Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of Claims
The present application is being examined under the claims filed July 9, 2026. The status of
the claims are as follows:
Claims 1-5, 7-19, and 21-22 are pending and under examination;
Claims 1-5, 7-11, 13, 15, 18, and 19 have been amended;
Claims 6 and 20 have been canceled;
Claims 21 and 22 are newly presented.
Response to Amendment
The Office action is in response to Applicant’s communication filed July 9, 2026, in response to
the Office action mailed May 13, 2026. Applicants’ remarks and amendments to the claims have been
considered with the results set forth below.
Response to Arguments
Applicant's arguments filed July 9, 2026, have been fully considered but they are not persuasive.
Regarding rejections under 35 U.S.C. § 112(b)
Applicant argues that claim 3 and 11 were amended to resolve the indefiniteness issues identified in the previous Office action. See Remarks, pg. 10 of 13)
Applicant’s amendments with respect to claims 3 and 11 have been fully considered and are persuasive. The previous rejection of claims 3 and 11 under 35 U.S.C. § 112(b) has therefore been withdrawn.
Regarding rejections under 35 U.S.C. § 103
Applicant’s arguments at pages 10-12 of the Remarks, with respect to the previous rejection of claims 1-20 under 35 U.S.C. § 103 over Freeman in view of Esser and Hu, have been fully considered and are persuasive with respect to the previously stated rejection. Accordingly, the previous § 103 rejection of claims 1-20 over Freeman in view of Esser and Hu is withdrawn. However, upon further consideration of the claims as amended, new grounds of rejection under 35 U.S.C. § 103 are made as set forth above.
Applicant argues, in particular, that Freeman maps document vectors to topic nodes but does not store the input documents themselves within those nodes, and therefore does not teach inserting a data neuron comprising the input data into the NoK. Applicant further argues that Freeman’s nodes represent topics rather than data neurons having matching features. See Remarks, pg. 10-11 of 13.
The present rejection does not rely on Freeman alone to teach these limitations. Rather, Freeman is relied upon for the similarity-based selection of a location within a topological neural structure, including selection of a winning node based on similarity and mapping or clustering documents according to the selected node. Shinn is relied upon for storing the received input itself within a memory graph. Shinn expressly teaches that an input sentence may be inserted into the knowledge store and that “nodes corresponding to the input sentence are inserted into a memory graph”. Shinn further teaches that related nodes may be associated and/or shared. See Shinn ¶ [0098]. Shinn additionally describes first and rest nodes used for storing received input sentences. Shinn teaches that a first sentence is recorded under a free first node and subsequent sentences are recorded under the corresponding rest node are created for insertion of the new sentence. See Shinn ¶¶ [0100]-[0101]. Shinn further teaches that a new input sentence tree is inserted at a first node and, in a subsequent example, identifies a previously inserted input sentence tree and a newly inserted input sentence tree, with the newly inserted sentence located at a newly appended first node and a rest node provided for a future input sentence. See Shinn ¶¶ [0104]-[0106].
Accordingly, Applicant’s observation that Freeman itself does not store the input document as a data neuron does not overcome the present rejection. The present rejection relies on the combined teachings of Freeman and Shinn. Freeman supplies the similarity-based/topological determination of the search location, while Shinn supplies insertion of the received input itself as graph-resident data and the sequential first/rest-node insertion structure, as explained in detail in the rejection of claim 1 above.
Applicant also argues that Hu’s disclosed disclosure of dynamically updating a knowledge graph does not teach modifying the graph based on the temporal order of insertion or based on the location of the last insertion. See Remarks, pg. 11-12 of 13.
This argument has been considered. The present rejection of claim 1 does not rely on Hu to teach the temporal-order/last-insertion limitation. Rather, Shinn is relied upon for the sequential insertion structure discussed above, including previously inserted and newly inserted input is inserted. See Shinn ¶¶ [0100]-[0101], [0104]-[0106]. Thus, Applicant’s argument concerning the scope of Hu’s graph-update teaching does not overcome the present rejection of claim 1.
The additional references applied to the dependent claims are relied upon only for the additional limitations identified in their respective rejections. In particular, Chakraborty is relied upon for retrieval-driven memory-strength reinforcement and modification of memory-network search organization; Chitturi is relied upon for the claimed data-limit functionality; Esser is relied upon for the recited temporal relationships, temporal-distance criteria, and sentence/path functionality; and Hu is relied upon for the additional limitations of the dependent claims to which Hu is applied. The particular teachings and reasons for combination are set forth in the respective rejections below.
Accordingly, although Applicant’s arguments are persuasive as to the previous Freeman-Esser-Hu rejection, they do not overcome the new grounds of rejection set forth in this Office action.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-4, 12, 21, and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Richard T. Freeman et al. (Adaptive topological tree structure for document organisation and visualisation) in view of Hong Shik Shinn et al. (US20190065625A1).
Regarding claim 1, Freeman in view of Shinn, teach a computing entity comprising a memory system and one or more processors communicatively coupled to the memory system, the one or more processors configured to:
“receive one or more storage parameters, the one or more storage parameters comprising input data and a set of one or more input features associated with the input data;” – Freeman teaches this limitation. Freeman teaches receiving and processing input documents represented by document vectors containing term features. Freeman explains that:
“a document is stored as
[
f
1
,
…
,
f
N
]
, where
f
i
is the frequency of term
i
” (Freeman, pg. 1260, § 3.2)
And that a document vector may be stored as vocabulary-term/weight pairs. Freeman further uses a term-frequency/inverse-document-frequency weighting scheme for the terms associated with each document:
“The typical information retrieval weighting scheme for terms (term frequency multiplied by the inverse document frequency (Salton, 1989)) is used.” (Freeman, pg. 1260, § 3.2)
Thus, Freeman’s input document/document vector reads on the claimed input data, and the corresponding weighted terms of the document vector read on the claimed set of one or more input features associated with the input data.
“determine a store procedure cue neuron search location in a neural memory network (NoK) ” – Freeman teaches this limitation in part. Freeman teaches an Adaptive Topological Tree Structure (ATTS) comprising adaptive self-organizing chains of nodes and teaches searching the nodes for a best-matching unit based on the input-document vector:
“There are two major steps in the GC training: search for the best matching unit and update of the winner and its neighbourhood.” (Freeman, pg. 1261, § 3.3)
Freeman further teaches that an input document vector x is mapped to a chain containing n nodes, each node having a weight vector
w
j
:
“At time
t
, an input document vector x is mapped to a chain consisting of n nodes. The weight vector of node
j
,
w
j
” (Freeman, pg. 1261, § 3.3)
Thus, Freeman teaches determining a search location corresponding to a node/neuron in a neural topological network. Freeman does not expressly teach that the location is determined for purposes of inserting the input data itself as a data neuron.
“(i) the store procedure cue neuron search location corresponds to a cue neuron of a set of one or more cue neurons,” – Freeman teaches this limitation. Freeman searches among the nodes of the self-organizing chain and identifies a winning or best-matching node:
“The best matching unit c(x) is the node with maximum
S
d
o
t
amongst all the nodes with respect to the document x(t)” (Freeman, pg. 1261, § 3.3)
Freeman expresses the winner as:
“
c
=
a
r
g
m
a
x
j
S
d
o
t
x
t
,
w
j
)” (Freeman, pg. 1261, Eq. 4)
Under the broadest reasonable interpretation, Freeman’s searchable ATTS nodes read on the claimed set of one or more cue neurons, and the winning node reads on the claimed store procedure cue neuron search location.
”(ii) the set of one or more cue neurons are associated with a set of one or more matching features of ” – Freeman teaches this limitation in part. Freeman’s nodes represent clusters having associated term features and documents. Freeman teaches recording associations and their frequencies based on:
“the document vectors belonging to a particular cluster.” (Freeman, pg. 1264, § 3.6)
Freeman further teaches that similar documents are located close to the winning leaf node and may be retrieved from the corresponding nodes/cluster:
“The topology ensures that similar documents are located close to the winning leaf node. Hence, these nodes can also be used to return relevant documents, as well as those on the winning path. This is a key property of using the topological trees.” (Freeman, pg. 1265, § 3.6)
Freeman additionally teaches that, when a node spans a child chain:
“all documents mapped to the node form a new subset of documents” (Freeman, pg. 1263, § 3.5)
Thus, Freeman teaches cue-like nodes associated with matching features of corresponding underlying documents. Freeman does not expressly teach that the underlying documents themselves comprise data neurons within the NoK.
“(iii) the store procedure cue neuron search location is determined based on the cue neuron comprising a similarity to the set of one or more input features associated with the input data that is highest among the set of one or more cue neurons;” – Freeman teaches this limitation. Freeman compares the input-document vector x with each node weight vector
w
j
, using a modified dot product
S
d
o
t
(
x
(
t
)
,
w
j
), and selects the node having the maximum similarity:
“The best matching unit c(x) is the node with maximum
S
d
o
t
amongst all the nodes with respect to the document x(t)” (Freeman, pg. 1261, § 3.3)
Thus, Freeman expressly teaches determining the search location based on the node having the highest similarity to the features of the input data.
“insert ” – Freeman teaches this limitation in part. Freeman teaches adaptive node insertion within the network and teaches inserting new nodes in proximity to an identified active node:
“Once a chain is judged to have sufficiently converged via the stabilisation of AvgSim, a new node is inserted.” (Freeman, pg. 1262, § 3.4)
Freeman teaches that the new node is inserted:
“The new node is inserted in the proximity of the node with the largest activity, to its left or right, depending on which leads to a higher AvgSim.” (Freeman, pg. 1262, § 3.4)
Thus, Freeman teaches insertion of a node based on characteristics of an identified node/location. However, Freeman’s inserted node is a newly generated topic/prototype node having interpolated or extrapolated weights, rather than a data neuron comprising the input data. Freeman therefore does not teach the struck portion of this limitation.
“” – Freeman does not teach this limitation.
“and modify the NoK based on a pattern of a search for the store procedure cue neuron search location ” – Freeman teaches this limitation in part. Freeman expressly teaches modifying the neural network based on the best-matching-unit search. Freeman states that, once the winner node c(x) is found:
“its weight and the weights of its neighbourhood are updated” (Freeman, pg. 1261, § 3.3)
Freeman also expressly identifies the search trajectory through its hierarchical network. The search begins at the root chain, proceeds through the child chain generated from the winning root node, and continues through descendant chains until a winning leaf node is identified. Freeman states:
“This path through the tree can be referred to as the winning path.” (Freeman, pg. 1264-1265, § 3.6)
Thus, Freeman teaches modifying the NoK based on the pattern/result of the search for the winning cue-neuron location. Freeman does not teach modifying the NoK additionally based on a temporal order of the data neuron corresponding to the location of last insertion.
Freeman does not teach these limitations and/or portions of:
“”
“”
“”
“”
“”
Shinn, however, teaches these limitations and/or portions of:
“to insert the input data, wherein:” – Shinn teaches this portion. Shinn teaches receiving an input sentence and updating a knowledge store by inserting the input sentence into a memory graph. In particular, Shinn teaches:
“the input sentence is inserted into the knowledge store.” (Shinn, pg. 9, ¶ [0098])
Shinn further teaches:
“nodes corresponding to the input sentence are inserted into a memory graph.” (Shinn, pg. 9, ¶ [0098])
Shinn additionally teaches that:
“a new input sentence tree is inserted into a memory graph data structure at a first node.” (Shinn, pg. 10, ¶ [0104])
Thus, unlike Freeman’s insertion of a newly generated topic/prototype node, Shinn expressly teaches inserting received input data into the memory graph.
“a set of one or more data neurons within the NoK and” – Shinn expressly characterizes noes of its memory graph as data nodes:
“relevant data nodes are identified.” (Shinn, pg. 7, ¶ [0083])
Shinn further teaches that the knowledge stores conversation history and is implemented using a memory graph data structure:
“data nodes of a knowledge store, such as a memory graph data structure, are identified and made candidates for retrieval.” (Shinn, pg. 7, ¶ [0083])
Shinn’s memory graph further contains nodes corresponding to the actual received input. For example, the sentence:
“Input sentence tree 516 includes the nodes node-number node 517, contains node 519, contents nodes 521, object node 523, ‘apples’ node 524, predicate node 525, ‘love’ node 526, subject node 527, and ‘I’ node 528.” (Shinn, pg. 10, ¶ [0103])
“I love apples” is represented in input sentence tree 516 by node-number node 517, contains node 519, contents node 521, object node 523, “apples” node 524, predicate node 525, “love” node 526, subject node 527, and “I” node 528. Shinn explains that contains node 519 encapsulates the contents associated with the input sentence, while contents node 521 provides access to those contents:
“a data neuron comprising the input data into the NoK based on the store procedure cue neuron search location;” – Shinn teaches that the received sentence is inserted into the knowledge store by inserting nodes corresponding to the input sentence into the memory graph:
“the input sentence is inserted into the knowledge store.” (Shinn, pg. 9, ¶ [0098])
“nodes corresponding to the input sentence are inserted into a memory graph.” (Shinn, pg. 9, ¶ [0098])
Shinn further teaches inserting an input sentence tree into the memory graph at a first node:
“a new input sentence tree is inserted into a memory graph data structure at a first node.” (Shinn, pg. 10, ¶ [0104])
For example, Shinn’s input sentence tree 516 comprises nodes representing the actual input sentence “I love apples” and its constituent subject, predicate, object, and terms. Shinn therefore teaches inserting into the memory graph a data structure comprising the received input data.
Thus, Freeman supplies the determination of the store procedure cue neuron search location, while Shinn supplies insertion of a data neuron comprising the input data, and their combination teaches “insert a data neuron comprising the input data into the NoK based on the store procedure cue neuron search location”.
“temporally link the data neuron with a location of last insertion;” – Shinn teaches a sequential first/rest-node insertion structure in which the graph location used for a subsequent insertion is linked to the structure produced by the preceding insertion. Shinn explains that a first sentence is recorded under a free first node and:
“subsequent sentences are recorded under its corresponding rest node” (Shinn, pg. 9, ¶ [0101])
And when no free first node is available, a new first node and rest node are created under the existing empty rest node, after which the new sentence is inserted using the newly created first node:
“Subsequent new sentences utilize a similar process using the newly created rest node.” (Shinn, pg. 9, ¶ [0101])
Shinn provides a concrete example in which memory graph data structure 530 contains previously inserted input sentence tree 536 corresponding to “I love apples” and newly inserted input sentence tree 540 corresponding to “I like skiing”. Shinn teaches:
“First node 538 and rest node 539 are new nodes appended to rest node 537 …” (Shinn, pg. 10, ¶ [0106])
“Input sentence tree 540 is inserted at first node 538 and a future input sentence can be inserted at rest node 539.” (Shinn, pg. 10, ¶ [0106])
Thus, the location for the newly inserted data structure is structurally linked to the rest-node location associated with the preceding insertion, and the new insertion in turn produces the location used for the next insertion. Under the broadest reasonable interpretation, this teaches temporally linking the data neuron with a location of last insertion.
“and a temporal order of the data neuron corresponding to the location of last insertion.” – Shinn’s first/rest structure expressly distinguishes input data according to the sequence in which the input is inserted into the memory graph. Shinn identifies sentence tree 536 as “previously inserted”, sentence tree 540 as “newly inserted”, and then identifies rest node 539 as the location at which a future input sentence can subsequently be inserted (see Shinn, pg. 10, ¶ [0106]).
Shinn repeats the same ordering in the subsequent embodiment. Memory graph 550 includes previously inserted sentence trees 556 and 560 and newly inserted sentence tree 562. First node 561 and rest node 563 are added to the prior rest node 559; sentence tree 562 is inserted at first node 561; and the next input is thereafter inserted using the newly created rest node 563:
“Memory graph data structure 550 includes previously inserted input sentence tree 556
corresponding to the sentence ‘I love apples,’ previously inserted input sentence tree 560 corresponding to the sentence ‘I like skiing,’ and newly inserted input sentence tree 562 corresponding to the query sentence ‘What fruit do I like?.’ First node 561 and rest node 563 are new nodes added to rest node 559 as a location for saving additional sentences. Input sentence tree 562 is inserted at first node 561 and a future input sentence can be inserted” (Shinn, pg. 10-11, ¶ [0110])
Accordingly, Shinn’s graph structure records the temporal insertion sequence of the data structures (previously inserted, newly inserted, and subsequently/future inserted) and modifies the memory graph by appending the new first/rest structure to the location associated with the prior insertion. Thus, Shinn teaches a temporal order of the data neuron corresponding to the location of last insertion.
It would have been obvious to one of ordinary skill in the art before the effective filing data of the claimed invention to modify Freeman’s adaptive topological tree structure according to Shinn such that the input documents associated with Freeman’s similarity-selected winning nodes are themselves stored as data nodes/data structures within the memory network and are sequentially linked according to Shinn’s first/rest insertion structure.
Freeman teaches organizing documents based on similarity by determining a winning node and associating documents with the corresponding topological location. Shinn teaches the complementary technique of storing the actual received input as data nodes within a memory graph and maintaining the order of successive inputs through a first/rest structure. Shinn expressly teaches that its knowledge store contains data nodes and that nodes corresponding to received input are inserted into and linked within the memory graph.
A person of ordinary skill would have been motivated to incorporate Shinn’s data-node storage and sequential insertion mechanism into Freeman’s similarity-organized structure so that the actual underlying input information, rather than merely a prototype representing that information, would be directly represented and retrievable in the network while retaining both semantic similarity organization and insertion-order relationships. Such a modification would have predictably resulted in Freeman’s similarity search determining the storage location, Shinn’s input-data node being inserted at or associated with that selected location, and the memory network being modified according to both Freeman’s search pattern and Shinn’s temporal insertion sequence.
Regarding claim 2, Freeman in view of Shinn, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“determine that the input data is salient with respect to data in proximity to the store procedure cue neuron search location based on a saliency threshold value;” – Freeman teaches this limitation. As discussed with respect to claim 1, Freeman maps an input document vector to a chain of nodes, determines a winning/best-matching node based on similarity, and updates the winner and its neighborhood. Thus, Freeman’s winning node corresponds to the claimed store procedure cue neuron search location, and documents clustered to the winning node and neighboring nodes correspond to data in proximity to the store procedure cue neuron search location. Freeman expressly evaluates the similarity of documents associated with the nodes using an average-similarity measure (“AvgSim”). Freeman states:
“A measure termed average similarity (AvgSim) is recorded during the training” (Freeman, pg. 1262, § 3.4)
and defines AvgSim using the similarity
S
d
o
t
(
x
i
,
w
j
) between documents
x
i
clustered to node
j
and the weight vector
w
j
of that node. Freeman further teaches:
“Once a chain is judged to have sufficiently converged via the stabilisation of AvgSim, a new node is inserted.” (Freeman, pg. 1262, § 3.4)
Freeman therefore evaluates whether input/document data is sufficiently represented relative to the data clustered around the corresponding node using a similarity-based criterion.
Freeman additionally expressly teaches, in the same document-organization context, a threshold-based determination of whether an input sample is sufficiently represented by the network. In discussing adaptive resonance theory (ART) networks, Freeman explains:
“a vigilance parameter determines if the network represents a chosen sample sufficiently well.” (Freeman, pg. 1257, § 2.1)
Freeman further teaches that, if the sample is sufficiently represented, the existing weights are updated; otherwise:
“a new node is added, until a threshold is reached.” (Freeman, pg. 1257, § 2.1)
Under the broadest reasonable interpretation, the claimed determination that input data is “salient” encompasses determining that the input data is sufficiently distinct from, or insufficiently represented by, the data associated with the selected/proximate network location to warrant a new node. Freeman’s vigilance parameter therefore reads on the claimed saliency threshold value, and Freeman’s similarity comparison to the winning node and its associated/neighboring clustered data reads on determining saliency with respect to data in proximity to the store procedure cue neuron search location.
“and insert the data neuron into the NoK based on a determination that the input data is salient.” – Freeman teaches this limitation. Freeman expressly teaches that the result of the threshold-based representativeness determination controls whether an additional node is introduced into the network. As Freeman explains regarding the vigilance criterion, when the network does not represent the chosen sample sufficiently well:
“a new node is added” (Freeman, pg. 1257, § 2.1)
rather than merely updating the weights of an existing node.
Freeman’s own ATTS similarly performs adaptive node insertion after evaluating the similarity characteristics of the data clustered around the nodes. Freeman teaches:
“Once a chain is judged to have sufficiently converged via the stabilisation of AvgSim, a new node is inserted.” (Freeman, pg. 1262, § 3.4)
The new node is then:
“inserted in the proximity of the node with the largest activity, to its left or right, depending on which leads to a higher AvgSim..” (Freeman, pg. 1262, § 3.4)
Thus, Freeman teaches using a similarity/significance determination concerning the input and the locally represented data to determine whether to insert a neuron into the network. In the Freeman-Shinn combination set forth with respect to claim 1, the inserted neuron is the Shinn-style data neuron comprising the input data, while Freeman applies the claimed saliency-based criterion controlling whether that neuron is inserted.
Regarding claim 3, Freeman in view of Shinn, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“determine that the input data is not salient with respect to data in proximity to the store procedure cue neuron search location based on a saliency threshold value;” – Freeman teaches this limitation. As discussed with respect to claim 1, Freeman maps an input document vector to a chain of nodes, determines a winning node based on similarity between the input document vector and respective node weight vectors, and associates documents with the winning node and neighboring nodes. Freeman expressly teaches that the topology causes similar documents to be located close to the winning node and observes that documents in neighboring nodes are generally more relevant.
Freeman further describes adaptive resonance theory (ART) networks in which a vigilance parameter determines whether the existing network represents an input sample sufficiently well:
“a vigilance parameter determines if the network represents a chosen sample sufficiently well.” (Freeman, pg. 1257, § 2.1)
Freeman teaches that if the sample is sufficiently represented:
“the weights resonate and are updated” (Freeman, pg. 1257, § 2.1)
otherwise:
“a new node is added” (Freeman, pg. 1257, § 2.1)
Under the broadest reasonable interpretation, an input that is already sufficiently represented by the existing network is not salient relative to the data represented at and near the corresponding winning node because the input does not warrant creation of a distinct new representation. Freeman’s vigilance parameter therefore reads on the claimed saliency threshold value, and the comparison of the input with the winning node and locally associated data reads on determining whether the input data is not salient with respect to data in proximity to the store procedure cue neuron search location.
“increase memory strength of data in proximity to the store procedure cue neuron search location based on a determination that the input data is not salient;” - Freeman teaches this limitation. Freeman’s ART disclosure expressly makes the weight-update operation conditional upon the determination that the sample is sufficiently represented by the existing network. Freeman states:
“If this is the case then the weights resonate and are updated (typically winner-takes-all). Otherwise, a new node is added” (Freeman, pg. 1257, § 2.1)
Freeman’s ATTS likewise teaches that, once the best-matching node c(x) is identified, the strength values associated with the winning node and nearby nodes are adaptively reinforced toward the input:
“Once the winner node c(x) is found its weight and the weights of its neighbourhood are updated” (Freeman, pg. 1261, § 3.3)
Freeman’s Eq. (5) performs the update using the input vector, a learning rate, and a neighborhood function:
“
w
j
t
+
1
=
w
j
t
+
α
t
h
j
,
c
x
t
x
(
t
)
w
j
t
+
α
t
h
j
,
c
x
t
x
(
t
)
” (Freeman, pg. 1261, Eq. (5))
Thus, under the broadest reasonable interpretation, Freeman’s node/feature weights read on the claimed memory strength, and Freeman’s winner and neighborhood read on the claimed data in proximity to the store procedure cue neuron search location. Freeman therefore teaches strengthening/updating the existing local network representation when the input is sufficiently represented, i.e., when the input is determined to be not salient, rather than creating a distinct representation for the input.
“” – Freeman teaches this limitation in part. Freeman’s ATTS performs a “search for the best matching unit”, wherein an input document vector is mapped to a chain of nodes and a winning node is selected based on similarity:
“There are two major steps in the GC training: search for the best matching unit and update of the winner and its neighbourhood.” (Freeman, pg. 1261, § 3.3)
Freeman further teaches searching through the topological tree until a particular winning leaf node is identified:
“The search through the tree begins at the root chain, where the winning node is identified. The child chain generated from that root node is then searched for the winning node and this process continues for the descendent chains until the winning leaf node is identified.” (Freeman, pg. 1265, § 3.6)
Freeman refers to this traversal as the “winning path”.
Thus, Freeman’s identified best-matching/winning node reads on the claimed store procedure cue neuron search location. Freeman does not teach “temporally link” that location “with the location of last insertion”.
Freeman does not teach these limitations and/or portions of:
“and temporally link … with the location of last insertion.”
Shinn, however, teaches these limitations and/or portions of:
“and temporally link … with the location of last insertion.” – Shinn teaches a sequential memory-graph structure in which locations for successively inserted input-data structures are linked through first and rest nodes. Shinn teaches that a new input sentence tree is inserted into the memory graph at a first node, that a first node may be nested under a rest node, and that a first node may be created under a free rest node:
“a new input sentence tree is inserted into a memory graph data structure at a first node. … a new sentence tree is inserted into a memory graph data structure at the first available first node. … a first node is nested under a rest node. … a first node is created under a free rest node.” (Shinn, pg. 10, ¶ [0104])
More particularly, Shinn describes memory graph 530 as including a “previously inserted input sentence tree 536” corresponding to “I love apples” and a “newly inserted input sentence tree 540” corresponding to “I like skiing”. Shinn then teaches:
“First node 538 and rest node 539 are new nodes appended to rest node 537 as a location for inserting additional sentences.” (Shinn, pg. 10, ¶ [0106])
Shinn further teaches:
“Input sentence tree 540 is inserted at first node 538 and a future input sentence can be inserted at rest node 539.” (Shinn, pg. 10, ¶ [0106])
Thus, Shinn teaches linking a current insertion location to the location associated with the preceding insertion, with the first/rest structure preserving the successive order of previously inserted, newly inserted, and future input data.
In the Freeman-Shinn combination set forth with respect to claim 1, Freeman supplies the store procedure cue neuron search location, and the input data is inserted at or associated with that Freeman-selected location using Shinn’s graph-resident storage technique. Applying Shinn’s disclosed successive-insertion linkage to that location would result in the current Freeman-selected store procedure cue neuron search location being linked to the location associated with the preceding insertion, thereby teaching “temporally link the store procedure cue neuron search location with the location of last insertion”.
Regarding claim 4, Freeman in view of Shinn, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“perform the search for the store procedure cue neuron search location until a search limit value is reached;” – Freeman teaches this limitation. As discussed with respect to claim 1, Freeman repeatedly performs a search for a winning/best-matching node by selecting an input document and determining the node having the greatest similarity to the input document. Freeman’s growing-chain algorithm expressly bounds this search-and-growth process by a maximum value
n
m
a
x
. Freeman teaches that
n
m
a
x
is:
“the maximum number of nodes used for the validation.” (Freeman, pg. 1262, § 3.4)
Freeman’s Table 1 further provides the following outer-loop structure:
“For n=2: n_max” (Freeman, pg. 1262, Table 1)
within which the algorithm repeatedly selects a document, “Find[s] winning node via Eq. (4)”, updates the chain weights, and calculates AvgSim. Freeman then records the validation result and inserts a new node before continuing the process. (Freeman, pg. 1261-1262, §§ 3.3-3.4, Table 1)
Freeman further expressly states:
“The chain is allowed to grow to a specified maximum number of nodes, e.g.
n
m
a
x
.” (Freeman, pg. 1262, § 3.4)
Thus, Freeman repeatedly performs the best-matching-node search used to identify the claimed store procedure cue neuron search location, while limiting that search/growth process according to the specified maximum value
n
m
a
x
. Under the broadest reasonable interpretation, Freeman’s
n
m
a
x
therefore reads on the claimed search limit value.
“insert ” – Freeman teaches this limitation in part. Freeman teaches determining a preferred location for insertion based on the nodes identified during the best-matching-unit search process. Freeman monitors each node’s activity using a counter that records the node’s winning times. Once the chain has sufficiently converged, Freeman teaches:
“The new node is inserted in the proximity of the node with the largest activity” (Freeman, pg. 1262, § 3.4)
and further determines whether to insert the new node to the left or right of that node:
“depending on which leads to a higher AvgSim.” (Freeman, pg. 1262, § 3.4)
Freeman likewise initializes the new node using interpolation or extrapolation depending on which alternative increases AvgSim more. (Freeman, pg. 1262, § 3.4).
Thus, Freeman evaluates locations encountered through the repeated winner-search process and inserts a node proximate to the node having the greatest accumulated winning activity, at the side producing the higher average similarity. Under the broadest reasonable interpretation, this teaches inserting a node in a best location of the NoK encountered during the search.
Freeman does not, however, teach that the node inserted at that location is the claimed data neuron comprising the input data, as discussed with respect to claim 1.
Freeman does not teach these limitations and/or portions of:
“the data neuron”
“and temporally link the data neuron to the location of last insertion.”
Shinn, however, teaches these limitations and/or portions of:
“the data neuron” – Shinn teaches inserting received input data itself into a memory graph. Shinn teaches that the received sentence “I love apples” is represented by input sentence tree 516 comprising nodes corresponding to the sentence contents, subject, predicate, object, and actual terms, and further teaches that:
“a new input sentence tree is inserted into a memory graph data structure at a first node.” (Shinn, pg. 10, ¶ [0104])
Thus, as set forth more fully with respect to claim 1, it would have been obvious to employ Shinn’s data-containing node/tree as the node inserted at Freeman’s identified best insertion location, such that the node inserted at Freeman’s best location is the claimed data neuron.
“and temporally link the data neuron to the location of last insertion.” – Shinn expressly describes a memory graph containing a previously inserted input sentence tree 536 and a newly inserted input sentence tree 540. Shinn teaches that new first node 538 and rest node 539 are appended to prior rest node 537 as locations for additional sentences, that sentence tree 540 is inserted at first node 538, and that a future sentence may thereafter be inserted at rest node 539. (Shinn, pg. 10, ¶ [0106]).
Shinn similarly teaches generally that a new sentence tree is inserted at a first node, that a first node may be created under a free rest node, and that subsequent sentences use the newly created rest node. (Shinn, pg. 10, ¶ [0104])
Thus, Shinn’s first/rest structure links the newly inserted data-containing structure to the graph location associated with the preceding insertion and establishes the location for the next insertion. Under the broadest reasonable interpretation, Shinn therefore teaches temporally linking the data neuron to the location of last insertion.
In the Freeman-Shinn combination set forth with respect to claim 1, Freeman supplies the similarity-based search, the bounded search/growth process, and selection of the best insertion location, while Shinn supplies the data-containing neuron. The combination therefore teaches performing the search until a search limit value is reached, inserting the data neuron in a best location encountered during the search, and temporally linking that data neuron to the location of last insertion.
Regarding claim 12, Freeman in view of Shinn, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“mark key data point locations in the NoK;” – Freeman teaches this limitation. Freeman teaches assigning descriptive key terms to the nodes/clusters of its ATTS:
“a number of most descriptive key words have to be assigned to each cluster.” (Freeman, pg. 1264, § 3.6)
Freeman further explains that the resulting labels summarize the content of the clusters and identify the underlying topics at respective locations in the topology.
Under BRI, Freeman’s labeled cluster/node locations read on the claimed key data point locations in the NoK, and assigning the key terms to those locations reads on marking those locations.
“assign a score to a data point at each of the key data point locations;” – Freeman teaches this limitation. For the terms associated with each node
j
, Freeman expressly calculates a term score:
“The term score
t
s
j
i
is calculated for the remaining terms I in node j …” (Freeman, pg. 1264, § 3.6)
Freeman then ranks the terms using frequency of occurrence and the term score
t
s
i
as a primary key.
Thus, Freeman assigns a score to respective term/data points associated with each identified node/cluster location.
“and fetch one or more top key data points based on a search limit and respective scores of the one or more top key data points.” – Freeman teaches this limitation. After calculating and ranking the term scores, Freeman expressly teaches:
“The first few terms (e.g. 5) with the highest ranking values are selected to label the cluster.” (Freeman, pg. 1264, § 3.6)
Thus, Freeman selects the top-ranked data points based on their respective scores and limits the selection to a specified number, e.g., the first five terms. Under BRI, the specified number of terms reads on the claimed search limit, and the selected highest-ranked terms read on the claimed one or more top key data points.
Regarding newly added claims 21 and 22
Claims 21 and 22 are rejected under 35 U.S.C. § 103 as being unpatentable over Freeman in view of Shinn for substantially the same reasons set forth above with respect to claim 1.
Regarding claim 21
Claim 21 recites substantially the same storage, search, data-neuron insertion, temporal-linking, and NoK-modification operations of claim 1 in computer-implemented method form. Freeman expressly describes an implemented application package comprising clustering, organization, search, and related functionality, and Shinn likewise teaches that its disclosed processes may be implemented on a “computer programmed system”, including remote computer systems. Thus, the Freeman-Shinn teachings and rationale applied to the substantive limitations of claim 1 are equally applicable to the corresponding method steps of claim 21.
Regarding claim 22
Claim 22 recites substantially the same operations as claim 1 in the form of processor-executable instructions stored on the one or more non-transitory computer-readable media. Shinn expressly teaches a system comprising:
“a processor; and a memory coupled with the processor” (Shinn, pg. 21, claim 20)
where the memory provides the processor with instructions that, when executed, cause the processor to perform operations including receiving input, processing the input, and storing the resulting information in an artificial-intelligence memory graph.
Thus, Shinn teaches the processor-executable-instruction/storage-medium implementation recited by claim 22, while the substantive operations performed by those instructions are taught or rendered obvious by Freeman in view of Shinn for the reasons set forth with respect to claim 1.
Accordingly, Freeman in view of Shinn renders claims 21 and 22 obvious.
Claims 5 and 7 are rejected under 35 U.S.C. 103 as being unpatentable over Freeman in view of Shinn and further in view of Prabuddha Chakraborty et al. (Neural Storage: A New Paradigm of Elastic Memory).
Regarding claim 5, Freeman in view of Shinn and further in view of Chakraborty, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“receive one or more retrieval parameters, the one or more retrieval parameters comprising a set of one or more search features associated with a data request;” – Freeman teaches this limitation. Freeman teaches receiving a query document and using features of that query to search the ATTS for related documents. Freeman explains that:
“each document in the test datasets were used as a query to search for related documents of the same topic.” (Freeman, pg. 1269, § 4)
Freeman further teaches matching the query document against the network to identify the most similar nodes and clusters.
Freeman’s query document therefore reads on the claimed data request, and the term/vector features of the query document used to perform the similarity search read on the claimed set of one or more search features associated with the data request.
“fetch data based on a determination that a retrieve procedure cue neuron search location is in proximity to data matching the set of one or more search features;” – Freeman teaches this limitation. Freeman expressly teaches identifying a winning leaf node for a query and retrieving relevant documents located at or near that node:
“In order to search for related documents in a topological tree, the leaf node that best matches the query needs to be identified” (Freeman, pg. 1264, § 3.6)
Freeman further teaches that the search begins at the root chain, identifies a winning node, proceeds through corresponding child chains, and continues until the winning leaf node is identified. Freeman then explains:
“The topology ensures that similar documents are located close to the winning leaf node. Hence, these nodes can also be used to return relevant documents, as well as those on the winning path. This is a key property of using the topological trees.” (Freeman, pg. 1265, § 3.6)
Freeman further teaches that the most similar clusters to the query document are identified and:
“all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Thus, Freeman’s winning leaf node reads on the claimed retrieved procedure cue neuron search location, and Freeman teaches/returns relevant documents based on their location in or near the winning node and their similarity to the search features of the query.
“” – Freeman does not teach this limitation.
“” – Freeman does not teach this limitation.
Freeman does not teach these limitations and/or portions of:
“increase memory strength of the fetched data;”
“and modify the NoK based on another pattern of search for the retrieve procedure cue neuron search location.”
Chakraborty, however, teaches these limitations and/or portions of:
“increase memory strength of the fetched data;” – Chakraborty’s Retrieve Algorithm searches candidate data neurons and, when a target data neuron matches the fine-grained search cues, invokes the REACTION procedure to reflect the successful retrieval and then returns the target neuron’s data:
“
R
E
A
C
T
I
O
N
(
M
H
,
T
a
r
g
e
t
D
N
,
P
a
t
h
,
η
,
1
,
C
,
u
p
,
k
)
R
e
t
D
a
t
a
=
T
a
r
g
e
t
D
N
→
D
a
t
a
” (Chakraborty, Algorithm 2 Retrieve, lines 23-24)
The corresponding Reaction Algorithm expressly provides, upon a successful match:
“
I
n
c
r
e
a
s
e
_
M
e
m
_
S
t
r
T
a
r
g
e
t
D
N
” (Chakraborty, Algorithm 4 Reaction, line 8)
and updates the memory search order:
“
U
P
D
A
T
E
_
M
E
M
O
R
Y
_
S
E
A
R
C
H
_
O
R
D
E
R
(
M
H
” (Chakraborty, Algorithm 4 Reaction, line 12)
Neural Storage’s worked example confirms the operation. After DN1 matches and is successfully retrieved, “memory strength of DN1 is restored back to 100”. Later retrievals likewise restore the successfully matched DN0 to 100:
“The fine-grained cue and data feature of DN1 matches leading to a successful retrieval. The weight of <Wolf, DN1> association is increased. Also, note that the memory strength of all data neurons decreases and memory strength of DN1 is restored back to 100” (Chakraborty, § Appendix D)
“The next operation is of type retrieve … Memory strength of DN0 is restored back to 100 and the memory strengths of all remaining data neurons decreases.” (Chakraborty, § Appendix D)
Thus, Neural Storage directly teaches increasing the memory strength of the data neuron whose data is fetched during retrieval.
“and modify the NoK based on another pattern of search for the retrieve procedure cue neuron search location.” – Chakraborty’s retrieval operation uses search cues to obtain a search-order list and traverses candidate paths/data neurons according to that order. When a candidate successfully or unsuccessfully matches, the Reaction procedure updates the NMN using reinforcement learning:
“Using the Reaction procedure (Algo. 4), the NMN is updated using reinforcement learning to reflect a success (line 26). Otherwise, the procedure (Algo. 4) updates the Reaction NMN using reinforcement learning considering a failed merge attempt (line 28).” Chakraborty, § Appendix B-1)
More particularly, Chakraborty explains that the Reaction procedure is used during store and retrieve operations:
“The REACTION (Algo. 4) subroutine is a reinforcement learning guided procedure which is used during the store and retrieve operation for creating new associations, changing association weights, and updating search order for cues.” (Chakraborty, § Appendix B-1)
Chakraborty further teaches calculating the search order by selecting paths to each data neuron according to their highest average association strength, with the NewSearchOrder replacing the previous search order:
“the paths to each data-neuron D, with the highest average association strength is selected and stored in decreasing order of average association strength. The NewSearchOrder replaces the previous search order for the MH” (Chakraborty, Appendix B-3-C)
Thus, Chakraborty teaches modifying the neural memory network according to the pattern and outcome of the retrieval search. In the Freeman-Shinn-Neural Storage combination, Freeman supplies the claimed retrieve procedure cue neuron search location, while Neural Storage supplies modification of the NoK based on the separate retrieval-search pattern used to reach and evaluate that location.
It would have been obvious to one of ordinary skill in the art to modify the Freeman-Shinn retrieval system according to Chakraborty’s Neural Storage so that successful retrieval reinforces the fetched data and updates the network’s search organization. Chakraborty expressly teaches that its learnable memory parameters are updated based on feedback from memory operations and identifies a learning objective of:
“Increase memory search speed by learning the right NMN organization given current circumstances and access-pattern bias.” (Chakraborty, § III-C-2)
A POSITA would therefore have been motivated to apply Chakraborty’s retrieval-driven reinforcement and search-order updating to Freeman’s similarity-based retrieval network so that successful retrievals make relevant stored data and paths more prominent for subsequent searches, with the predictable result of increasing the memory strength of fetched data and adapting the NoK based on retrieval-search behavior.
Regarding claim 7, Freeman in view of Shinn and further in view of Chakraborty, teach the computing entity of claim 5, wherein:
“” – Freeman does not teach this limitation.
“and ii the one or more processors are further configured to select a next candidate cue neuron from the set of one or more cue neurons for a potential next search step based on NoK connectivity ” – Freeman teaches this limitation in part. Freeman teaches a hierarchical topological search in which the search proceeds from one selected node to a connected portion of the network for a subsequent search step. Freeman expressly teaches:
“The search through the tree begins at the root chain, where the winning node is identified. The child chain generated from that root node is then searched for the winning node and this process continues for the descendent chains until the winning leaf node is identified.” (Freeman, pg. 1265, § 3.6)
Freeman further teaches that this traversal is the “winning path” and that, rather than searching all nodes:
“the topology property has been used in this search mechanism.” (Freeman, pg. 1265, § 3.6)
Thus, after identifying a winning node at one level, Freeman selects and searches the child chain generated from what the winning node at the next level. Accordingly, Freeman’s nodes read on the claimed set of one or more cue neurons, the winning node is the subsequently searched child chain reads on a next candidate cue neuron for a potential next search step, and Freeman’s parent/child topological relationship reads on selection based on NoK connectivity.
Freeman further teaches that this process continues until the winning leaf node is identified and that:
“The topology ensures that similar documents are located close to the winning leaf node. Hence, these nodes can also be used to return relevant documents, as well as those on the winning path” (Freeman, pg. 1265, § 3.6)
Accordingly, Freeman expressly teaches both (1) continuing the successive connected-node search until the winning leaf node is identified and (2) similar/relevant documents being located close to that winning leaf node. Under the broadest reasonable interpretation, Freeman therefore teaches or suggests performing the successive search steps until the data matching the query/search feature is in proximity to the resulting retrieve procedure cue neuron search location.
