DETAILED ACTION
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 .
Response to Amendment/Status of Claims
Claims 1, 9, 10, and 18 were amended.
Claims 1-18 are pending and examined herein.
Claims 1, 3-6, 8, 10, 12-15, and 17 are rejected under 35 U.S.C. 102.
Claims 2, 7, 9, 11, 16, and 18 are rejected under 35 U.S.C. 103.
Response to Arguments
Applicant's arguments filed 03/17/2026 have been fully considered but they are not persuasive. Applicant argues, see pages 5-8, that "Gao is completely silent on the use of symbolic reasoning in neurons of a neural network, let alone the specific limitations of amended Claim 1." Examiner respectfully disagrees.
Wikipedia, “Symbolic artificial intelligence”, states "Symbolic artificial intelligence is the term for the collection of all methods in artificial intelligence research that are based on high-level "symbolic" (human-readable) representations of problems, logic and search." Gao teaches, see page 3, "We attempt to improve this process, and our core strategy is to link the network architecture with human understandable concepts. In general, the human understanding of concepts is naturally hierarchical. For instance, the appearance of Maybach in an image could trigger a series of concepts like Maybach
→
car
→
motor vehicle
→
vehicle
→
artifact
→
physical object." Gao further states "As mentioned above, by adopting a concept hierarchy suitable for a given image collection, we can associate each concept in the hierarchy with a neuron in the hidden layers and force the neuron connectivity to match the concept organization hierarchy." Therefore, as the architecture performs reasoning and links neurons to human-readable concepts, Gao teaches the use of symbolic reasoning in neurons of a neural network. See amended 35 U.S.C. 102 and 35 U.S.C. 103 rejections below for a full analysis.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim(s) 1, 3-6, 8, 10, 12-15, and 17 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Gao (“An interpretable deep architecture for similarity learning build upon hierarchical concepts”, January 2020).
Regarding claim 1, Gao teaches
A computer-implemented method for training an artificial neural network, the method comprising: (Page 2 states "We propose a similarity neural network (SNN) built upon novel designs of interpretable network architecture and neuron operations, and propose various approaches for visualizing and understanding the SNN neurons." Section B describes the training of the neural network. Therefore, the method includes training an artificial neural network.)
obtaining a training sample for training the artificial neural network; (Page 6 states "To improve this, we have previously proposed an unsupervised multi-view training algorithm in [22], [62]. It pre-trains a CNN to preserve knowledge offered by multiple image feature extraction methods that characterize heterogeneous properties of the image content. To facilitate the practitioners, we explain in Appendix A how the multi-view pre-training works. By taking the image representation vector
ϕ
i
computed by the CNN as the input, we further optimize the SNN model variables by minimizing the loss function in Eq. (2) through stochastic gradient descent, where $ is used to store the SNN variables instead. So far, the separate training of CNN and SNN divides the model architecture into two independent components: (1) the unsupervised feature learning, and (2) the supervised relation learning. A fine-tuning procedure is conducted based on the ranking loss, to further jointly optimize all the model variables, including
η
for CNN,
ϖ
for SNN and
w
for relevance prediction." The image content is interpreted as the training sample.)
determining multiple sub concepts within the training sample; (Page 4 states "The cluster-based concept hierarchy can be obtained by applying a hierarchical clustering algorithm [59] over the high-level representations of images
ϕ
i
i
=
1
n
returned by the CNN network. The resulting concepts correspond to image clusters and are therefore naturally observable. Fig. 1(b) demonstrates a three-level concept hierarchy based on image clusters, where each concept is represented by an image that is closest to the cluster center in the CNN feature space. The hierarchical organization of the extracted concepts can be modelled by a tree structure. Let
l
0
denote the number of leaf concepts, corresponding to level 0. The
h
-th level contains the parents of the concepts from the previous level, and the number of parent concepts in the
h
-th level
(
h
=
1
,
2
,
.
.
.
,
H
)
is denoted by
l
h
." As the clusters are using the image representations from the CNN that receives the image content, the leaf concepts and their parents (interpreted as sub concepts) are from within the training sample.)
processing the sub concepts to obtain differential neurons associated with the sub concepts, wherein the differential neurons are configured to utilize symbolic reasoning to provide a relative distinction between the sub concepts; (Page 4 states "Therefore, we force the neuron organization hierarchy to be identical to a concept organization hierarchy. Each object pair is connected to all leaf neurons. Specifically, the
t
-th concept in the
h
-th level corresponds to the
t
-th neuron in the
h
+
1
-th hidden layer, where
h
=
0,1
,
…
,
H
. In Fig. 2, we demonstrate the network architectures built upon the two example cases in Fig. 1." Therefore, the neurons are associated with the sub concepts. Page 5 further states "It can be seen that the concept hierarchy results in sparse connections between the hidden layers, where neuron connections in dashed lines are removed to ensure alignment between the neurons and the pre-constructed concepts." Therefore, the neuron and its connections are interpreted as the differential neuron, as the neuron connections also differentiate between concepts. Page 5 states "The input layer of our SNN contains a pair of image vectors
ϕ
i
and
ϕ
j
that are computed by a CNN. A similarity score between an input image and the leaf concept can be simply modelled by
ϕ
i
T
c
t
+
b
, where b is a bias parameter. Two similar images with their relevance triggered specifically by the concept
c
t
would be expected to have a high score of
s
=
ϕ
i
T
c
t
+
b
ϕ
j
T
c
t
+
b
3
"The full parametric formulation in Eq. (7) is expected to model a wider range of similarity patterns between images, but meanwhile the initialization by attempts to maintain a connection between the neurons and
c
t
the pre-constructed concepts. Specifically, the pre-constructed concepts largely affect which local optimum to be arrived during the training." Therefore, the neurons are differential, as they provide a distinction between the sub concepts (when the input is similar to the neuron/concept, the relevance score will be higher, thus distinguishing between two inputs with different concept similarities). Wikipedia, “Symbolic artificial intelligence”, states "Symbolic artificial intelligence is the term for the collection of all methods in artificial intelligence research that are based on high-level "symbolic" (human-readable) representations of problems, logic and search." Gao teaches, see page 3, "We attempt to improve this process, and our core strategy is to link the network architecture with human understandable concepts. In general, the human understanding of concepts is naturally hierarchical. For instance, the appearance of Maybach in an image could trigger a series of concepts like Maybach
→
car
→
motor vehicle
→
vehicle
→
artifact
→
physical object." Gao further states "As mentioned above, by adopting a concept hierarchy suitable for a given image collection, we can associate each concept in the hierarchy with a neuron in the hidden layers and force the neuron connectivity to match the concept organization hierarchy." Therefore, as the architecture performs reasoning and links neurons to human-readable concepts, Gao teaches the use of symbolic reasoning in neurons of a neural network. As the differential neuron is linked to a concept and forms a hierarchy (reasoning), the differential neuron is configured to utilize symbolic reasoning. )
integrating the differential neurons to obtain sub concepts neurons, wherein the sub concepts neurons are configured to utilize symbolic reasoning to provide an absolute distinction of sub concepts; and (Page 4 states "The parent concepts from the same level are uniquely distinguishable by their connections to the concepts from the previous level." Therefore, the parents of the leaf concepts are interpreted as the sub concepts neurons. Page 6 states "Regarding to the messages received by a parent concept, it is natural to assume that a parent concept can receive only the messages passed by its child concepts. Based on this, an accumulated message received by the
t
-th concept at level
h
can be formulated as …". Accumulating the message is interpreted as integrating the neurons. As the sub concepts neuron is linked to a concept and forms a hierarchy (reasoning), the differential neuron is configured to utilize symbolic reasoning.)
integrating the sub concepts neurons to obtain concept neurons, wherein the concept neurons are configured to utilize symbolic reasoning to form an output of the neural network. (Page 7 states "As compared to the leaf concepts, the parent concepts represent more abstract patterns. They can be understood as an accumulation of the selected leaf patterns and may therefore not appear visually meaningful when observing their computed visual patterns." Therefore, the final layer parent concept neurons are interpreted as the concept neurons, which as above, will be accumulated with the past messages, interpreted as the integration. Fig. 2 shows that the parent neuron at the
H
-th level (green triangles) form the output. Fig. 8 also shows that the last layer parent concept forms an output of the neural network. As the concept neuron is linked to a concept and forms a hierarchy (reasoning), the differential neuron is configured to utilize symbolic reasoning.)
Regarding claim 3, the rejection of claim 1 is incorporated herein. Gao teaches
wherein the determining of the sub concepts within the training sample includes obtaining various subsets of the training sample and distinguishing between the various subsets. (Page 13 states "Specifically, we extract images from the 3 categories of “dog”, “car” and “bird” in CIFAR-10 dataset, and group these images to 6 clusters using k-means clustering based on their pre trained image representation vectors. The 6 images that are the closest to the clustering centers are selected. Then, we randomly select 4 patches from each image and resize these patches to the same as that of the original image. The pre-trained image representation vectors of these
6
×
4
=
24
patches are used to initialize the concepts
c
t
t
=
1
24
.” The categories are interpreted as the subsets, and the grouping using k-means clustering is interpreted as distinguishing between the various subsets. The concepts are interpreted as the sub concepts.)
Regarding claim 4, the rejection of claim 1 is incorporated herein. Gao teaches
wherein unsupervised learning is used to determine hierarchical structure of the sub concepts. (Page 4 states "When there is no explicit external knowledge available to help building a concept hierarchy, an alternative way is to seek latent topics contained within the images by exploring their visual content, which usually corresponds to clusters of images. Such an approach is based on the assumption that interactions between images exhibit the same underlying topic structure as that revealed by individual images. A similar strategy has been pursued in [58], which assumes that both individual documents and document pairs are generated from the same set of topic distributions. The cluster-based concept hierarchy can be obtained by applying a hierarchical clustering algorithm [59] over the high-level representations of images
ϕ
i
i
=
1
n
returned by the CNN network. The resulting concepts correspond to image clusters and are therefore naturally observable. Fig. 1(b) demonstrates a three-level concept hierarchy based on image clusters, where each concept is represented by an image that is closest to the cluster center in the CNN feature space." One of ordinary skill in the art would recognize that this clustering is unsupervised, as it does not use external knowledge/labeled data.)
Regarding claim 5, the rejection of claim 1 is incorporated herein. Gao teaches
wherein the sub concepts are overlapping or hierarchically structured. (Page 4 states "As mentioned above, by adopting a concept hierarchy suitable for a given image collection, we can associate each concept in the hierarchy with a neuron in the hidden layers and force the neuron connectivity to match the concept organization hierarchy" The concepts are interpreted as sub concepts, which are hierarchically structured.)
Regarding claim 6, the rejection of claim 1 is incorporated herein. Gao teaches
wherein one or more of the differential neurons are pruned before the integrating of the differential neurons to obtain sub concepts neurons. (Page 5 states "It can be seen that the concept hierarchy results in sparse connections between the hidden layers, where neuron connections in dashed lines are removed to ensure alignment between the neurons and the pre-constructed concepts." Additionally, the caption of Fig. 2 states "The dashed lines correspond to the eliminated neuron connections due to the concept hierarchy matching." Therefore, the differential neurons along a path would be removed, meaning that when the message is passed, it will not be integrated to obtain the sub concept neuron that it is no longer connected to.)
Regarding claim 8, the rejection of claim 1 is incorporated herein. Gao teaches
tuning the artificial neural network after the training to improve its performance. (Page 6 states "The complete similarity learning process includes the training of a CNN to obtain the neural content code for characterizing an image, the training of the proposed SNN to obtain the neural relevance code characterizing the image relevance, and finally a fine-tuning process for optimizing the two connected CNN-SNN networks together.")
Regarding claim 10, Gao teaches
A system for training a neural network, the system comprising a processor and an associated memory, the processor being configured to: (Page 2 states "We propose a similarity neural network (SNN) built upon novel designs of interpretable network architecture and neuron operations, and propose various approaches for visualizing and understanding the SNN neurons." Section B describes the training of the neural network. Therefore, the method includes training an artificial neural network. One of ordinary skill in the art would realize that this method is implemented on a computer, which necessarily includes a processor and an associated memory that performs the method.)
The remainder of claim 10 recites substantially similar subject matter to claim 1 and is rejected with the same rationale, mutatis mutandis.
Claims 12, 13, 14, 15, and 17 recite substantially similar subject matter to claims 3, 4, 5, 6, and 8 respectively and are rejected with the same rationale, mutatis mutandis.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 2, 9, 11, and 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gao (“An interpretable deep architecture for similarity learning build upon hierarchical concepts”, January 2020) as applied to claim 1 above, and further in view of Shi (“Neural Logic Networks”, October 2019).
Regarding claim 2, the rejection of claim 1 is incorporated herein. Gao does not appear to explicitly teach
wherein the training sample is designed to teach one or more rules.
However, Shi—directed to analogous art—teaches
wherein the training sample is designed to teach one or more rules. (Page 4 states "Our prototype task is defined in this way: given a number of training logical expressions and their T/F values, we train a neural logic network, and test if the model can solve the T/F value of the logic variables, and predict the value of new expressions constructed by the observed logic variables in training." The new expressions constructed by the observed logic variables are interpreted as rules. As the training logical expressions (training samples) are used to train the neural network to construct rules, the training sample is designed to teach the rules.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Gao and Shi because, as stated by Shi on page 2, "By encoding logical structure information in neural architecture, NLN can flexibly process an exponential amount of logical expressions." Additionally, as stated by Shu on page 8, "The integration of logical inference and neural network reveals a promising direction to design deep neural networks for both abilities of logical reasoning and generalization"
Regarding claim 9, the rejection of claim 1 is incorporated herein. Gao does not appear to explicitly teach
wherein the symbolic reasoning provides symbolic outputs that are interpretable as algorithms.
However, Shi—directed to analogous art—teaches
wherein the symbolic reasoning provides symbolic outputs that are interpretable as algorithms. (Page 4 states "In our experiments, the AND module is implemented by multi-layer perceptron (MLP) with one hidden layer:", and "The OR module is built in the same way, and the NOT module is similar but with only one vector as input:". Page 2 states "The three modules can be implemented by various neural structures, as long as they have the ability to approximate the logical operations. Figure 1 is an example of the neural logic network corresponding to the expression
v
i
∧
v
j
∨
¬
v
k
. The red left box shows how the framework constructs a logic expression. Each intermediate vector represents part of the logic expression, and finally, we have the vector representation of the whole logic expression
e
=
v
i
∧
v
j
∨
¬
v
k
.” The modules implement logical expressions, which are human-readable symbols. Thus, the modules perform symbolic reasoning to construct a logic expression. Therefore, as the modules perform symbolic reasoning (logical expressions are human-readable symbols), the symbolic reasoning provides symbolic outputs. One of ordinary skill would realize that expressions of propositional logic such as the one taught by Shi would be interpretable as an algorithm.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Gao and Shi for the reasons given above in regards to claim 2.
Claims 11 and 18 recite substantially similar subject matter to claims 2 and 9 respectively and are rejected with the same rationale, mutatis mutandis.
Claim(s) 7 and 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gao (“An interpretable deep architecture for similarity learning build upon hierarchical concepts”, January 2020) as applied to claim 1 above, and further in view of Miconi (“Differential plasticity: training plastic neural networks with backpropagation”, 2018).
Regarding claim 7, the rejection of claim 1 is incorporated herein. Gao does not appear to explicitly teach
wherein neurons of the artificial neural network are deliberative, temporarily changing their parameters.
However, Miconi—directed to analogous art—teaches
wherein neurons of the artificial neural network are deliberative, temporarily changing their parameters. (A connection between any two neurons
i
and
j
has both a fixed component and a plastic component. The fixed part is just a traditional connection weight
w
i
,
j
. The plastic part is stored in a Hebbian trace
Hebb
i
,
j
, which varies during a lifetime according to ongoing inputs and outputs (note that we use “lifetime” and “episode” interchangeably)." Varying is interpreted as changing, and the Hebbian trace is interpreted as the parameters that are changed.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Gao and Miconi because, as Miconi states on page 1, "Endowing artificial agents with lifelong learning abilities is essential to allowing them to master environments with changing or unpredictable features, or specific features that are unknowable at the time of training"
Claim 16 recites substantially similar subject matter to claim 7 and is rejected with the same rationale, mutatis mutandis.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 JESSICA THUY PHAM whose telephone number is (571)272-2605. The examiner can normally be reached Monday - Friday, 9 A.M. - 5:00 P.M..
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, Li Zhen can be reached at (571) 272-3768. 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.
/J.T.P./ Examiner, Art Unit 2121
/Li B. Zhen/ Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights