log of frequency counts are also stored at each node to psrmit efficient probability computations lS in the log domain. However this is an implementation issue and the math has been left in the non-log domain. The value ci determines equivalently: the level in the treel the context in which the objec~ Yi is matched, and the conditional probability used for the object. It is chosen to maximize the confidence value of the decision which as explained above is e~ual to the product o-the probability and the support coefficient. Equation (17) shows that the context level for the next object is limited to be no more than one level deeper than the best level of the current object.
The derivation of the second term containing the second probabilistic knowledge type will now be discussed. The frequency counts stored at a node in the COM object tree for a particular object n-gram are first divided between the states that were learned when the n gram appeared (knowledge type 1) and then further divided between the various positions in which the n-gram appeared within each state (knowledge type 2). The position is given by two values: the number of objects ~3~3~ T. B. SLACR ET AL 1-6 preceding the last object of the n-gram; and the number of objects following the the last object plus 1. We call these values the "distance to beginning" and the "distance to ending'~.
The sum of these values will always equal the length of the state (i.e., number of objects) at the time the pattern was learned.
Let fi, and gi be the distance to beginning and ending respectively for the last object Yi of n-grams appearing in patterns learned for state X. The probability that a object sequence y(b:e) is a complete pattern for state X (i.e., both begins and ends the state) is estimated by P(y(b:e)~b~e~lx) = (p(yb~fb-0~gb e P(yi~f~ b~gi=e-i~lly(i-ci~ f(i-ci:i~ g(i-cioi-l)~x)l/~
b<i<e (18) wher~ L=e-b~l and ci-takes on the same values o equation (17). The conditional probabilities returned by the tree are bounded below by a small po~itive value since in many cases there will be no learned examples of a particular n-gram, in a particular position, for a particular state. The effect of this "deficiency" in the training is removed by replacing zero probabilities with small probabilities, and normalizing by length by taken the Lth root of the product. These calculations take place in the log domain.
We can now compute the second term of equation ~16) with P(y(br:t) ,br,t¦Xr) P(br¦Y(br t) ~t~Xr) ~ P(y(i:t),i,t¦xr) br<i<t (19) 25The Predict probability (the second term of (7)) can be rewritten as P(xr,Lr¦x(r-c:r-l),L(r-c:r-l)) =
P(x~,br~l=t~l¦x(r-c:r-l~,b(r-c:r-l)) (2~) ~23Z.3S9 T. B. SLACK ET AL 1-6 where Lr-~r~l-br is the length of state xr. This probability can be computed based on state predictions from probabilistic knowledge type P3 as P(xr~Lrlx(r~c:r~ L(r-c r 1)~ =
P(Lrjxr,x(r-cl:r-l),L(r-c':r-l)) * P(xr ¦xr c:r 1) (21) .

or based on length predictions from probabilistic knowledge type P4 as P(xrlLrlx(r-c:r-l),L(r-c r-l)) =
P(xr¦Lr,x(r-c':r-l),L(r-c':r-l)) * P(Lr¦Lr c:r 1) (22) For each state and length pair the method is chosen to give the maximum con~idence value for the decision. The first term in each case is derived from the state-length COM tree structure probabilistic knowledge type P5 by ~umming over all lengths for a given state or all states for a given leng~h as appropriate.
In equation (22) the first term is derived from the equation:

- P(xi,Lj¦x~r-c':r-l),L(r-c' or-l) P(Lj¦xi,x(r-c':r-l),L(r-c':r-l)) =
(xi~Lilx(r-c~:r~ L(r-cl:r-l) (21A) In equation (21) the irst term is derived from the equation.
.

P(xi,Ljlx~r-c':r-l),L(r-c':r-l)) P(xi¦Lj,x(r-c':r l),L(r-c':r-l)) =
P(xi,Ljlx~r-c':r-l)!L~r-c':r-l)`
(22A) 3~ T. B. SLACK ET AL 1-6 - 26 ~

The context level c' of this tree is typically less than the conte~t levels of the stale and length prediction trees.
If c=c' there is no advantaye in combining in the state or length prediction information. In all trees the context level is chosen to maximize the confidenc2 values of the conditional probabilities.
The following is a description of a physical embodiment of a PLE constructed in accordance with the present invention.
Referring to Figure 5 there is shown a block diagram of a PLE
comprising four major modules, namely input module 28, predict module 30, decide module 32 and output module 34. A comparator 36 is also provided having one input connected to a ~ariable divider circuit 38. An OR-gate 40 is provided having one input connected to the output of the comparator 36 and a second input connected to receive a learn signal.
Input informationr in the form of objects enter the input module 28 at an input terminal 42. The input module uses the input objects to provide two kinds o~ probability information based on previously learned object sequences. At a terminal 44, the input module provides a signal PE corresponding to the probability that some state ends at the present time and this probability will be known as End-Of-State Probability. At a terminal 46y module 28 provid~s a signal PI correspo~ding to - the probability that an input object sequence b~ginning at a time b, will occur and span a state given that it will end at time t and that the state will occur. This probability~will be known as Input Probability PI and i5 derived using the previously discussed equation ~16).
The predict module 30 receives Options Information at an input 48 from the decide module 32 and uses this information in conjunction with other information to calculate the most likely state-length pairs and their probabilities~ which probability information is provided as signal Pp at an output 50. The state-length pair probability ioformation shall be known as Predict Probability and may be derived using the previously discussed equations (20)~ (21), and (22).
The decide module 32 includes a first input 52 for receiving the Input Probability sign~l PI from the input module 28 and a second input 54 for receiving the Predict Probability signal Pp from the predict module 30, The decide module combines the Input and the Predict Probabilities to form 3~ 3~ T. B. SLACK ET AL 1-6 ~ 27 -the previously mentioned options Information which is provided at terminal 56. The Options Information is derived using the previously discussed equation5 (5), (6~ and (7~ implementing the Viterbi Al~srithm.
The outpu~ module 34 includes ~wo inputs, 58 and 60 for receiving the End-Of-State Probability signal PE and the Options Information respectively. The output module accumulates the Options Information and utilizes the End-O~-State Probability to decide when the Options Information is sufficiently good to output a recognized state and its preceding states at a final time T at a terminal 62. The output module al~o provid~s an output at a terminal 64 corresponding to the probability that the state occured in a particular position and this signal is known as the Confidence Factor derived using equation (9) and the probability vector as previously discussed. The ou~put module provides one additionaloutput at a terminal 66 corresponding to the positions of the recognized state and the preceding states.
The recognized states are fed back ~o the input and predict modules at terminals 68 and 70 respectively while the position information is fed back to the input and predict modules at terminals 72 and 74 respectivelyO
The Confidence Factor is applied to a second input of the comparator 36 so that when the level of the Confidence Factor exceeds a threshold established by the divider 38 a self learn signal is provided from the comparator 36 to an input of the Or-gate 40, which in response thereto provides an update signal to inputs 76 and 78 of the input and predict modules respectively. The second input of the OR-gate 40 is adapted to receive a learn signal~ The learn signal maybe from a human interface processor, such as the one shown in Figure l, A
human interface processor maybe used to provide human initiated reinforecement when the PLE is learning. Such a processor and its reinforcing function maybe used with a single PLE or a PLS, as shown in Figure 1. The learn signal may also come from another PLE when a PLS is used.
rrhe OR-gate 40 in response to either a learn or a self learn signal will cause an update signal to be provided to terminals 76 and 78 of the input and predict modules~ When an update signal is received at terminal 76 and 78 of the input and predict modules, the current information being received from the output module and the objects that were received and T . B . SLAC}C ET AL 1-6 stored in the input module will be accepted as true and recognized states. The state and position information will be used to update COM's contained in the input and predict modules.
Referring to Figure 6, there i5 shown a more detailed block diagram oE the input module 28 of Figure 5. An object database 80 includes short term memories and a plurality of COM's for long term memory as previously discussed. The object database has input terminals 42, 68, 72 and 76 for receiving the object information, the state inormation, and the position information and update signal respectively. The received object information is stored in the short term memory and is used to search in the long term memories. A first COM, called an alltree, within object database 80 stores the previously described type 1 knowledge, namely the frequency of object n-grams forming parts of all possible states. From this COM we receive pointers to appropriate single~rees from the nodes of which we receive the first type of probabilistic knowledge Pl at terminal ~2, namely the conditional probability that object Yt occuxs given the previous ob~ect concext or n-gram and state xi. This conditional probability identified as Pl is independant of position within the state and is calculated for all significant states. The attribute lists of the singletrees are used to pr~vide an output at terminal 84 corresponding to the conditional probability that object Yt with beginning position f and ending position g will occur given the previous object context or n-gram with consistant positioning and state xi. This conditional probability P2 is derived from the type 2 knowledge, namely the positional frequency of object n-grams within states and i5 calculated for all significant states and times that such states could end~
The conditional probability Pl from terminal 82 is provided to a spanned-length function module 86 by way of an input terminal a8. Module 86 also receives at a terminal 90 signal DB from an end-time state-length function module 92 having an output terminal 94. Said signal DB corresponds to the distance back (DB) or to the begin time for each significant state-length. The spanned-length function module 86 stores the previously received Pl value and combines the currently received Pl value with the stored previous value. ThP sum is then stored and indexed by time to develop accummulated probabilities stored for vaious times. The module uses the DB input to calculate the ~3~3~9 T. B. SLACK ET AL 1-6 difference between ~he accummulated probabiliky at the current time and the accummulated probability at ~he time ~B before the current tima. This difference i5 then outputted at terminal 96 as a probability P6 that the sequence of objects between the begin time and the end time occurs for a given state. This probability is calculated using the previously discussed equation (17).
The end-time state-length function module 92 receives at terminal 98 the conditional probability P2 outputted from terminal 84. Module 92 outputs at terminal 100 the accummulated probability values as each end-time passes, said accummulated probability being the probability that the sequence back to some begin time occurs in the given state. This probability P7 is derived using the product found in equation (18), previously discussed.
The maximum value of the P7 pxobability will give the probability that some state ends at the present time. This maximum value of P7 is determined by the maximum value function module 102 which includes the output terminal 44 which provides the End-Of-State Probability PE.
A length normalizer module 104 receives the outputs of module 92 and provides at a terminal 106 a signal P8 corresponding to the probability that the begin time is correct given the sequence of objectsr the end-time and the state. This probability is calculated in accordance with the previously discussed equation (19).
The outputs of modules 86 and 104 are multiplied together to provide at terminal 46 the previously discussed Input Probability calculated in accordance with equation (16) wherein 3~ the results of equations (17) and (19) are multiplied together.
The end-time state-length function module 92 receives the previously discussed second type of conditional probabilistic knowledge P2 from the object database 800 The positional information stored in the database provides values for the number of objects preceeding the last object of the n-gram and the number of objects following the last object plus l. These values are called the "distance to beginning" and the "distance to ending" and the sum of these values will always equal the length of the state at the time that the pattern was learned.
The probability P7 that an object sequence is a complete pattern for a state is determined by the product found within ~3~3~9 To B. SLACK ET AL 1-6 the previously discussed equation (18), which defines the signal provided at terminal 100 of module 92.
Referrin~ to Figure 7, there i5 shown a detailed block diagram of the end-~ime state-length function module.
Conditional probabilistic information P2 arrives at terminal 98 of a decoder 1~8. The decoder ~unctions ~o separate the information received at the input and to provide a timing signal each time a New P2 signal enters at terminal 98. The decoder outputs the timing signal called New P2, a DE signal, a DB signal and a probability value at terminals 110, 112, 114 and 116 respectively.
A matrix of ~11 possible state-lengths would be exceedingly large and most nodes would have zero entries. Dealing with such a large matrix would tax ~he memory capacity of the database;
therefore, the significant states including their DB and DE
information will be indexed by a common parameter ~. Thus, at a given object time the information provided by the decoder 108 includes the New P2 timing signal, DE (q, state), DB (q~ state) and probability value (q, state).
The New P2 timing signal i5 provided to a counter 118, a multiplexer 120 and a latch 124. The counter 118 increments each time a New P2 signal is received and provides a current time output which is a modular number based on how many addresses are needed in the memory to span the distance from the beginning`to the end of the longest state.
An adder 113 is provided to add the current time to DE to provide a signal corresponding to end-time, i.e~ current ~ime plus distance to end equals "end-time". The ~E signal is added to the DB signal by another adder 115 to provide a signal corresponding to "length'~O The probability value is multiplied in a multiplexer 117 by an accumulated probability value to provide a 'Iproduct''O The "end-time", "length" and "pxoduct"
signals are applied to multiplexer 120 on the left side marked "1".
The top side of the multiplexer marked "0'~ receives three signals, two corresponding to 0 and one being the current time.
The multiplexer provides output signals on the right side marked "out'l. When the multiplexer receives a high level signal at a select input "S" from the New P2 ~ignal~ the multiplexer selects from the left side marked "1".

T. B. SLACK ET ~L 1 6 Memory 122 has an address input which receives a time signal corresponding to end time or current time depending on the multiplexer output. Two data signal~ are inputted to the memory from the left by the multiplexer. ~he first data signal is either "length" or zero and the second is the "product" or zero depending upon whether New P2 is high or low.
When New P2 is high the multiplexer selects from the left and the memory address receives the value of the time when the state ends i.e. "end time". The memoxy stores the "length" (q, state) and the "product" (q, state). A read modify write operation is done on the memory to develop the accumalated value which is stored at the addressed "end-time".
When the New P2 signal goes low, the multiplexer selects from the top. Thus, the memory address input receives the current time and a second read modify wri~e is done. Latch 124 is responsive to the low signal on New P2 so that the data values at the current time are latched and will be available at the outputs. The write operation provides a clearing of the information in the memory at the current time address since "0"s are written in. This prepares the memory for the next cycle of information. It should be noted tha~ the data values were actually written for the "end-times" of the states so that when the current time reaches the "end-time" the '~length'~ of the state is the same as the DB and the length information outputted from the memory corresponds to DB.
Referring now to Figure 8, there is shown a detailed block ~iagram of the spanned-length function module 86. As previously discussed terminal 88 receives the conditional probabilistic information Pl which enters a decoder 126~ The decoder provides a timing signal New Pl at an output 128 when ~ew probability information Pl is entered. The New Pl signal is provided to a counter 130, a multiplexer 132, a d~lay circuit 134, a memory 136, a latch 138 and another latch 140. The counter 130 in response to the timing signal New Pl generates a current time sign~l in a manner similar to that generated in the end-time state-length function module. The current ti~e signal is applied to one input of the multiplexer 132 and to an add circuit 129. The DB signal from the end-time state length function module 92 is provided to a terminal 90 which is an inverted input of the add circuit. Thus, the add circuit effectively subtracts tha DB from the current time to output a ~3~ 9 T. B. SLACK ET AL 1-6 - 3~ -begin time signal which is provided to another input of the multiplexer.
The multiplexer is controlled by the New Pl signal to provide either the current time or the begin time to an address input of memory 136 When the New Pl signal is high the multiplexer 132 selects from the left and the memory is in a write mode. At this time, the memory is addressed by the value of the current time signal provided from counter 130.
Decoder 126 provides at a second output 142 a signal correspooding to the conditional probability Pl, which output is connected to a first input of a multiplier 144 which multiplier has a second input connected to its output through the delay circuit 134 so that the output of the multiplier corresponds to the produc~ of the probability value P1 multiplied by the accummulated value resulting from the product of previous inputted probability values. The output of multiplier 144 is connected to an input of memory 136 and an input of latch 140 where the current accumulated value i5 latched for use during the next New Pl low period and stored in the memory 136 and indexed at the current time.
When the timing signal New Pl is low the multiplexer selects the begïn time signal which is the value of the count signal outputed from counter 130 minus the DB (q, state) 25 received at terminal 90. At this time the memory 136 is reading and latch 138 holds the information corresponding to - the accumulated value at the begin time that is read. The outputs of latches 138 and 140 are provided to an add circuit 141 with the output of latch 138 going to an inverted input 143 30 so that the output of the add circuit 141 on terminal 96 is really the difference between the inputs. Thus, the output ak terminal 96 is the difference betweent the current accumulated value and the accumulated value -that existed at a distance DB
in the past i.e. at the begin time. The output at terminal 96 is derived in accordance with the previously discussed equation (17) and is identified as P6. It must be kept in mind that we are only interested in the difference and it is assumed that the borrow is possible and the value of the data in the memory may be allowed to overflow without penalty, provided that the memory was all the same value when the first object arrived and that the size of the data is higher than the largest difference possible.

3 ~ 9 T. B. SLACK ET AL 1-6 Referring to Figure 9, there is shown a detailed block diagram of the length normalizer 104, which recelves the probability information P7 and the distance to begin DB
information f~om module 92 and provides an output P8 in accordance with equation (l9)o Both the probability value P7 and the distance to begin value DB are provided to a module 146 which provides an output equivalent to the XY or P7Y in accordance with equation (18). The output of module 146 is provided to a module 144 where all probability values for each (q, state) are added together to provide an output that is indexed only by (~, state). In order to do this summation the value of the probability which is a log function must be exponentiated after which the sum is taken. The log i5 then applied before the value ;s passed on~ The outputs of modules 144 and 1~6 are provided to a module 148 where the output of module 146 is divided by the output of module 144. The-result of this division is provided to an encoder 150 where it is encoded with the distanct to begin or length information.
To provide an output P8 at terminal 106 in accordance with equation (19). The probability P8 i~ indexed by length and state with the parameter q being eliminated by the encoder.
Referring to Figure 10, there is shown a datailed block diagram of the predict module 30 including a length database 152, a both database 154 and state database 156 all comprising separate COM's for storing the type 4, type 5 and type 3 knowledge respectively. A decoder 168 receives options information at terminal 48, state information at terminal 70, position information at terminal 74 and an update signal at terminal 78. The decode module separates the received information and provides at an output 170 length information and at an output 172 state information. The length da-tabase 152 receives the length information in the form of number~ of objects that make up states. The length information is organized by sequences of lengths for storage in a COM. The state database 156 receives state information which is organized by sequences of states in a COM. The both database 154 receives both length and state information which i~ stored in a COM and is organized by state-length pairs. The state database 156 provides a first output comprising all possible next states and their probabilities. The probabilities for the states are in the forms of the previously discussed type 3 conditional ~L~32~ ig T ~ B . SLACK ET AL 1 6 probabilistic information P3. The output of database 156 is provided to an input of a multiplexer 158. The length database 152 provides an output comprising all possible`next lengths and their probabilities in the form of the type 4 conditional probabilistic in~ormation P4. The length database output is connected to another input of multiplexer 158. Database 154 provides an output comprising all possible next state-length pairs and their probabllities which probabilities are in the form of the type 5 conditional probabilistic information P5 previously discussed. The output information from databases 152 and 156 each include support coefficients corresponding to the usefulness of the probability information being provided by the respective database. The support coefficients are derived using equation (9).
The P5 information from the both database 154 is provided to summing circuits 153 and 155 where the probabilities of all states and all lengths are summed respectively. This is the same type of summing across that was done in the length normalizer. The outpu~s of the summing circuits 153 and 155 are provided to divider circuits 157 and 159 respectively. The P5 signal i5 also provided to dividers 157 and 159 so that the dividers each output a signal in accordance with equations (21A) and (22A) respectively~
The outputs of dividers 157 and 159 are provided to multipliers 161 and 163 respectively as are the P4 and P3 signals. Multipliers 161 and 163 output signals to multiplexer 158 in accordance with equations (22) and (21) respectively.
The output information including the probabilities and the support coefficients from multipliers 161 and 163 are provided to module 166 where the probabilities are multiplied by the support coefficients to provide confidence factors for both the state and length information provided to multiplexer 158. The confidence factor signals for state and length information are provided to a comparator 160. Comparator 160 provides an output depending upon which confidence factor is higher, this output controls the multiplexer 158 so that the output signal ep is selected from either equation (21) or (22) depending upon which has the higher confidence factor.
Referring to Figure 11, there is shown a de~ailed block diagram of the decide module 32. The Input Probability PI
calculated in accordance with equation (16) is received at 3 ~9 T. B. SLACK ET AL 1~6 termlnal 52 of a decoder circuit 1740 The decoder circuit separates the Input Probability into its value and length and urther provides a clock signal New PI when a new Input Probability arrives. The clock signal is provided to a counter 176, a multiplexar 178, an option memory 180 and a prediction memory 182. The clock signal New PI clocks the counter so that it pxovides an output corresponding to current time. The length information from the decoder 174 is provided to an inverting input of a summing circuit 175 where it is effectively subtracted from the current time to provide a signal corresponding to begin time which is provided to an input on the left or "1" side of the multiplexer 178.
Multiplexer 178 also receives on the left side pas~ options information from the option memory 180. Th~ top or the "0"
side the multiplexer receives current time and options information. The outputs from the multiplexer 178 are provided to both the option memory 180 and the prediction memory 182 The prediction memory 182 is addressed by the time, and the option data from the multiplexer.
Multiplexer 178 is clocked by signal New PI and first selects from the left when Ne~ PI is high which causes the option memory to be addressed by the current time minus the length information or the begin time. The output of the option memory is a list of options that were available at the addressed time or begin time. This list includes states, positions and probabilities at the addressed begin time. The output of the option memory is looped back and provided as a second input to the multiplexer 178 so that the past options data may be used to aid in the addressing of the prediction memory 182. The time from the multiplexer and the past options data both address the prediction memory for storage of the Predict Probability Pp data received at terminal 54. The Pp data consists of sets of states, lengths and probabilities.
The value information provided by decoder 174 containing input probability, the past options data from the option memory 180, and ~he past predictions data Erom the prediction memory 182 are multiplied together in 183 to implement equation (6) using equation (7)~ The first term in equa-tion (7) is input probability, the second term of equation (7) is past predictions. Equation (7) is the second term of equation (6) and the first term is the past options from the option memory.

3~ 3'-~ T. B. SL~CK ET AL 1-6 The product of this multiplication is provided to a maximum N
function circuit 184. The maximum N function circuit chooses the N best options based on their confidence levelsO These options are outputted at terminal 56.
When the New PI timing signal goes low the mul~iplexer 178 selects from the top and the option memory is addressed at the current time. The write input of the option memory 180 is activated so that the current options from the maximum N
function circuit 184 are written into option memory 180 through multiplexer 178 and are addressed at the current time. These current options and the current time also address the prediction memory 182 which is also write enabled by the New PI low to store the ~redict Probability data for future use.
The size of both of the memories 180 and 182 and the counter 176 must be sufficient to span the length of the longest state plus the options needed to specify their history.
Referring to Figure 12, there is shown a detailea block diagram of the output module 34. An option decoder 188 r~ceives the options from the decide module 32 at terminal 60. The options including states, lengths and probabilities are stored in decoder 188 and are addressed by time. The output module uses tha end of state probability signal which is received at terminal 58 to decide when the data in the list of options is sufficiently good ~o output as a recognition of the input objects as a recognized state. The end of state probability is smoothed by circuit 186 to avoid false triggers. The smoothed function is provided to circuit 190 where the maximum value of the smoothed function is stored. A divider 192 is provided to select a predetermined perc~ntage of the stored maximum value.
30 The output of the smoother 186 and the divider 192 are provided to a comparator 194 so that when the peak value of the signal coming from the smoother 186 drops below the prede~ermined percentage of the stored maximum value comparator 194 provides an output to option decoder 188 to trigger said decoder.
The End-Of-State Probability signal PE is also provided to a maximum end-time circuit 196 which stores the time of the actual maximum end of state probability value. This maximum end-time value is also provided to option decoder 188 so that when the decoder 188 is triggered ît may select the best options that were stored and addressed at the maximum end-time.
These best options signals are then provided as confidence, ~ ~323~ T. B~ SLACK ET AL 1 6 state and position output signals. At this time an output changed signal is provided by the decoder 188 which is used to reset the maximum functian circuit 190 and the maximum end-time function circui~ 196 so that a new maximum function and maximum end time may be sought.
Referring to Figure 13, there is shown the use of a probabilistic learning system in a character recognition application. ~ video camera 200 is focused on a hand printed word 202 namely "HELL0'9 that is to be recognized. The signals from the video camera 200 are provided to a video processor ~04 which provides outputs to a video monitor 2060 A workstation 208 may be provided particularly for use during the learning mode for providing external reinforcement. The PLS is similar to that shown in Figure 1 in that it comprises an array 12 consisting of eight individual PLE's 14a to 14h, an input processor 22~ an output processor 16 and an interface circuit 20.
The input to the PLS is terminal 11 from the video processor while the PLS output 13 is passed through the user interface 20 and on to the video processor 2040 The video representation of ~he hand printed word 'IHELLO'' i5 shown in the upper portion of the video monitor a~ 210. The video representation is digitized as shown at 212 on the video monitor~
The digiti~ed word "HELL0" is shown more clearly in Figure 14 where each vertical slice is a time frame containing a predetermined number of pixels of information a~ for example 10 pixels as shown in Figure 14. The digitized-word is scanned from left to right and input objects in the form of time slices containing 10 pixels each are provided to the PLS. The sequences of objects could be provided to a single high capacity PLE or to a PLS comprised of an-array as in the case of Figure 13.
The power of using a PLS comprising an array may be illustrated by referring to Figures 1 and 13. Inputting objects containing 10 pi~els presents a rather complex recognition problem which would require a PLE with considerable capacity.
The array provides the advantage of parallelism to speed up the execu~ion oE the recognition task by permitting the input information to be partitioned between a plurality of individual PLE's. The information is partitioned in an overlapping or redundant manner to enhance the reliability of the system. Due to the redundant overlapping a breakdown in a portion of the system will not affect the overall system operation.

~ 3~3~ T. ~O SLACK ET AL 1-6 Referring to Figure 13, there is shown, that the input preprocessor 22 receives 10 pixels of information and partitions these pixels 50 that pixels 1 to 4 are provided to ~he first PLE
14a, pixels 3 to 6 are provided to PLE 14b, pixels 5 to 8 are provided to PLE 14c and pixels 7 to 10 are provided to PLE 14d.
Each PLE performs a recognition function on the i~puts it receives to identify output states in the form of a certain type of feature. This is not to be confused with a feature extraction steps but is a true pattern classification step and illustrates the generalized aspect of the PLE which allows it to recognize and learn anything such as an abstract feature as opposed to such things as letters, numbers, etc. It might be said that the features that are recognized are slices of output states. Thus, the first bank of four PLE's i.e. PLE's 14a to 15 14d receives a total of 16 bits of information, 4 bits to each PLE in overlapping relationship. Each PLE in the first bank outputs 4 bits identifying a par~icular feature out of 16 possible features.
The 4 bit feature representation outputted from PLE 14a of the first bank is provided to the inputs of the PLE's of the second bank i.e. 14e to 14h. In like manner, the 4 bit representation of a feature at the output of the second PLE 14b of the first bank is provided to the inputs of each PLE of the second bank PLE's Thus, each PLE of the second bank receives four, 4 bit feature inputs. Each PLE in the second ban~
provides a 4 bit output which comprise one fourth of a 16 bit redundantly coded name for a recogni7ed character or output state. Thus, the recognition task is simplified in that each PLE in the second bank must only recogni~e the first 4 bits of a 16 bit coded name for a characterO The output processor 16 receives the 16 bit redundantly coded representation of the output state and reduces the 16 bits to 8 bits using a BCH
decoding system. The 8 bit output provided at 13 is in the form of an ~SCII Code for a character recognition system. The ASCII
Code has the capability of representing 256 possible letters or characters.
By using the 16 to 8 bit reduction, significant data overlap is provided so that many of the 16 bits could be in error or missing and the output would still be correctO Thus, one PLE could fail and the system would continue to function without error.

~32~9 T. B. SLACK ET AL 1-6 Training of ~he array takes place in a manner similar to that of an individual PLE in that external reinforcement learn signals may be providad through a human in~erface. In addition, the PLE's of an of array are iDterconnected so that the self learn signal from an individual PLE is provided to the learn input of its source P1E's. Thus, when the PLE of a second bank provides an output with a high confidence levell thls indication will be relayed back to the source PLE's in the first bank. All of this training is of course in addition to the internal self learning of each individual P1E.
The array shown in Figures 1 and 13 comprises a 4 by 2 arrangement. The next size array would be 16 by 3 Eor a to~al of 64 PLE's comprising the array. Larger arrays may be built using this progres~ion in size; however, while a larger array would provided for more parallelism in its operation and greater reliability, its speed would be reduced due to the number of PLE's through which the data must flow from the inpu~ to the output.
Figure 14 shows an expanded concept of using a plurality of PLS's wherein pixels 216 are first used to recognize charact~rs 218 as output states from PLS 12. The characters 218 may become input objects to a PLS 220 which is used to recognize words 222 as output states. The words 222 become input objects to a PLS
224 to recognize scripts 226 as output states.
It should also be remembered that the PLS is no~ limited to use in an optical character reader but rather may be used in many other applications, such as voice recognition. The PL5 is appropriate for use wherever sequential patterns are to be recognized.

Claims (14)

THE EMBODIMENTS OF THE INVENTION IN WHICH AN EXCLUSIVE PROPERTY
OR PRIVILEGE IS CLAIMED ARE DEFINED AS FOLLOWS:
1. A probabilistic learning element, that sequentially receives objects and outputs sequences of recognized states, said learning element comprising:
means for sequentially receiving objects;
means for storing, said received objects, sequences of received objects, previously learned sequences of states, states contained in said previously learned sequences of states, and predetermined types of knowledge relating to, said previously learned sequences of states, said states contained in said previously learned sequences of states, objects contained in said previously learned sequences of states, and sequences of objects contained in said previously learned sequences of states, whereby current object information relating to said received objects and said sequences of received objects is stored as well as statistical information relating to previously learned sequences of states and said states, objects and sequences of objects contained in said previously learned sequences of states;
means for correlating said stored current object in-formation with said stored statistical information for assigning probabilities to possible next states in the sequence of recognized states;
means, responsive to said probabilities of possible next states, for determining a most likely next state;
means, responsive to the stored current object infor-mation and statistical information, for providing a signal cor-responding to the probability that a state has ended; and means, responsive to said end of state signal, for outputting said most likely next state as a recognized next state in a recognized state sequence.
2. A probabilistic learning element as claimed in claim 1, additionally comprising means for providing a rating of confidence in said recognized next state.
3. A probabilistic learning element as claimed in claim 2, additionally comprising:
means for accumulating the ratings of confidence of the recognized states of the recognized state sequence; and means, responsive to said accumulated ratings of con-fidence, for causing said means for storing to store the recog-nized state sequence as a previously learned sequence of states, and to store the objects, sequences of objects and states in the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequen-ces of states forming said recognized state sequence as statis-tical information relating to previously learned sequences of states when the accumulated ratings exceed a predetermined thres-hold level.
4. A probabilistic learning element as claimed in claim 1, additionally comprising learning supervision means, responsive to external reinforcement signals, for causing said means for storing to store the recognized state sequence as a previously learned sequence of states, and to store the objects, sequences of objects and states in the recognized state sequence and the pre-determined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said recognized state sequence as statistical information relating to previously learned sequences of states.
5. A probabilistic learning element as claimed in claim 1, additionally comprising:
means for providing a rating of confidence in said recognized next state;
learning supervision means, adapted to receive said rating of confidence and an external reinforcement signal, for providing an output signal in response to accumulated ratings of confidence of the recognized states of the recognized state sequ-ence and the external reinforcement signal when either the accumu-lated ratings of confidence exceed a predetermined threshold level or an external reinforcement signal is received; and means, responsive to the output signal from the learning supervision means, for causing said means for storing to store the recognized state sequence as a previously learned sequence of states, and to store the objects, sequences of objects, and states in the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said recognized state sequence as statistical information relating to previously learned sequences of states.
6. A probabilistic learning element as claimed in claim 1, wherein the predetermined types of knowledge that are stored include the number of occurrences of: each stored object, each stored sequence of objects, each stored state and each stored sequence of states.
7. A probabilistic learning element as claimed in claim 6, wherein the predetermined types of knowledge that are stored additionally include state lengths and the number of their occur-rences and sequences of state lengths and the number of their occurrences.
8. A probabilistic learning element as claimed in claim 7, wherein the predetermined types of knowledge that are stored additionally include state-length pairs and the number of their occurrences and sequences of state-length pairs and the number of their occurrences.
9. A probabilistic learning element as claimed in claim l, wherein the means for correlating includes a first means for determining the probabilities that possible states will span an object sequence having a particular begin time and end time and second means for determining the probabilities of state-length pairs given the previous state-length pair context.
10. A probabilistic learning element as claimed in claim 9, additionally comprising means responsive to the previously men-tioned probabilities for implementing an algorithm to provide probabilities of possible states, with a particular length that span an object sequence given the previous state-length pair context.
11. A probabilistic learning element as claimed in claim 1, wherein the means for storing stores information on a plurality of previously learned states and the objects contained therein, whereby several levels of context are stored and the means for correlating implements an nth order Markov process correlating several levels of stored context.
12. A probabilistic learning element as claimed in claim 4 additionally comprising means for correcting a recognized state -44a-sequence prior to initiating an external reinforcement signal.
13. A probabilistic learning element as claimed in claim 5 additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
14. A method of recognizing a sequence of states from a sequence of inputted objects utilizing a probabilistic learning element, comprising the steps of:

correlating said sequence of inputted objects and predetermined types of knowledge relating to said sequence of inputted objects with stored information relating to previously learned states, objects and sequences of objects and predetermined types of knowledge relating to said previously learned states, objects and sequences of objects including the number of occurrences of each;

providing probabilities for possible next states in a sequence of states based on said correlations;

determining a most likely next state in the sequence of states each time a new object is received;

deriving from stored information a signal corresponding to the probability that a state has ended; and outputting the most likely next state as a recognized next state in the sequence of recognized states in response to the end of state probability.
CA000472108A 1984-01-16 1985-01-15 Probabilistic learning element Expired CA1232359A (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US06/571,230 US4593367A (en) 1984-01-16 1984-01-16 Probabilistic learning element
US571,230 1984-01-16

Publications (1)

Publication Number Publication Date
CA1232359A true CA1232359A (en) 1988-02-02

Family

ID=24282833

Family Applications (1)

Application Number Title Priority Date Filing Date
CA000472108A Expired CA1232359A (en) 1984-01-16 1985-01-15 Probabilistic learning element

Country Status (6)

Country Link
US (1) US4593367A (en)
JP (1) JPS60229182A (en)
AU (1) AU3666784A (en)
BR (1) BR8500194A (en)
CA (1) CA1232359A (en)
ES (1) ES8700464A1 (en)

Families Citing this family (86)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2607457B2 (en) * 1984-09-17 1997-05-07 株式会社東芝 Pattern recognition device
US4718094A (en) * 1984-11-19 1988-01-05 International Business Machines Corp. Speech recognition system
US4816994A (en) * 1984-12-04 1989-03-28 Tektronix, Inc. Rule acquisition for expert systems
JPS61218323A (en) * 1985-03-20 1986-09-27 株式会社東芝 Fault identification
US4670848A (en) * 1985-04-10 1987-06-02 Standard Systems Corporation Artificial intelligence system
US4759068A (en) * 1985-05-29 1988-07-19 International Business Machines Corporation Constructing Markov models of words from multiple utterances
US4967340A (en) * 1985-06-12 1990-10-30 E-Systems, Inc. Adaptive processing system having an array of individually configurable processing components
US4697243A (en) * 1985-07-25 1987-09-29 Westinghouse Electric Corp. Methods of servicing an elevator system
US4754489A (en) * 1985-10-15 1988-06-28 The Palantir Corporation Means for resolving ambiguities in text based upon character context
US4763277A (en) * 1986-01-17 1988-08-09 International Business Machines Corporation Method for obtaining information in an expert system
JPS62231993A (en) * 1986-03-25 1987-10-12 インタ−ナシヨナル ビジネス マシ−ンズ コ−ポレ−シヨン Voice recognition
US4827521A (en) * 1986-03-27 1989-05-02 International Business Machines Corporation Training of markov models used in a speech recognition system
US4882759A (en) * 1986-04-18 1989-11-21 International Business Machines Corporation Synthesizing word baseforms used in speech recognition
US4783828A (en) * 1986-06-02 1988-11-08 Honeywell Inc. Two-dimensional object recognition using chain codes, histogram normalization and trellis algorithm
US4821333A (en) * 1986-08-22 1989-04-11 Environmental Research Inst. Of Michigan Machine learning procedures for generating image domain feature detector structuring elements
FR2604004B1 (en) * 1986-09-11 1990-05-11 Anatex MANUSCRIPT WRITING RECOGNITION METHOD
JPH0642268B2 (en) * 1986-10-31 1994-06-01 日本電気株式会社 Character recognition device
CA1287175C (en) * 1987-03-12 1991-07-30 Analog Intelligence Corporation Back propagation system
US5014327A (en) * 1987-06-15 1991-05-07 Digital Equipment Corporation Parallel associative memory having improved selection and decision mechanisms for recognizing and sorting relevant patterns
US4852173A (en) * 1987-10-29 1989-07-25 International Business Machines Corporation Design and construction of a binary-tree system for language modelling
US4855923A (en) * 1988-01-06 1989-08-08 Xerox Corporation Parts assembly using signature analysis
US5014220A (en) * 1988-09-06 1991-05-07 The Boeing Company Reliability model generator
US4941122A (en) * 1989-01-12 1990-07-10 Recognition Equipment Incorp. Neural network image processing system
US5140530A (en) * 1989-03-28 1992-08-18 Honeywell Inc. Genetic algorithm synthesis of neural networks
US5077677A (en) * 1989-06-12 1991-12-31 Westinghouse Electric Corp. Probabilistic inference gate
US5056037A (en) * 1989-12-28 1991-10-08 The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration Analog hardware for learning neural networks
US5119425A (en) * 1990-01-02 1992-06-02 Raytheon Company Sound synthesizer
US5274714A (en) * 1990-06-04 1993-12-28 Neuristics, Inc. Method and apparatus for determining and organizing feature vectors for neural network recognition
US5161204A (en) * 1990-06-04 1992-11-03 Neuristics, Inc. Apparatus for generating a feature matrix based on normalized out-class and in-class variation matrices
WO1991019989A1 (en) * 1990-06-21 1991-12-26 Reynolds Software, Inc. Method and apparatus for wave analysis and event recognition
US6683697B1 (en) * 1991-03-20 2004-01-27 Millenium L.P. Information processing methodology
US5258855A (en) * 1991-03-20 1993-11-02 System X, L. P. Information processing methodology
US5335291A (en) * 1991-09-20 1994-08-02 Massachusetts Institute Of Technology Method and apparatus for pattern mapping system with self-reliability check
US5526444A (en) * 1991-12-10 1996-06-11 Xerox Corporation Document image decoding using modified branch-and-bound methods
US5903454A (en) 1991-12-23 1999-05-11 Hoffberg; Linda Irene Human-factored interface corporating adaptive pattern recognition based controller apparatus
US10361802B1 (en) 1999-02-01 2019-07-23 Blanding Hovenweep, Llc Adaptive pattern recognition based control system and method
US6400996B1 (en) 1999-02-01 2002-06-04 Steven M. Hoffberg Adaptive pattern recognition based control system and method
US6418424B1 (en) 1991-12-23 2002-07-09 Steven M. Hoffberg Ergonomic man-machine interface incorporating adaptive pattern recognition based control system
US6850252B1 (en) 1999-10-05 2005-02-01 Steven M. Hoffberg Intelligent electronic appliance system and method
US8352400B2 (en) 1991-12-23 2013-01-08 Hoffberg Steven M Adaptive pattern recognition based controller apparatus and method and human-factored interface therefore
US7242988B1 (en) 1991-12-23 2007-07-10 Linda Irene Hoffberg Adaptive pattern recognition based controller apparatus and method and human-factored interface therefore
US5649066A (en) * 1992-01-03 1997-07-15 The Florida State University For And On Behalf Of The Florida Board Of Regents Method and apparatus for refinement of learning in expert networks
WO1993018483A1 (en) * 1992-03-02 1993-09-16 American Telephone And Telegraph Company Method and apparatus for image recognition
US5598511A (en) * 1992-12-28 1997-01-28 Intel Corporation Method and apparatus for interpreting data and accessing on-line documentation in a computer system
US5524176A (en) * 1993-10-19 1996-06-04 Daido Steel Co., Ltd. Fuzzy expert system learning network
US8364136B2 (en) 1999-02-01 2013-01-29 Steven M Hoffberg Mobile system, a method of operating mobile system and a non-transitory computer readable medium for a programmable control of a mobile system
US7904187B2 (en) 1999-02-01 2011-03-08 Hoffberg Steven M Internet appliance system and method
US6574596B2 (en) * 1999-02-08 2003-06-03 Qualcomm Incorporated Voice recognition rejection scheme
US6519562B1 (en) * 1999-02-25 2003-02-11 Speechworks International, Inc. Dynamic semantic control of a speech recognition system
US7167587B2 (en) * 2002-08-30 2007-01-23 Lockheed Martin Corporation Sequential classifier for use in pattern recognition system
US7412842B2 (en) 2004-04-27 2008-08-19 Emerson Climate Technologies, Inc. Compressor diagnostic and protection system
US7275377B2 (en) 2004-08-11 2007-10-02 Lawrence Kates Method and apparatus for monitoring refrigerant-cycle systems
US7583819B2 (en) * 2004-11-05 2009-09-01 Kyprianos Papademetriou Digital signal processing methods, systems and computer program products that identify threshold positions and values
US8590325B2 (en) 2006-07-19 2013-11-26 Emerson Climate Technologies, Inc. Protection and diagnostic module for a refrigeration system
US7930564B2 (en) * 2006-07-31 2011-04-19 Intel Corporation System and method for controlling processor low power states
US20080216494A1 (en) 2006-09-07 2008-09-11 Pham Hung M Compressor data module
US20090037142A1 (en) 2007-07-30 2009-02-05 Lawrence Kates Portable method and apparatus for monitoring refrigerant-cycle systems
US9140728B2 (en) 2007-11-02 2015-09-22 Emerson Climate Technologies, Inc. Compressor sensor module
US8332348B1 (en) 2009-07-10 2012-12-11 United Services Automobile Association (Usaa) Knowledge extraction and prediction
US8405531B2 (en) * 2010-08-31 2013-03-26 Mitsubishi Electric Research Laboratories, Inc. Method for determining compressed state sequences
US8706653B2 (en) * 2010-12-08 2014-04-22 Microsoft Corporation Knowledge corroboration
US9285802B2 (en) 2011-02-28 2016-03-15 Emerson Electric Co. Residential solutions HVAC monitoring and diagnosis
US8964338B2 (en) 2012-01-11 2015-02-24 Emerson Climate Technologies, Inc. System and method for compressor motor protection
US9310439B2 (en) 2012-09-25 2016-04-12 Emerson Climate Technologies, Inc. Compressor having a control and diagnostic module
US9665088B2 (en) 2014-01-31 2017-05-30 Fisher-Rosemount Systems, Inc. Managing big data in process control systems
US9397836B2 (en) 2014-08-11 2016-07-19 Fisher-Rosemount Systems, Inc. Securing devices to process control systems
US10678225B2 (en) 2013-03-04 2020-06-09 Fisher-Rosemount Systems, Inc. Data analytic services for distributed industrial performance monitoring
US9804588B2 (en) 2014-03-14 2017-10-31 Fisher-Rosemount Systems, Inc. Determining associations and alignments of process elements and measurements in a process
US10649424B2 (en) 2013-03-04 2020-05-12 Fisher-Rosemount Systems, Inc. Distributed industrial performance monitoring and analytics
US10223327B2 (en) * 2013-03-14 2019-03-05 Fisher-Rosemount Systems, Inc. Collecting and delivering data to a big data machine in a process control system
US10909137B2 (en) 2014-10-06 2021-02-02 Fisher-Rosemount Systems, Inc. Streaming data for analytics in process control systems
US10386827B2 (en) 2013-03-04 2019-08-20 Fisher-Rosemount Systems, Inc. Distributed industrial performance monitoring and analytics platform
US10282676B2 (en) 2014-10-06 2019-05-07 Fisher-Rosemount Systems, Inc. Automatic signal processing-based learning in a process plant
US10866952B2 (en) 2013-03-04 2020-12-15 Fisher-Rosemount Systems, Inc. Source-independent queries in distributed industrial system
US9558220B2 (en) 2013-03-04 2017-01-31 Fisher-Rosemount Systems, Inc. Big data in process control systems
US9823626B2 (en) 2014-10-06 2017-11-21 Fisher-Rosemount Systems, Inc. Regional big data in process control systems
US10649449B2 (en) 2013-03-04 2020-05-12 Fisher-Rosemount Systems, Inc. Distributed industrial performance monitoring and analytics
US9803902B2 (en) 2013-03-15 2017-10-31 Emerson Climate Technologies, Inc. System for refrigerant charge verification using two condenser coil temperatures
US10031489B2 (en) 2013-03-15 2018-07-24 Fisher-Rosemount Systems, Inc. Method and apparatus for seamless state transfer between user interface devices in a mobile control room
CA2904734C (en) 2013-03-15 2018-01-02 Emerson Electric Co. Hvac system remote monitoring and diagnosis
US9551504B2 (en) 2013-03-15 2017-01-24 Emerson Electric Co. HVAC system remote monitoring and diagnosis
EP2973242B1 (en) 2013-03-15 2020-12-23 Fisher-Rosemount Systems, Inc. Modelling and adjustment of process plants
US9765979B2 (en) 2013-04-05 2017-09-19 Emerson Climate Technologies, Inc. Heat-pump system with refrigerant charge diagnostics
US10168691B2 (en) 2014-10-06 2019-01-01 Fisher-Rosemount Systems, Inc. Data pipeline for process control system analytics
US10503483B2 (en) 2016-02-12 2019-12-10 Fisher-Rosemount Systems, Inc. Rule builder in a process control network
CN112417002B (en) * 2020-11-19 2023-04-07 华中师范大学 Information literacy data mining method and system applied to education informatization

Family Cites Families (37)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
BE608415A (en) * 1960-09-23 1900-01-01
US3103648A (en) * 1961-08-22 1963-09-10 Gen Electric Adaptive neuron having improved output
US3196399A (en) * 1962-10-01 1965-07-20 Ibm Specimen identification techniques employing selected functions of autocorrelation functions
US3267431A (en) * 1963-04-29 1966-08-16 Ibm Adaptive computing system capable of being trained to recognize patterns
US3446950A (en) * 1963-12-31 1969-05-27 Ibm Adaptive categorizer
US3457552A (en) * 1966-10-24 1969-07-22 Hughes Aircraft Co Adaptive self-organizing pattern recognizing system
US3581281A (en) * 1967-03-28 1971-05-25 Cornell Aeronautical Labor Inc Pattern recognition computer
US3440617A (en) * 1967-03-31 1969-04-22 Andromeda Inc Signal responsive systems
US3562502A (en) * 1967-08-14 1971-02-09 Stanford Research Inst Cellular threshold array for providing outputs representing a complex weighting function of inputs
US3601811A (en) * 1967-12-18 1971-08-24 Matsushita Electric Ind Co Ltd Learning machine
US3588823A (en) * 1968-03-28 1971-06-28 Ibm Mutual information derived tree structure in an adaptive pattern recognition system
US3613084A (en) * 1968-09-24 1971-10-12 Bell Telephone Labor Inc Trainable digital apparatus
JPS5039976B1 (en) * 1968-11-20 1975-12-20
FR2051725B1 (en) * 1969-07-14 1973-04-27 Matsushita Electric Ind Co Ltd
US3623015A (en) * 1969-09-29 1971-11-23 Sanders Associates Inc Statistical pattern recognition system with continual update of acceptance zone limits
US3725875A (en) * 1969-12-30 1973-04-03 Texas Instruments Inc Probability sort in a storage minimized optimum processor
US3716840A (en) * 1970-06-01 1973-02-13 Texas Instruments Inc Multimodal search
US3715730A (en) * 1970-06-01 1973-02-06 Texas Instruments Inc Multi-criteria search procedure for trainable processors
US3678461A (en) * 1970-06-01 1972-07-18 Texas Instruments Inc Expanded search for tree allocated processors
US3702986A (en) * 1970-07-06 1972-11-14 Texas Instruments Inc Trainable entropy system
US3700866A (en) * 1970-10-28 1972-10-24 Texas Instruments Inc Synthesized cascaded processor system
US3772658A (en) * 1971-02-05 1973-11-13 Us Army Electronic memory having a page swapping capability
US3701974A (en) * 1971-05-20 1972-10-31 Signetics Corp Learning circuit
US3753243A (en) * 1972-04-20 1973-08-14 Digital Equipment Corp Programmable machine controller
CH591726A5 (en) * 1973-07-30 1977-09-30 Nederlanden Staat
US3934231A (en) * 1974-02-28 1976-01-20 Dendronic Decisions Limited Adaptive boolean logic element
US3950733A (en) * 1974-06-06 1976-04-13 Nestor Associates Information processing system
NL165863C (en) * 1975-06-02 1981-05-15 Nederlanden Staat DEVICE FOR RECOGNIZING SIGNS.
US3988715A (en) * 1975-10-24 1976-10-26 International Business Machines Corporation Multi-channel recognition discriminator
JPS5272504A (en) * 1975-12-15 1977-06-17 Fuji Xerox Co Ltd Device for recognizing word audio
US4286330A (en) * 1976-04-07 1981-08-25 Isaacson Joel D Autonomic string-manipulation system
US4189779A (en) * 1978-04-28 1980-02-19 Texas Instruments Incorporated Parameter interpolator for speech synthesis circuit
CA1102451A (en) * 1979-06-29 1981-06-02 Percy E. Argyle Apparatus for pattern recognition
US4384273A (en) * 1981-03-20 1983-05-17 Bell Telephone Laboratories, Incorporated Time warp signal recognition processor for matching signal patterns
US4450530A (en) * 1981-07-27 1984-05-22 New York University Sensorimotor coordinator
US4507760A (en) * 1982-08-13 1985-03-26 At&T Bell Laboratories First-in, first-out (FIFO) memory configuration for queue storage
US4504970A (en) * 1983-02-07 1985-03-12 Pattern Processing Technologies, Inc. Training controller for pattern processing system

Also Published As

Publication number Publication date
ES539618A0 (en) 1986-09-01
AU3666784A (en) 1985-07-25
BR8500194A (en) 1985-08-20
ES8700464A1 (en) 1986-09-01
JPS60229182A (en) 1985-11-14
US4593367A (en) 1986-06-03

Similar Documents

Publication Publication Date Title
CA1232359A (en) Probabilistic learning element
US4620286A (en) Probabilistic learning element
CA1232687A (en) Probabilistic learning element employing context driven searching
US4599693A (en) Probabilistic learning system
Sen et al. Multistage stochastic decomposition: a bridge between stochastic programming and approximate dynamic programming
Kangin et al. Autonomously evolving classifier TEDAClass
CN111709493B (en) Object classification method, training device, object classification equipment and storage medium
Chen et al. Generative inverse deep reinforcement learning for online recommendation
Gokcesu et al. Recursive experts: An efficient optimal mixture of learning systems in dynamic environments
CN115017178A (en) Training method and device for data-to-text generation model
Guinet et al. Pareto-efficient acquisition functions for cost-aware Bayesian optimization
US20210374545A1 (en) Method and apparatus of increasing knowledge based on uncertainty in neural networks
US20220277031A1 (en) Guided exploration for conversational business intelligence
KR20190103011A (en) Distance based deep learning
CN113296755A (en) Code structure tree library construction method and information push method
CN112508177A (en) Network structure searching method and device, electronic equipment and storage medium
EP0157080A2 (en) Probabilistic learning element
US20220366315A1 (en) Feature selection for model training
US20210264259A1 (en) Learning device, inference device, learning method, and inference method
Ferrarotti et al. The benefits of sharing: a cloud-aided performance-driven framework to learn optimal feedback policies
Castro et al. Learning probability distributions generated by finite-state machines
Saito et al. Numeric Law Discovery Using Neural Networks.
KR102421987B1 (en) Electronic device, method, and computer readable medium for enhanced batch normalization in deep neural network
Thonglek et al. Automated quantization and retraining for neural network models without labeled data
Nath Fixed Point Propagation: A New Way To Train Recurrent Neural Networks Using Auxiliary Variables

Legal Events

Date Code Title Description
MKEX Expiry