Freeman does not teach these limitations and/or portions of:
“(i) the one or more retrieval parameters comprise gist information associated with the data request”
“… and the gist information …”
Shinn, however, teaches these limitations and/or portions of:
“(i) the one or more retrieval parameters comprise gist information associated with the data request” – Shinn teaches receiving a natural-language query and processing the query to identify structured information characterizing its content and meaning. Shinn explains:
“The natural language input is processed to classify components of the natural language input. For example, the voice is processed into a structured format that identifies a subject, object, and predicate” (Shinn, pg. 2. ¶ [0034])
“one or more lemma are identified as well as a performative classification, a sender of the input, and a receiver of the input.” (Shinn, pg. 2. ¶ [0034])
Shinn additionally teaches that receiving sentence input may include:
“identified parts of speech, lemmas (e.g., subject, predicate, and/or object lemma), coreference relationships, and performative classification, among others.” (Shinn, pg. 7, ¶ [0079])
Under the broadest reasonable interpretation, Shinn’s classified semantic information (including subject, object, predicate, lemmas, parts of speech, coreference relationships, and performative classification) reads on the claimed gist information associated with the data request, because the information represents structured semantic/contextual information derived from the query.
“… and the gist information …” – Shinn teaches using the gist information to identify relevant candidate data nodes during a memory-graph search:
“A starting node of an artificial intelligence memory graph data structure is selected to begin a search for one or more supporting knowledge data nodes associated with the classified components,” (Shinn, pg. 2, ¶ [0034])
Shinn further teaches:
“Starting at the starting node, the artificial intelligence memory graph data structure is searched using a lexical database to identify the one or more supporting knowledge data nodes.” (Shinn, pg. 2, ¶ [0034])
Shinn also expressly claims that:
“the one or more supporting knowledge data nodes are searched for by comparing subject predicate-object lemma, parts of speech, and coreference results between the classified components and one or more data nodes of the short-term artificial intelligence memory graph data structure or the long-term artificial intelligence memory graph data structure.” (Shinn, pg. 21, claim 10)
Thus, Shinn teaches using the semantic/classified information derived from the query to determine which memory-graph data nodes are relevant to the query.
Freeman teaches a hierarchical retrieval search that proceeds from a winning root node through descendent chains to a winning leaf node, with documents topologically close to the winning leaf node and winning path used to return relevant documents. Shinn, in turn, teaches processing a received query using semantic information including parts of speech, subject/predicate/object lemmas, coreference relationships, and performative classification, and using the processed query to identify relevant data nodes in a memory-graph knowledge store.
A POSITA would have been motivated to apply Shinn’s semantic query-processing and relevance-identification teachings to Freeman’s hierarchical retrieval search so that the content and semantic characteristics of the received query could be used to identify and evaluate relevant stored data encountered during the hierarchical search, thereby predictably improving the relevance or retrieved results.
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Freeman in view of Shinn further in view of Chakraborty and further in view of Suresh Chitturi et al. (US20100198854A1).
Regarding claim 8, Freeman in view of Shinn further in view of Chakraborty and further in view of Chitturi, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“receive one or more retrieval parameters comprising a set of one or more search features, a search limit value, ” – Freeman teaches this limitation in part. Freeman teaches using a query document to search for related documents. In its retrieval experiments, Freeman states that:
“each document in the test datasets were used as a query to search for related documents of the same topic.” (Freeman, pg. 1269, § 4)
Freeman further performs retrieval using the features of the query document, consistent with its vector-space representation of documents.
Freeman also expressly uses a retrieval-search scope value. Freeman states:
“For the SOM and ATTS the best number of leaf clusters is tested from 1 node to 8 nodes.” (Freeman, pg. 1269, § 4)
Freeman then finds the most similar clusters to the query document and ranks the documents in those clusters.
Thus, Freeman teaches search features associated with a query and a value limiting the number of leaf clusters considered during retrieval, which reads on the claimed search limit value. Freeman does not expressly teach the claimed gist information, as discussed with respect to claim 7.
With respect to the claimed data limit value, Freeman further limits which retrieved documents are included in the search results by applying a cutoff:
“Those falling below a cut-off similarity to query threshold were ignored (e.g. 0.025)” (Freeman, pg. 1269, § 4)
This expressly teaches a retrieval parameter that limits the data included in the result set. Under the broadest reasonable interpretation, the cutoff value reads on a data limit value because it limits which data are included in the retrieved results.
“generate a results data structure;” – Freeman teaches this limitation. Freeman expressly generates search results by selecting best-matching leaf clusters:
“The search results are based on selecting the best matching leaf clusters in each structure.” (Freeman, pg. 1269, § 4)
Freeman then states:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Under the broadest reasonable interpretation, Freeman’s organized/ranked collection of retrieved documents reads on the claimed results data structure.
“determine that a multi-retrieve procedure cue neuron search location is in proximity to matching data that matches the set of one or more search features;” – Freeman teaches this limitation. Freeman searches the ATTS for a winning leaf node corresponding to the query, as discussed with respect to claims 5 and 7. Freeman then expressly teaches:
“The topology ensures that similar documents are located close to the winning leaf node.” (Freeman, pg. 1265, § 3.6)
Freeman further teaches that nearby nodes may be used to return relevant documents. Thus, Freeman’s winning leaf node reads on the claimed multi-retrieve procedure cue neuron search location, and Freeman expressly teaches matching/similar data being located in proximity to that search location.
Freeman’s experimental retrieval disclosure further states that:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
after which the documents in those clusters are ranked by relevance.
“fetch and append the matching data to the results data structure;” – Freeman teaches this limitation. Freeman retrieves documents from the best-matching clusters and forms the search results from those retrieved documents:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Documents below the cutoff similarity threshold are excluded.
Thus, Freeman teaches fetching matching documents from the selected clusters and including those documents in the generated ranked search results. Under broadest reasonable interpretation, including the retrieved matching documents in the search-result collection reads on appending the matching data to the results data structure.
“” – Freeman does not teach this limitation.
“”
“and return the results data structure comprising a plurality of data elements as output.” – Freeman teaches this limitation. Freeman expressly defines B as:
“the total number of documents retrieved from the search.” (Freeman, pg. 1269, § 4)
And teaches ranking the multiple documents returned from the most similar clusters:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Freeman does not teach these limitations and/or portions of:
“gist information, and a data limit value;”
“increase memory strength of the matching data appended to the results data structure;”
“modify the NoK based on”
Shinn, however, teaches these limitations and/or portions of:
“gist information” – As discussed with respect to claim 7, Shinn processes a natural-language input into classified semantic information including subject, object, predicate, lemmas, performative classification, sender, and receive, and uses those classified components to search its memory graph:
“The natural language input is processed to classify components of the natural language input. For example, the voice is processed into a structured format that identifies a subject, object, and predicate” (Shinn, pg. 2. ¶ [0034])
“one or more lemma are identified as well as a performative classification, a sender of the input, and a receiver of the input.” (Shinn, pg. 2. ¶ [0034])
“identified parts of speech, lemmas (e.g., subject, predicate, and/or object lemma), coreference relationships, and performative classification, among others.” (Shinn, pg. 7, ¶ [0079])
“A starting node of an artificial intelligence memory graph data structure is selected to begin a search for one or more supporting knowledge data nodes associated with the classified components,” (Shinn, pg. 2, ¶ [0034])
Under BRI, this semantic/contextual information associated with the query reads on gist information.
Shinn does not teach these limitations and/or portions of:
“and a data limit value;”
“increase memory strength of the matching data appended to the results data structure;”
“modify the NoK based on”
Chakraborty, however, teaches these limitations and/or portions of:
“increase memory strength of the matching data appended to the results data structure;” – Chakraborty’s Neural Storage Retrieve Algorithm determines whether a target data neuron matches the fine-grained search cues, invokes the Reaction procedure upon a successful match, and then returns the target neuron’s data. The Reaction Algorithm expressly provides:
“
I
n
c
r
e
a
s
e
_
M
e
m
_
S
t
r
(
T
a
r
g
e
t
D
N
)
” (Chakraborty, Algorithm 4 Reaction, line 8)
upon a successful search.
The worked example confirms that successful retrieval stores the matching neuron’s memory strength: following successful retrieval of DN1, its memory strength is “restored back to 100”. (Chakraborty, § Appendix D)
Thus, Freeman supplies fetching and appending the matching data to the multi-result structure, while Chakraborty teaches increasing the memory strength of that successfully retrieved matching data.
“modify the NoK based on” – During retrieval, Chakraborty’s retrieval procedure obtains a search order, traverses candidate data neurons according to that order, evaluates whether each candidate matches the retrieval cues, and invokes a Reaction procedure based on the success or failure of that search. Chakraborty expressly states that, upon a successful match, the Reaction procedure:
“Using the Reaction procedure (Algo. 4), the NMN is updated using reinforcement learning to reflect a success (line 26).” (Chakraborty, § Appendix B-1)
And likewise updates the NMN in response to a failed match:
“Otherwise, the procedure (Algo. 4) updates the Reaction NMN using reinforcement learning considering a failed merge attempt (line 28).” (Chakraborty, § Appendix B-1)
Chakraborty further teaches that the Reaction procedure is used during retrieval for:
“creating new associations, changing association weights, and updating search order for cues.” (Chakraborty, § Appendix B-1)
and identifies the Path input as the path to the target data neuron and the flag input as indicating search success or failure. Algorithm 4 correspondingly increases or decreases association weights along the search path and may update the memory search order. (Chakraborty, Algorithm 4, lines 4, 6, and 12)
Chakraborty additionally teaches replacing the previous order with a new search order based on paths to the data neurons and their average association strengths. (Chakraborty, § Appendix B-1)
Thus, Chakraborty expressly teaches modifying the neural memory network based on the path and outcome of a retrieval search. Freeman supplies the claimed multi-result retrieval context and the multi-retrieve procedure cue neuron search location. In combination, these teachings satisfy modifying the NoK based on another pattern of search for that multi-retrieve procedure cue neuron search location.
Chakraborty does not teach these limitations and/or portions of:
“and a data limit value;”
Chitturi, however, teaches these limitations and/or portions of:
“and a data limit value;” – Chitturi’s received search request includes a maxResults attribute. Its example provides:
“<Keyword maxResults-"50">example </Keyword>” (Chitturi, pg. 5, ¶ [0051])
and similarly provides an XQuery having maxResults=”50”. Chitturi explains that maxResults:
“an attribute ‘maxResults’ which defines the maximum number of results that can be returned” (Chitturi, pg. 5, ¶ [0055])
And that once the maximum is reached, the search cuts off or ignores the remaining response
“In the example above, the maximum result is defined as 50 indicating that at most fifty results can be returned at which point the search cuts off, discards or ignores the remaining response.” (Chitturi, pg. 5, ¶ [0055])
Thus, Chituri’s maxResults value is a retrieval parameter defining the maximum quantity of data returned by the search and directly reads on the claimed data limit value.
It would have been obvious to one of ordinary skill in the art to modify the Freeman-Shinn retrieval system according to Neural Storage and Chitturi to provide adaptive retrieval-based learning while also controlling the quantity of data returned by a multi-result search. Chakraborty teaches updating learnable memory parameters based on feedback generated during memory operations, including reinforcement of successfully retrieved data and modification of the memory search organization, with the objective of increasing memory-search speed according to current circumstances and access-pattern bias. Chitturi further teaches including a maxResults parameter in a search request to define the maximum number of results that may be returned and terminating or disregarding additional results once that limit is reached.
A POSITA would have been motivated to apply Chakraborty’s retrieval-driven reinforcement and adaptive search-order updating to Freeman’s similarity-based retrieval system so that successfully retrieved data and corresponding search paths become more prominent for subsequent retrievals, thereby predictably improving retrieval efficiency. A POSITA would further have been motivated to incorporate Chitturi’s maximum-results parameter into that multi-result retrieval system to control the volume of returned data and avoid unnecessary retrieval and processing beyond a requested result limit. The combined modification would predictably provide a retrieval system that both adapts its memory organization in response to successful retrievals and limits the quantity of matching data returned according to a specified data limit value.
Claims 9 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Freeman in view of Shinn further in view of Chakraborty and further in view of Stefan Esser (Multi-Dimensional Event Data in Graph Databases Representing, Querying, and Process Mining) .
Regarding claim 9, Freeman in view of Shinn further in view of Chakraborty and further in view of Esser, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“receive one or more retrieval parameters comprising a set of one or more search features, a search limit value, – Freeman teaches this limitation in part. Freeman teaches using each test document as a query to search for related documents and matching the query to the most similar nodes. Freeman further teaches limiting the retrieval search by varying the number of leaf clusters considered:
“For the SOM and ATTS the best number of leaf clusters is tested from 1 node to 8 nodes.” (Freeman, pg. 1269, § 4)
Freeman then finds the most similar clusters to the query document and ranks the documents in those clusters.
Thus, Freeman teaches retrieval parameters comprising search features of the query and a search limit value defining the number of leaf clusters considered during the search.
“generate a results data structure;” – Freeman teaches this limitation. Freeman teaches generating search results from the selected best-matching leaf clusters:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Under the broadest reasonable interpretation, Freeman’s ranked collection of retrieved documents reads on the claimed results data structure.
“determine that a ” – Freeman teaches this limitation in part. Freeman teaches determining a winning/best-matching leaf node during retrieval:
“the leaf node that best matches the query needs to be identified.” (Freeman, pg. 1264, § 3.6)
Freeman further teaches:
“The topology ensures that similar documents are located close to the winning leaf node.” (Freeman, pg. 1265, § 3.6)
“determine the matching data ” – Freeman teaches this limitation in part. Freeman teaches determining the matching data by finding the most similar clusters to the query and ranks documents according to relevance:
“The search results are based on selecting the best matching leaf clusters in each structure. … The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
”fetch and append the matching data to the results data structure;” – Freeman teaches this limitation. Freeman teaches finding the most similar clusters to the query document and including the documents from those clusters in the ranked search results:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Freeman further excludes documents falling below a similarity cut-off.
Thus, Freeman teaches retrieving matching documents and including them in the search-result collection, which under BRI reads on fetching and appending the matching data to the results data structure.
“” – Freeman does not teach this limitation.
“modify the NoK “ – Freeman teaches this limitation in part. In the retrieval context, Freeman expressly teaches:
“A top-down search … was used to seek for the related documents” (Freeman, pg. 1269, § 4)
Freeman further states:
“For the ATTS and bisecting k-means the search begins with the winning node in the root chain and then proceeds down the hierarchy.” (Freeman, pg. 1269, § 4)
And explains that:
“the selection of relevant clusters will be topologically close to the path from the winning root node to the winning leaf node.” (Freeman, pg. 1269, § 4)
Freeman therefore teaches a distinct retrieval-oriented pattern of search that proceeds through the network from a winning root node toward a winning leaf node, with the resulting path used to identify relevant clusters. Freeman does not expressly teach modifying the NoK based on that retrieval-search pattern, nor does Freeman teach the temporal component of the claimed spatio-temporal cue-neuron search location.
“and return the results data structure comprising ” – Freeman teaches this limitation in part. Freeman teaches returning the results data structure as output and expressly defines B as:
“the total number of documents retrieved from the search.” (Freeman, pg. 1269, § 4)
Freeman further teaches:
“The most similar clusters to query documents are found and all the documents in these cluster are ranked by their relevance.” (Freeman, pg. 1269, § 4)
Freeman’s implemented application also presents the retrieval output to the user; Figure 5 is described as including the query and its:
“respective ranked retrieved documents.” (Freeman, pg. 1265, Fig. 5 text)
Thus, Freeman teaches returning a result collection comprising a plurality of retrieved documents as outputs. Freeman does not expressly characterize those returned data elements as the claimed “sequence of data” in the temporal sense.
Freeman does not teach these limitations and/or portions of:
“gist information, a temporal limit value, and a near or far value”
“spatio-temporal”
“is temporally consistent with the gist information based on the near or far value;”
“increase memory strength of the matching data appended to the results data structure;”
“based on another pattern of search”
“a sequence of data as output.”
Shinn, however, teaches these limitations and/or portions of:
“gist information” – As discussed with respect to claim 7 and 8, Shinn processes a natural-language input into classified semantic information including subject, object, predicate, lemmas, performative classification, sender, and receive, and uses those classified components to search its memory graph:
“The natural language input is processed to classify components of the natural language input. For example, the voice is processed into a structured format that identifies a subject, object, and predicate” (Shinn, pg. 2. ¶ [0034])
“one or more lemma are identified as well as a performative classification, a sender of the input, and a receiver of the input.” (Shinn, pg. 2. ¶ [0034])
“identified parts of speech, lemmas (e.g., subject, predicate, and/or object lemma), coreference relationships, and performative classification, among others.” (Shinn, pg. 7, ¶ [0079])
“A starting node of an artificial intelligence memory graph data structure is selected to begin a search for one or more supporting knowledge data nodes associated with the classified components,” (Shinn, pg. 2, ¶ [0034])
Under BRI, this semantic/contextual information associated with the query reads on gist information.
Shinn does not teach these limitations and/or portions of:
“a temporal limit value, and a near or far value”
“spatio-temporal”
“is temporally consistent with the gist information based on the near or far value;”
“increase memory strength of the matching data appended to the results data structure;”
“based on another pattern of search”
“a sequence of data as output.”
Chakraborty, however, teaches these limitations and/or portions of:
“increase memory strength of the matching data appended to the results data structure;” – Chakraborty’s Retrieve Algorithm determines whether a target data neuron matches the search cues, invokes the Reaction procedure upon a successful match, and then returns the target neuron’s data. The Reaction Algorithm expressly performs, upon a successful retrieval:
“
I
n
c
r
e
a
s
e
_
M
e
m
_
S
t
r
(
T
a
r
g
e
t
D
N
)
” (Chakraborty, § Algorithm 4, line 8)
Thus, Freeman supplies fetching/appending the matching data to the result collection, while Chakraborty teaches increasing the memory strength of that successfully retrieved matching data.
“based on another pattern of search” - Chakraborty teaches retrieval according to a search order and updating the NMN based on the successor failure of the retrieval search. Its Reaction procedure is expressly used for:
“creating new association weights, and updating search order for cues” (Chakraborty, § Appendix B-1)
The resulting new search order replaces the prior search order according to paths and association strengths.
Chakraborty does not teach these limitations and/or portions of:
“a temporal limit value, and a near or far value”
“spatio-temporal”
“is temporally consistent with the gist information based on the near or far value;”
“a sequence of data as output.”
Esser, however, teaches these limitations and/or portions of:
“a temporal limit value, and a near or far value” – Esser expressly queries temporal relationships using durations and path lengths, calculates elapsed time between events, and alternatively measures event distance by counting path hops:
“For Q5 we want to query temporal relations in the form of durations and path lengths” (Esser, pg. 36, § 6.1-Q5)
“calculate the elapsed time between them” (Esser, pg. 36, § 6.1-Q5)
“In case we want to retrieve the distance wrt. the number of activities, we can aggregate over the nodes along the path between the two events with eventually-follows relation and count the hops with the \Length()" function” (Esser, pg. 36, § 6.1-Q5)
Esser further validates its Q5 query using a:
“the Follower filter with constraints on the time between eventually following events” (Esser, pg. 38, § 6.2)
thereby teaching application of a temporal-distance constraint.
Esser additionally distinguishes temporally immediate relationships from more remote ones. A directly-follows relationship concerns two consecutive events, whereas an eventually-follows relationship permits a path of arbitrary length between the events:
“by considering 2 consecutive events. Directly-follows relations of events in a case are an important characteristic of event logs as they represent the case internal temporal order of events and many of today's process mining techniques rely on these relations.” (Esser, pg. 34-35, § 6.1-Q3)
“Eventually-follows relations are also related to the case internal order of events. Event y eventually follows event x if y occurs after x in the same case, that is, if x and y are connected through a path of directly-follows relations of arbitrary length.” (Esser, pg. 35, § 6.1-Q3)
Thus, Esser teaches or suggests a temporal limit and a criterion distinguishing temporally nearer from farther data.
“spatio-temporal” – Esser expressly represents event nodes having timestamps/order attributes and directly/eventually-follows graph relationships defining temporal order:
“all (event and entity) attributes are encoded as properties of event nodes.” (Esser, pg. 34, § 6.1-Q1)
“a timestamp or ordering attribute defining the order of events.” (Esser, pg. 2, § 1)
“Directly-follows relations of events in a case are an important characteristic of event logs as they represent the case internal temporal order of events” (Esser, pg. 35, § 6.1-Q2)
And explains that an eventually-follows relationship exists when y occurs after event x in the same case, such that the events are connected through a path of directly-follows relationships of arbitrary length. (Esser, pg. 35, § 6.1-Q3)
“is temporally consistent with the gist information based on the near or far value;” – Esser expressly teaches directly-follows relationships for consecutive events, explaining that such relationships:
“represent the case internal temporal order of events” (Esser, pg. 35, § 6.1-Q2)
and further teaches eventually-follows relationships in which a later event occurs after an earlier event and the two are connected through a path of directly-follows relationships of arbitrary length. (Esser, pg. 35, § 6.1-Q3)
Esser also expressly teaches querying temporal relationships:
“in the form of durations and path lengths” (Esser, pg. 36, § 6.1-Q5)
using timestamps to calculate the elapsed time between events and, alternatively, measuring distance by counting the hops along the path using the Length() function. (Esser, pg. 36, § 6.1-Q5)
Esser further teaches applying temporal constraints to such relationships, validating Q5 using a:
“Follower filter with constraints on the time between eventually following events” (Esser, pg. 38, § 6.2)
Accordingly, in the combination, Shinn supplies semantic gist information used to identify relevant/matching data, while Esser supplies temporal ordering, temporal-distance measurement, and temporal constraints for determining whether graph-resident data satisfy a selected temporal relationship. Applying Esser’s temporal criteria to the data identified according to Shinn’s gist information would have taught or suggested determining whether the matching data is temporally consistent with its gist information according to the claimed near/far temporal criterion.
“a sequence of data as output.” – Esser expressly states:
“A case variant is the sequence of activities of a case.” (Esser, pg. 35, § 6.1-Q4)
Esser then teaches querying the graph to return a path of events for the case and explains that the returned event paths:
“can be turned in a list of activity sequences” (Esser, pg. 36, § 6.1-Q4)
Thus, Freeman supplies the multi-element result collection as discussed above, while Esser supplies returning temporally ordered graph data as a sequence of data.
It would have been obvious to one of ordinary skill in the art to modify the Freeman-Shinn retrieved system according to Chakraborty and Esser to provide adaptive retrieval-based learning and temporal criteria for retrieval of semantically relevant data. Chakraborty teaches reinforcing successfully retrieved data and modifying the memory-network search organization based on retrieval feedback, with the objective of increasing memory-search speed according to access-pattern bias. Esser teaches representing and querying temporal relationships among graph-resident data using direct/eventually-follows relationships, elapsed duration, path distance, and temporal constraints.
A POSITA would have been motivated to apply Chakraborty’s retrieval-driven reinforcement to Freeman’s similarity-based retrieval system so that successfully retrieved data and corresponding search paths become more prominent for subsequent searches, predictably improving retrieval efficiency. A POSITA would further have been motivated to apply Esser’s temporal graph-query techniques so that semantically relevant data identified using Shinn’s query information could additionally be selected according to temporal proximity or separation, predictably enabling retrieval and return of temporally related data in sequence.
Regarding claim 10, Freeman in view of Shinn further in view of Chakraborty and further in view Esser, teach the computing entity of claim 9, wherein the one or more processors are further configured to:
“determine whether the matching data is within a distance of the temporal limit value from other elements in the results data structure based on the near or far value comprising a near value;” – Freeman does not teach this limitation. Esser, however, teaches or suggests this limitation in part. Esser expressly teaches querying temporal relationships in terms of “durations and path lengths”. For graph query Q5, Esser calculates the elapsed time between two event nodes from their timestamps and alternatively determines distance by counting hops along the path between the events using the Length() function. Esser further validates Q5 using:
“the Follower filter with constraints on the time between eventually following events for Q5” (Esser, pg. 38, § 6.2)
thereby expressly teaching applying a temporal constraint to the temporal distance between related event data.
Thus, Esser teaches determining the temporal distance between event data and applying a temporal constraint thereto. Esser does not expressly characterize the constraint as the claimed near value or state the comparison using the exact phrase “within a distance of the temporal limit value”.
“and determine whether the matching data is at least the distance of the temporal limit value from the other elements in the results data structure based on the near or far value comprising a far value.” – Freeman does not teach this limitation. Esser, however, teaches or suggests this limitation in part. Esser’s graph query Q5 expressly calculates the elapsed temporal distance between events and demonstrates ordering events according to the calculated temporal distance:
“
O
R
D
E
R
B
Y
t
i
m
e
D
E
S
C
L
I
M
I
T
1
” (Esser, pg. 36, § 6.1-Q5, lines 4-5)
such that:
“Only the result with the longest duration is returned.” (Esser, pg. 36, § 6.1-Q5)
Esser therefore teaches determining and distinguishing greater temporal separation between graph elements based on a measured temporal distance. Its disclosed temporal filtering additionally applies constraints to the time between events. In validating its Q5 temporal-distance query, Esser expressly states that it used a:
“Follower filter with constraints on the time between eventually following events for Q5” (Esser, pg. 38, § 6.2)
Esser does not expressly call such greater separation “far” or expressly state the claimed “at least the distance of the temporal limit value” comparison.
Esser does not teach these limitations and/or portions of:
“within a distance of the temporal limit value” and characterization of the constraint as the claimed near value
“at least the distance of the temporal limit value”
However, claim 9 already provides the temporal limit value and the near or far value, and Esser expressly teaches (1) calculating a numerical temporal distance between graph elements and (2) applying temporal constraints to that distance. Once those teachings are combined, a POSITA would have found it obvious to implement the claimed near and far classifications by comparing the calculated temporal distance with the specified temporal limit: treating data whose temporal distance is within the limit as “near”, and data whose temporal distance is at least the limit as “far”.
These are predictable complementary uses of a known numerical threshold for classifying a measured distance relative to a boundary. The modification would predictably permit the Freeman-Shinn-Chakraborty-Esser system to distinguish matching data having temporal proximity from matching data having temporal separation according to the near/far retrieval parameter already recited by claim 9.
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Freeman in view of Shinn, further in view of Esser and further in view of Luhui Hu (US20210406779A1).
Regarding claim 11, Freeman in view of Shinn further in view of Esser and further in view of Hu, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“receive ” – Neither Freeman nor Shinn teach this limitation. Esser, however, teaches this limitation in part. Esser teaches determining graph-path distance according to the number of activities along a path, stating that the distance may be retrieved by aggregating the nodes along the path and counting the hops using the Length() function:
“In case we want to retrieve the distance wrt. the number of activities, we can aggregate over the nodes along the path between the two events with eventually-follows relation and count the hops with the ‘Length()’ function” (Esser, pg. 36, § 6.1-Q5)
Esser therefore teaches a quantitative graph-path distance corresponding to the claimed activity highlight distance value.
Esser does not teach the claimed background identifier threshold.
“” – Freeman, Shinn, nor Esser teach this limitation.
“perform a plurality of traces comprising unique paths to the background data neuron ” – Neither Freeman nor Shinn teach this limitation. Esser, however, teaches this limitation in part. Esser expressly queries graph paths between event nodes and teaches:
“This way we get a unique path for every Offer that meets the criteria.” (Esser, pg. 37, §§ 6.1-Q6)
Esser further states that Figure 22 shows:
“2 of the 218 paths of the query's output” (Esser, pg. 37, §§ 6.1-Q6)
Esser therefore expressly teaches generating a plurality of traces comprising unique graph paths to selected event nodes. Esser also teaches measuring path distance as discussed above. Esser does not expressly teach limiting those paths according to the claimed activity highlight distance value.
“and return the plurality of traces comprising referential activities as output.” – Neither Freeman nor Shinn teach this limitation. Esser, however, teaches this limitation. Esser teaches returning event paths and further states:
“The paths of events returned by the above query can be turned in a list of activity sequences …” (Esser, pg. 36, §§ 6.1-Q4)
using nodes() to translate each path into a list of event nodes and mapping each event node to its activity property.
Thus, Esser teaches returning a plurality of traces comprising activities represented by the event nodes along the paths.
Neither Freeman, Shinn, nor Esser teach these limitations and/or portions of:
“a background identifier threshold”
“select the data neuron from the NoK as a background data neuron in accordance with the background identifier threshold;”
“within the activity highlight distance value;”
Hu, however, teaches these limitations and/or portions of:
“a background identifier threshold”, “select the data neuron from the NoK as a background data neuron in accordance with the background identifier threshold;”, “within the activity highlight distance value;” – Hu expressly teaches selecting graph features according to threshold criteria. Hu determines a correlation metric and:
“compare[s] the determined correlation metric to one or more pre-determined criteria (e.g., pre-determined thresholds).” (Hu, pg. 7, ¶ [0044])
When the criterion is met, the feature is selected/recommended; when it is not met, the feature is not selected. (Hu, pg. 6-7, ¶¶ [0044]-[0046])
Thus, applying Hu’s threshold-based selection to the Shinn data neurons teaches selecting a data neuron as a background data neuron in accordance with the background identifier threshold.
Hu further teaches graph-distance criteria. Hu states that:
“the distance between two nodes in the knowledge graph may be determined based on the number of edges between these two nodes” (Hu, pg. 5, ¶ [0034])
And separately teaches grouping graph features that are:
“within a threshold distance” (Hu, pg. 5, ¶ [0036])
Thus, Hu teaches or suggests applying a threshold distance to graph nodes. In combination with Esser’s unique-path generation and path-distance measurement, this teaches limiting the generated traces to paths terminating at the selected background data neuron within the activity distance value.
It would have been obvious to apply Hu’s threshold-based node-selection and distance criteria to the Freeman-Shinn-Esser system so that path tracing is performed for selected data nodes satisfying a predetermined criterion and is restricted according to a desired graph distance. Hu expressly teaches threshold-based selection and threshold-distance grouping, while Esser teaches path generation and path-distance measurement. The combination would predictably permit the system to select the data neuron to be traced and constrain the resulting traces according to a specified distance.
Claims 13-19 are rejected under 35 U.S.C. 103 as being unpatentable over Freeman in view of Shinn and further in view of Hu.
Regarding claim 13, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 1, wherein
“the NoK comprises one or more compute neurons, wherein the one or more compute neurons are configured to perform one or more operations on the set of one or more data neurons.” – Freeman nor Shinn teach this limitation. Hu, however, teaches this limitation. Hu teaches a knowledge graph comprising nodes corresponding to machine-learning models and nodes corresponding to features:
“Each node of the knowledge graph may correspond to a ML model or a feature in the knowledge graph.” (Hu, pg. 3, ¶ [0029])
Hu further teaches that each feature may be associated with one or more ML models:
“Each feature in the knowledge graph may be associated with one or more ML models” (Hu, pg. 3, ¶ [0029])
Hu also teaches performing operations involving the machine-learning models and their associated features. In particular, the system determines correlation metrics between a particular ML model and features in the knowledge graph and determines recommended features for the model based on those correlations. Hu further teaches that:
“The ML model with these recommended features may be deployed to an application for recommending Ads for users of a social network platform.” (Hu, pg. 8, ¶ [0050])
Under BRI, Hu’s knowledge-graph nodes corresponding to ML models read on the claimed compute neurons, while the feature/data nodes associated with those models read on the claimed data neurons. The disclosed ML models perform inference/recommendation operations using their associated features, thereby teaching or suggesting compute neurons configured to perform operations on the set of data neurons.
It would have been obvious to one of ordinary skill in the art to further modify the Freeman-Shinn memory graph according to Hu to include model/computational nodes associated with the stored data/feature nodes. Hu teaches representing ML models and their associated features as interconnected nodes in the same knowledge graph and using the models with those features to perform inference operations.
A POSITA would have been motivated to incorporate such computational nodes into the Freeman-Shinn memory structure to permit operations to be performed directly using the data/features represented within the graph, with the predictable result of providing graph-resident compute neurons configured to operate on the graph-resident data neurons.
Regarding claim 14, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 13, wherein
“the one or more operations comprise one or more of creating new knowledge, updating data, or generating responses to queries.” – Freeman nor Shinn teach this limitation. Hu, however, teaches this limitation. Hu expressly teaches using graph neural networks or intelligent logic to:
“learn new knowledge (e.g., relationships, similarity, importance, relevance, correlations)” (Hu, pg. 2, ¶ [0023])
about models and features, and further teaches that:
“The system may dynamically update the knowledge graph based on new knowledge learned through graph learning.” (Hu, pg. 2, ¶ [0023])
Hu also teaches receiving query information associated with a particular ML model and using the knowledge graph to determine/recommend features for that model. (Hu, pg. 2, ¶ [0023])
Thus, Hu expressly teaches at least creating new knowledge and updating data/the knowledge graph. Because claim 14 recites the alternatives disjunctively as “one or more of”, these teachings satisfy the additional limitation.
Regarding claim 15, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 13, wherein
“the one or more compute neurons are located at one or more regions of the NoK comprising a subset of data neurons of the set of one or more data neurons on which the one or more compute neurons are most likely to operate.” – Freeman nor Shinn teach this limitation. Hu, however, teaches or suggests this limitation. As discussed with respect to claim 13, Hu’s ML-model nodes read on the claimed compute neurons and Hu’s feature nodes read on the claimed data neurons.
Hu further teaches that its knowledge graph may contain:
“embedded sub-graphs corresponding to different layers of features or knowledge graphs in different sub-problem domains.” (Hu, pg. 6, ¶ [0043])
Hu’s FIG. 3F embodiment includes model 381 and associated features 382, 383, and 386. Hu further teaches that, for a particular model in a sub-domain, the system determines correlation metrics for particular features with respect to that model and, when the metrics satisfy predetermined criteria, correlates those features to that particular sub-domain model. (Hu, pg. 6, ¶ [0043])
Hu likewise teaches determining recommended features for a particular ML model based on correlation metrics between that model and features in the knowledge graph. (Hu, pg. 11, ¶¶ [0048]-[0049]).
Thus, under BRI, Hu’s embedded sub-graph/sub-problem domains read on the claimed regions of the NoK, the features associated with a particular model from a subset of data neurons, and Hu’s correlation-based selection of features for that particular model teaches or suggests locating/associating the compute neuron with the region containing the subset of data neurons on which that compute neuron is most likely to operate.
Regarding claim 16, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 1, wherein
“the NoK is distributed across a plurality of computing entities.” – Freeman nor Shinn teach this limitation. Hu, however, teaches or suggests this limitation. Hu’s disclosed feature knowledge graph system accesses and operates on a knowledge graph defining relationships among ML-model and feature nodes. Hu further teaches that the functionality described therein may be performed by “one or more computer systems 800” and that computer system 800 may:
“be unitary or distributed; span multiple locations; span multiple machines; span multiple data centers; or reside in a cloud,” (Hu, pg. 15, ¶¶ [0079]-[0080])
Hu additionally teaches a network environment comprising client systems, a social-networking system, servers, data stores, and third-party systems, and teaches that a server may be:
“a distributed server spanning multiple computers or multiple datacenters.” (Hu, pg. 10, ¶ [0061]; FIG. 6)
Thus, Hu expressly teaches implementing the disclosed graph-based system functionality across multiple distributed computing systems/machines. Under BRI, distributing the computing resources implementing and providing access to the knowledge graph across those computing systems teaches or suggests a NoK distributed across a plurality of computing entities.
It would have been obvious to one of ordinary skill in the art to implement the Freeman-Shinn memory graph using Hu’s disclosed distributed-computing architecture. Hu expressly teaches performing the disclosed graph-system functionality using distributed computer systems spanning multiple machines, locations, or data centers.
A POSITA would have been motivated to distribute the Freeman-Shinn memory system across multiple computing entities to accommodate increased storage and processing requirements and to make the graph-access functionality available across networked computing resources, with the predictable result of the NoK being distributed across a plurality of computing entities.
Regarding claim 17, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 16, wherein
“at least one of the plurality of computing entities is spatially aware of memory content of a second one of the plurality of computing entities via the NoK.” – Freeman nor Shinn teach this limitation. Hu, however, teaches or suggests this limitation. Hu taches a network environment including distinct computing entities, including a social-networking system, client systems, and third-party systems. A third-party system may include “one or more types of servers” and “one or more data stores”, and may operate in conjunction with the social-networking system over the network. (Hu, pg. 11, ¶ [0063]-[0065]).
Hu further teaches that the shared graph provides access to information across those computing entities:
“a social-networking system 660, client system 630, or third party system 670 may access social graph 700 and related social-graph information for suitable applications.” (Hu, pg. 12, ¶ [0069])
Hu also teaches that the nodes and edges of the social graph may be stored as data objects in a data store:
“The nodes and edges of social graph 700 may be stored as data objects, for example, in a data store (such as a social-graph database).” (Hu, pg. 12, ¶ [0069])
Hu additionally teaches that a resource represented by a concept node may be located on an external server, that the concept node may be associated with data objects corresponding to information about the resource, and that profile pages may be hosted on third-party websites while being accessible to the social-networking system. (Hu, pg. 12, ¶¶ [0071]-[0072]).
Thus, under BRI, Hu teaches or suggests one networking computing entity having awareness of/access to information or memory content residing at another computing entity through the graph and its graph relationships. Hu’s graph therefore reads on the claimed NoK, and the cross-system access to graph-associated content reads on at least one computing entity being spatially aware of memory content of a second computing entity via the NoK.
Regarding claim 18, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 16, wherein
“the NoK comprises one or more neurons configured to move ” – Freeman does not teach this limitation. Shinn, however, teaches this limitation in part. Shinn teaches short-term and long-term memory graph structures comprising stored components represented using memory-graph node structures. Shinn further teaches that components may be transferred from short-term memory to long-term memory:
“the components are only transferred to long-term memory when there is insufficient storage space in short-term memory.” (Shinn, pg. 19, ¶ [0252])
Shinn also teaches that removed short-term-memory components may be transferred to long-term memory (Shinn, pg. 19, ¶¶ [0252], [0255]-[0257])
More particularly, Shinn expressly bases movement of the stored memory content on access behavior:
“information from short-term memory is removed based on a least recently used metric.” (Shinn, pg. 19, ¶ [0256])
Shinn further teaches:
“a last accessed time may be used to determine the oldest accessed memory, which is then marked as a candidate for removal.” (Shinn, pg. 19, ¶ [0256])
and teaches removing information when it has not been accessed, after which the removal components may be transferred to long-term memory. (Shinn pg. 19, ¶¶ [0255]-[0257]).
Thus, under BRI, Shinn’s graph-resident stored components/nodes read on the claimed neurons, and transferring those components from short-term to long-term memory based on least-recently-used/last-accessed information teaches moving the neurons based on change in data access behavior.
Shinn further teaches that:
“long-term memory utilizes a remote database for storage and connectivity of different nodes of the saved components.” (Shinn, pg. 19, ¶ [0257])
Accordingly, Shinn teaches movement of memory-graph node content from one memory tier to a remotely implemented memory tier.
Shinn nor Freeman teach these limitations and/or portions of:
“across the plurality of computing entities”
Hu, however, teaches these limitations and/or portions:
“across the plurality of computing entities” – as set forth with respect to claim 16, Hu teaches distributed computing implementations spanning multiple machines, locations, and/or data centers:
“unitary or distributed; span multiple locations; span multiple machines; span multiple data centers; or reside in a cloud,” (Hu, pg. 15, ¶¶ [0079]-[0080])
“a distributed server spanning multiple computers or multiple datacenters.” (Hu, pg. 10, ¶ [0061]; FIG. 6)
In the Freeman-Shinn-Hu combination of claim 16, applying Shinn’s disclosed transfer or graph-resident memory components to Hu’s distributed implementation would result in those neurons moving across the plurality of computing entities.
Because claim 18 requires movement based on “one or more of” the recited alternatives, Shinn’s express teaching of transfer/removal based on least-recently-used and last-accessed information is sufficient to teach or suggest the claimed change in data access behavior alternative.
Regarding claim 19, Freeman in view of Shinn and further in view of Hu, teach the computing entity of claim 1, wherein the one or more processors are further configured to:
“reduce association strengths of a subset of cue neurons of the set of one or more cue neurons and the set of one or more data neurons in the NoK;” – Freeman does not teach this limitation. Hu, however, teaches or suggests this limitation. Hu teaches a knowledge graph in which relationships between graph nodes are represented by weighted edges. Hu further teaches adjusting those edge weights based on inference values and model/feature evaluation results. In particular, when performance is low:
“the system may cause the corresponding edges representing the relationships between these features and the corresponding ML model to have greater weight values indicating a higher level of relevance or importance.” (Hu, pg. 8, ¶ [0050])
In the Freeman-Shinn combination, Freeman supplies the cue nodes and Shinn supplies the graph-resident data nodes. Under BRI, Hu’s weighted edges read on association strengths between such graph nodes, and reducing selected edge weights teaches reducing association strengths between a subset of cue neurons and the data neurons.
“and reduce memory strength of the set of one or more data neurons in the NoK.” – Freeman does not teach this limitation. Hu, however, teaches or suggests this limitation. Hu expressly teaches reducing weights associated with graph features when those features/model combinations exhibit poor performance. The reduced weights indicate:
“lower level of relevance or importance” (Hu, pg. 8, ¶ [0050]).
Under BRI, the weight/relevance/importance associated with a graph-resident data representation reads on its memory strength. Thus, reducing those weights values teaches or suggests reducing the memory strength of the corresponding data neurons.
It would have been obvious to one of ordinary skill in the art to modify the Freeman-Shinn memory graph according to Hu to provide adjustable strengths for graph relationships and stored-data representations. Hu teaches decreasing graph weights when evaluation indicates lower relevance or importance. A POSITA would have been motivated to apply such adaptive weighting to the Freeman-Shinn graph so that less relevant relationships and data representations are correspondingly weakened, with the predictable result of reducing association strengths and memory strength within the NoK.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner
should be directed to Paul Coleman whose telephone number is (571)272-4687. The examiner can
normally be reached Mon-Fri.
Examiner interviews are available via telephone, in-person, and video conferencing using a
USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use
the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor,
David Yi can be reached at (571) 270-7519. The fax phone number for the organization where this
application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from
Patent Center. Unpublished application information in Patent Center is available to registered users. To
file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit
https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and
https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional
questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like
assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or
571-272-1000.
/PAUL COLEMAN/
Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights