DETAILED ACTION
This action is in response to the amendment filed 04/27/2026. Claims 1-4, 7, 9, 12, 14, 21-25, 28, 30, 33, 35, and 41-43 are pending and have been examined.
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 04/27/2026 has been entered.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-4, 7, 9, 12, 14, 21-25, 28, 30, 33, 35, and 41-43 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding Claim 1:
Subject Matter Eligibility Analysis Step 1:
Claim 1 recites a method and is thus a process, one of the four statutory categories of patentable subject matter.
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 1 recites
Determining, … based on at least the codeword and an initial graph, a preliminary reconstruction of the input data representation (This limitation is a mental process as it encompasses a human mentally determining a preliminary reconstruction of the input data representation and is thus an evaluation.)
Modifying the initial graph,… based on at least the preliminary reconstruction and the codeword, to generate a modified graph; (This limitation is a mental process as it encompasses a human mentally modifying a graph to generate a graph and is thus an evaluation.)
Determining,… based on at least the codeword and the modified graph, a refined reconstruction of the input data representation, (This limitation is a mental process as it encompasses a human mentally determining a refined reconstruction of the input data representation and is thus an evaluation.)
wherein the modified graph indicates topology information associated with the input data representation (This limitation is a mental process as it extends upon the mental process of determining a modified graph and is thus an evaluation.)
Therefore, claim 1 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 1 further recites additional elements of
implemented by a neural network-based decoder (NNBD), (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
obtaining or receiving, by the NNBD, a codeword, as a descriptor of an input data representation; (This element does not integrate the abstract idea into a practical application because it recites insignificant extra-solution activity of data gathering (see MPEP 2106.05(g)).)
by a first neural network (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
by a second neural network (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
Therefore, claim 1 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 1 do not provide significantly more than the abstract idea itself, taken alone and in combination because
implemented by a neural network-based decoder (NNBD) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
obtaining or receiving, by the NNBD, a codeword, as a descriptor of an input data representation is the well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)).
by a first neural network uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
by a second neural network uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 1 is subject-matter ineligible.
Regarding Claim 2:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 2 recites
wherein the modified graph is determined by combining the initial graph and an output of the second neural network. (This limitation is a mental process as it encompasses a human mentally combining the initial graph and an output of the second neural network and is thus an evaluation.)
Therefore, claim 2 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 2 does not further recite any additional elements. Therefore, claim 2 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
Since there are no additional elements, claim 2 does not provide significantly more than the abstract idea itself, taken alone and in combination. Therefore, claim 2 is subject-matter ineligible.
Regarding Claim 3:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 3 recites
wherein the modified graph is a locally connected graph. (This limitation is a mental process as it further expands upon the mental process of determining a modified graph from claim 1 and is thus an evaluation.)
Therefore, claim 3 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 3 does not further recite any additional elements. Therefore, claim 3 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
Since there are no additional elements, claim 3 does not provide significantly more than the abstract idea itself, taken alone and in combination. Therefore, claim 3 is subject-matter ineligible.
Regarding Claim 4:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 4 recites
generating a concatenation matrix for processing… by concatenating at least a replicated codeword, the initial graph and the preliminary reconstruction of the input data representation (This limitation is a mental process as it encompasses a human mentally generating a concatenation matrix and is thus an evaluation.)
Therefore, claim 4 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 4 further recites additional elements of
by one or more Convolutional Neural Networks (CNNs), (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
Therefore, claim 4 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 4 do not provide significantly more than the abstract idea itself, taken alone and in combination because
by one or more Convolutional Neural Networks (CNNs) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 4 is subject-matter ineligible.
Regarding Claim 7:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 7 recites
the determining of the refined reconstruction of the input data representation is performed via a plurality of iterative operations (This limitation is a mental process as it encompasses a human mentally determining the refined reconstruction of the input data representation through iterative operations and is thus an evaluation.)
Therefore, claim 7 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 7 further recites additional elements of
the NNBD is a Graph Conditioned NNBD, (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
of at least the first neural network, (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
Therefore, claim 7 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 7 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the NNBD is a Graph Conditioned NNBD uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
of at least the first neural network uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 7 is subject-matter ineligible.
Regarding Claim 9:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 9 recites
the modified graph and the refined reconstruction of the input data representation are further based on gradient information (This limitation is a mental process as it further expands upon the mental process of determining the modified graph in claim 1 and is thus an evaluation.)
Therefore, claim 9 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 9 further recites additional elements of
the NNBD includes one or more Multi-layer Perceptrons (MLPs); (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
generated by the one or more MLPs, (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
Therefore, claim 9 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 9 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the NNBD includes one or more Multi-layer Perceptrons (MLPs) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
generated by the one or more MLPs uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 9 is subject-matter ineligible.
Regarding Claim 12:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 12 recites
the initial graph and the modified graph are 2 dimensional (2D) point sets (This limitation is a mental process as it encompasses a human mentally making the initial graph and the modified graph 2D point sets and is thus an evaluation.)
the input data representation is a point cloud (This limitation is a mental process as it encompasses a human mentally making the input data representation a point cloud and is thus an evaluation.)
the determining of the preliminary reconstruction of the input data representation includes performing a deforming operation based on the codeword and the 2D point set that is initialized with a pre-determined sampling in a plane (This limitation is a mental process as it encompasses a human mentally determining the preliminary reconstruction of the input data representation by performing a deforming operation and is thus an evaluation.)
Therefore, claim 12 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 12 does not further recite any additional elements. Therefore, claim 12 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
Since there are no additional elements, claim 12 does not provide significantly more than the abstract idea itself, taken alone and in combination. Therefore, claim 12 is subject-matter ineligible.
Regarding Claim 14:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 14 recites
wherein the determining of the modified graph includes: performing a tearing operation, based on the preliminary reconstruction of the input data representation, the codeword and the initial graph to generate the modified graph. (This limitation is a mental process as it encompasses a human mentally determining the modified graph by performing a tearing operation and is thus an evaluation.)
Therefore, claim 14 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 14 does not further recite any additional elements. Therefore, claim 14 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
Since there are no additional elements, claim 14 does not provide significantly more than the abstract idea itself, taken alone and in combination. Therefore, claim 14 is subject-matter ineligible.
Regarding Claim 21:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 21 recites
wherein the determining of the modified graph includes: replicating the received or obtained codeword K times to generate a KxD codeword matrix, wherein K is a number of nodes in the initial graph and D is a length of the codeword; (This limitation is a mental process as it encompasses a human mentally determining the modified graph by replicating the codeword and is thus an evaluation.)
concatenating, the KxD codeword matrix and the initial graph, as a KxN matrix, to generate a Kx(D+ N) concatenated matrix; (This limitation is a mental process as it encompasses a human mentally concatenating the codeword matrix and the initial graph to generate a concatenated matrix and is thus an evaluation.)
generating…the modified graph (This limitation is a mental process as it encompasses a human mentally generating the modified graph and is thus an evaluation.)
Therefore, claim 21 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 21 further recites additional elements of
inputting, the concatenated matrix to the second neural network corresponding to one or more Convolutional Neural Networks (CNNs) or Multi-layer Perceptrons (MLPs); (This element does not integrate the abstract idea into a practical application because it recites insignificant extra-solution activity of data gathering (see MPEP 2106.05(g)).)
by the second neural network from the concatenated matrix (This element does not integrate the abstract idea into a practical application because it amounts to mere “apply it on a computer” (see MPEP 2106.05(f)).)
Therefore, claim 21 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 21 do not provide significantly more than the abstract idea itself, taken alone and in combination because
inputting, the concatenated matrix to the second neural network corresponding to one or more Convolutional Neural Networks (CNNs) or Multi-layer Perceptrons (MLPs) is the well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)).
by the second neural network from the concatenated matrix uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 21 is subject-matter ineligible.
Regarding Claim 22:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 22 recites
concatenating the codeword matrix to an output of a first set of CNN or MLP layers, as a concatenated intermediary matrix; (This limitation is a mental process as it encompasses a human mentally concatenating the codeword matrix to the output of MLP layers and is thus an evaluation.)
Therefore, claim 22 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 22 further recites additional elements of
inputting, the concatenated intermediary matrix to a next set of CNN or MLP layers following the first set of CNN or MLP layers (This element does not integrate the abstract idea into a practical application because it recites insignificant extra-solution activity of data gathering (see MPEP 2106.05(g)).)
Therefore, claim 22 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 22 do not provide significantly more than the abstract idea itself, taken alone and in combination because
inputting, the concatenated intermediary matrix to a next set of CNN or MLP layers following the first set of CNN or MLP layers is the well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)).
Therefore, claim 22 is subject-matter ineligible.
Regarding Claim 23:
Subject Matter Eligibility Analysis Step 1:
Claim 23 recites a neural network-based decoder is thus an apparatus, one of the four statutory categories of patentable subject matter.
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 23 recites
Determine based on at least the codeword and an initial graph, a preliminary reconstruction of the input data representation (This limitation is a mental process as it encompasses a human mentally determining a preliminary reconstruction of the input data representation and is thus an evaluation.)
Modify the initial graph, based on at least the preliminary reconstruction and the codeword, to generate a modified graph; (This limitation is a mental process as it encompasses a human mentally modifying a graph to generate a graph and is thus an evaluation.)
Determine based on at least the codeword and the modified graph, a refined reconstruction of the input data representation, (This limitation is a mental process as it encompasses a human mentally determining a refined reconstruction of the input data representation and is thus an evaluation.)
the modified graph indicates topology information associated with the input data representation (This limitation is a mental process as it extends upon the mental process of determining a modified graph and is thus an evaluation.)
Therefore, claim 23 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 23 further recites additional elements of
a neural network-based decoder (NNBD), (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
a receiver configured to receive or obtain a codeword, as a descriptor of an input data representation; (This element does not integrate the abstract idea into a practical application because it recites insignificant extra-solution activity of data gathering (see MPEP 2106.05(g)).)
a first neural network (NN) (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
a second NN (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
Therefore, claim 23 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 23 do not provide significantly more than the abstract idea itself, taken alone and in combination because
a neural network-based decoder (NNBD) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
A receiver configured to receive or obtain a codeword, as a descriptor of an input data representation is the well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)).
a first neural network (NN) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
a second neural network (NN) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 23 is subject-matter ineligible.
Regarding claim 24, claim 24 recites substantially similar limitations to claim 3, and is therefore rejected under the same analysis.
Regarding Claim 25:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 25 recites
generate a concatenation matrix using at least (1) a replicated codeword, (2) the initial graph and (3) the preliminary reconstruction of the input data representation; (This limitation is a mental process as it encompasses a human mentally generating a concatenation matrix and is thus an evaluation.)
process the concatenation matrix and to generate the modified graph or a refined modified graph. (This limitation is a mental process as it encompasses a human mentally processing the matrix to generate a modified graph and is thus an evaluation.)
Therefore, claim 25 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 25 further recites additional elements of
the second NN includes one or more Convolutional Neural Networks (CNNs) (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
the NNBD, (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
the one or more CNNs (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
Therefore, claim 25 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 25 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the second NN includes one or more Convolutional Neural Networks (CNNs) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
the NNBD uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
the one or more CNNs uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 25 is subject-matter ineligible.
Regarding claim 28, claim 28 recites substantially similar limitations to claim 7, and is therefore rejected under the same analysis.
Regarding Claim 30:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 30 recites
generate gradient information (This limitation is a mental process as it encompasses a human mentally generating gradient information and is thus an evaluation.)
Therefore, claim 30 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 30 further recites additional elements of
the first NN includes one or more Multi-layer Perceptrons (MLPs) (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
the second NN (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
output the modified graph based on the gradient information generated by the one or more MLPs. (This element does not integrate the abstract idea into a practical application because it recites insignificant extra-solution activity of data gathering (see MPEP 2106.05(g)).)
Therefore, claim 30 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 30 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the first NN includes one or more Multi-layer Perceptrons (MLPs) uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
the second NN uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
output the modified graph based on the gradient information generated by the one or more MLPs is the well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)).
Therefore, claim 30 is subject-matter ineligible.
Regarding Claim 33:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 33 recites
the initial graph and the modified graph are 2 dimensional (2D) point sets (This limitation is a mental process as it encompasses a human mentally making the initial graph and the modified graph 2D point sets and is thus an evaluation.)
the input data representation is a point cloud; (This limitation is a mental process as it encompasses a human mentally making the input data representation a point cloud and is thus an evaluation.)
perform a deforming operation based on the codeword and the 2D point set that is initialized with a pre-determined sampling in a plane (This limitation is a mental process as it encompasses a human mentally determining the preliminary reconstruction of the input data representation by performing a deforming operation and is thus an evaluation.)
Therefore, claim 33 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 33 further recites additional elements of
the first NN (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
Therefore, claim 33 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 33 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the first NN uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 33 is subject-matter ineligible.
Regarding Claim 35:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 35 recites
perform a tearing operation, based on the preliminary reconstruction of the input data representation, the codeword and the initial graph to generate the modified graph. (This limitation is a mental process as it encompasses a human mentally determining the modified graph by performing a tearing operation and is thus an evaluation.)
Therefore, claim 35 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 35 further recites additional elements of
the second NN (This element does not integrate the abstract idea into a practical application because it recites generic computing components on which to perform the abstract idea (see MPEP 2106.05(f)).)
Therefore, claim 35 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
The additional elements of claim 35 do not provide significantly more than the abstract idea itself, taken alone and in combination because
the second NN uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)).
Therefore, claim 35 is subject-matter ineligible.
Regarding Claim 41:
Subject Matter Eligibility Analysis Step 2A Prong 1:
Claim 41 recites
the initial graph is a 2D grid that includes a matrix of points, each point indicating a 2D position; (This limitation is a mental process as it encompasses a human mentally making the initial graph a 2D grid and is thus an evaluation.)
the 2D grid is associated with a manifold, each point indicating a fixed position on the manifold; (This limitation is a mental process as it encompasses a human mentally making the 2D grid associate with a manifold and is thus an evaluation.)
the 2D grid is a fixed set of sampled points from a 2D plane (This limitation is a mental process as it encompasses a human mentally making the 2D grid a fixed set of sampled points and is thus an evaluation.)
Therefore, claim 41 recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
Claim 41 does not further recite any additional elements. Therefore, claim 41 is not integrated into a practical application.
Subject Matter Eligibility Analysis Step 2B:
Since there are no additional elements, claim 41 does not provide significantly more than the abstract idea itself, taken alone and in combination. Therefore, claim 41 is subject-matter ineligible.
Regarding claim 42, claim 42 recites substantially similar limitations to claim 21, and is therefore rejected under the same analysis.
Regarding claim 43, claim 43 recites substantially similar limitations to claim 22, and is therefore rejected under the same 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, 7, 23-24, and 28 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Wang et al. (“Deep Cascade Generation on Point Sets”) (hereafter referred to as Wang).
Regarding claim 1, Wang teaches
A method implemented by a neural network-based decoder (NNBD), comprising: obtaining or receiving, by the NNBD, a codeword, as a descriptor of an input data representation (Wang, page 2, 1st column, 2nd paragraph, “we design a deep cascade model of two encoder-decoders” where “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input” (Wan, page 4, 1st column, 1st paragraph) and Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the codeword is the latent vector that was encoded, the input data representation is the point clouds or RGB Images, and the NNBD is the deep cascade model.);
determining, by a first neural network based on at least the codeword and an initial graph, a preliminary reconstruction of the input data representation (Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape.” Examiner notes that the first neural network is the autoencoder, the codeword is the latent vector, the initial graph is N-dimensional 2D fixed grid points, and the preliminary reconstruction is the collection of N x n 3D points to represent a course surface of 3D object shape.);
modifying the initial graph, by a second neural network based on at least the preliminary reconstruction and the codeword, to generate a modified graph representation (Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape” where “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set….the output layer of the encoder is a graph max pooling layer to take the maximum among the k vertex neighbors” (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph). Examiner notes that the graph G is the modified graph. The modified graph was created by using extracted features from the 3D point set, in which the 3D points, or preliminary reconstruction were outputted by the first neural network or autoencoder by modifying the initial graph, or N-dimensional 2D fixed grid points. Examiner further notes that the codeword or latent vector was used as input into the autoencoder or first neural network. Additionally, the second neural network is the autoencoder.);
and determining, by the first neural network based on at least the codeword and the modified graph, a refined reconstruction of the input data representation (Wang, page 4, 2nd column, section 3.3 An Ensemble of Point Decoders, “As shown in Figure 2, we employ a stack of decoders for a densely fine point-based surface, encouraged by the PointNet++ [Qi et al., 2017b] for 3D shape analysis in a hierarchical learning fashion. Specifically, given a course surface Pl-1 as an input, the surface output of the autoencoder at cascade level l is
∪
D
l
m
, where m is the size of point generators based on multi-layer perceptrons. We use the same network structure of the MLP in the AtlasNet, i.e. four 1D convolution layers with [1024, 512, 256, 3] hidden neurons respectively. Moreover, we apply residual skip-connections between two adjacent cascade levels, which ensures that the positions of coarser points can be propagated and updated through the entire network and incorporated for fine surface generation. Evidently, the size of points in such an ensemble learning manner is linearly proportional to the size m of stacked decoders, and thus evolves more dense surface with cascade levels l increases” and Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the fine point sets or dense surface is the refined reconstruction of the input data representation. Examiner further notes that the modified graph was created in the graph encoder in Figure 2. The decoder in the second stage, or first neural network, generated the fine point sets based on the latent vector or codeword from the first stage and the modified graph from the graph encoder. ),
wherein the modified graph indicates topology information associated with the input data representation (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph, “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set” Examiner notes that the edges and vertices are the topology information. Examiner further notes that the modified graph G was created by using extracted features from the 3D point set, in which the 3D points were outputted by the first neural network or autoencoder which used the input data representation as input.).
Regarding claim 2, Wang teaches
The method of claim 1, wherein the modified graph is determined by combining the initial graph and an output of the second neural network (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph, “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set….the output layer of the encoder is a graph max pooling layer to take the maximum among the k vertex neighbors” and “the encoder E(•) encodes input data into a latent vector θ which is then decoded into 3D geometry to approximate P*” (Wang, page 3, 2nd paragraph) and where “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input” (Wan, page 4, 1st column, 1st paragraph). Examiner notes that the graph G is the modified graph. Examiner further notes that the codeword or latent vector is output of the autoencoder or second neural network. Additionally, both the output of the second neural network and the initial graph are combined as input into the decoder which creates the modified graph.).
Regarding claim 3, Wang teaches
The method of claim 1, wherein the modified graph is a locally connected graph (Wang, page 4, 1st column, last paragraph, “A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set. We employ the edge convolution [Wang et al., 2018b] on such a k-NN graph. If there exists an edge eij connecting a vertex vi and tis neighbor vertex vj, we get an edge feature gij by applying a nonlinear function h{•,•} with learnable parameters Θ on vertex vi and edge eij. As a result, each vertex having k nearest neighbors will generate a P-dimensional feature as follows:
PNG
media_image2.png
70
348
media_image2.png
Greyscale
Where hΘ denotes a MLP mapping and N(i) is a set of local neighbors’ indexes around vertex vi.” Examiner notes that the graph uses k nearest neighbor and has local neighbors around vertices which makes a locally connected graph.).
Regarding claim 7, Wang teaches
The method of claim 1, wherein: the NNBD is a Graph Conditioned NNBD (Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the NNBD is the DCG net. Examiner further notes that since the NNBD uses a graph encoder, it is graph conditioned.);
and the determining of the refined reconstruction of the input data representation is performed via a plurality of iterative operations of at least the first neural network (Wang, page 4, 2nd column, last paragraph, “Specifically, in our two-stage cascade model (L=2), the weights w1 and w2 for their corresponding losses are as
PNG
media_image3.png
37
392
media_image3.png
Greyscale
where k is the current number of iterations during training.” Examiner notes that the iterations during training are the plurality of iterative operations. Examiner further notes that the first neural network is encompassed within the two-stage cascade model.).
Regarding claim 23, Wang teaches
A neural network-based decoder (NNBD), comprising: a receiver configured to receive or obtain a codeword, as a descriptor of an input data representation (Wang, page 2, 1st column, 2nd paragraph, “we design a deep cascade model of two encoder-decoders” where “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input” (Wan, page 4, 1st column, 1st paragraph) and Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the codeword is the latent vector that was encoded, the input data representation is the point clouds or RGB Images, the receiver unit is the decoder and the NNBD is the deep cascade model.);
a first neural network (NN) configured to: determine, based on at least the codeword and an initial graph, a preliminary reconstruction of the input data representation (Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape.” Examiner notes that the first neural network is the autoencoder, the codeword is the latent vector, the initial graph is N-dimensional 2D fixed grid points, and the preliminary reconstruction is the collection of N x n 3D points to represent a course surface of 3D object shape.);
and a second NN configured to modify the initial graph, based on at least the preliminary reconstruction and the codeword, to generate a modified graph (Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape” where “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set….the output layer of the encoder is a graph max pooling layer to take the maximum among the k vertex neighbors” (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph). Examiner notes that the graph G is the modified graph. The modified graph was created by using extracted features from the 3D point set, in which the 3D points, or preliminary reconstruction were outputted by the first neural network or autoencoder by modifying the initial graph, or N-dimensional 2D fixed grid points. Examiner further notes that the codeword or latent vector was used as input into the autoencoder or first neural network. Additionally, the second neural network is the autoencoder.),
wherein: the first NN is further configured to determine, based on at least the codeword and the modified graph, a refined reconstruction of the input data representation (Wang, page 4, 2nd column, section 3.3 An Ensemble of Point Decoders, “As shown in Figure 2, we employ a stack of decoders for a densely fine point-based surface, encouraged by the PointNet++ [Qi et al., 2017b] for 3D shape analysis in a hierarchical learning fashion. Specifically, given a course surface Pl-1 as an input, the surface output of the autoencoder at cascade level l is
∪
D
l
m
, where m is the size of point generators based on multi-layer perceptrons. We use the same network structure of the MLP in the AtlasNet, i.e. four 1D convolution layers with [1024, 512, 256, 3] hidden neurons respectively. Moreover, we apply residual skip-connections between two adjacent cascade levels, which ensures that the positions of coarser points can be propagated and updated through the entire network and incorporated for fine surface generation. Evidently, the size of points in such an ensemble learning manner is linearly proportional to the size m of stacked decoders, and thus evolves more dense surface with cascade levels l increases” and Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the fine point sets or dense surface is the refined reconstruction of the input data representation. Examiner further notes that the modified graph was created in the graph encoder in Figure 2. The decoder in the second stage, or first neural network, generated the fine point sets based on the latent vector or codeword from the first stage and the modified graph from the graph encoder. ),
and the modified graph indicates topology information associated with the input data representation (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph, “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set” Examiner notes that the edges and vertices are the topology information. Examiner further notes that the modified graph G was created by using extracted features from the 3D point set, in which the 3D points were outputted by the first neural network or autoencoder which used the input data representation as input.).
Regarding claim 24, claim 24 recites substantially similar limitations to claim 3, and is therefore rejected under the same analysis.
Regarding claim 28, claim 28 recites substantially similar limitations to claim 7, and is therefore rejected under the same analysis.
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.
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) 4, 12, 14, 21-22, 25, 33, 35, and 41-43 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang in view of Yang et al. (“FoldingNet: Interpretable Unsupervised Learning on 3D Point Clouds”) (hereafter referred to as Yang).
Regarding claim 4, Wang teaches the method of claim 1. Wang further teaches
generating a …matrix for processing by one or more convolutional neural networks (CNNs) (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally. We follow the settings in [Groueix et al., 2018] for point set generation and shape autoencoding, i.e., ResNet-18 [He et al., 2016] and PointNet [Qi et al., 2017a] for feature encoding on images or point clouds respectively. Specifically, the ResNet-18 contains four residual blocks followed by one fully connected layer, and each block consists of five 2D convolution layers, while the PointNet has four layers with three 1D convolution layers and on fully-connected layer. Inspired by the FoldingNet [Yang et al., 2018], we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input.” Examiner notes that the MLPs with convolutional layers are CNNs. Examiner further notes that the CNNs process 2D grids or matrices. )
Wang does not teach, but Yang does teach
generating a concatenation matrix for processing … by concatenating at least a replicated codeword, the initial graph and the preliminary reconstruction of the input data representation (Yang, page 4, 2nd column, 3rd – 4th paragraph, “Before we feed the codeword into the decoder, we replicate it m times and concatenate the m-by-512 matrix with an m-by-2 matrix that contains the m grid points on a square centered at the origin. The result of the concatenation is a matrix of size m-by-514. The matrix is processed row-wise by a 3-layer perceptron and the output is a matrix of size m-by-3. After that, we again concatenate the replicated codewords to the m-by-3 output and feed it into a 3-layer perceptron. This output is the reconstructed point cloud. The parameter n is set as per the input point cloud size, e.g. n = 2048 in our experiments, which is the same as [1].We choose m grid points in a square, so m is chosen as 2025 which is the closest square number to 2048. [4th paragraph] Notice that the concatenation of the codewords to the 2-dimensional grids, followed by a 3-layer perceptron essentially implements a folding operation” and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx512 matrix is the replicated codewords, the mx2 matrix is the initial graph. Examiner further notes that those two matrices are concatenated to form a mx514 matrix. Examiner further notes that the mx3 matrix or the preliminary reconstructed of the input data representation is also concatenated to the codewords. By feeding the mx514 matrix into a 3 layer perceptron and concatenating the codewords to the mx3 matrix output, the replicated codewords, initial graph and preliminary reconstructed of the input data representation are concatenated to form an mx515 matrix or concatenation matrix.)
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified the matrix in Wang to be a concatenation matrix. Doing so is advantageous because “this design makes the proposed decoder much smaller in parameter size” (Yang, page 2, 1st column, last paragraph).
Regarding claim 12, teaches the method of claim 1. Wang further teaches
the initial graph and the modified graph are 2 dimensional (2D) point sets ((Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape” and “A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set. We employ the edge convolution [Wang et al., 2018b] on such a k-NN graph. If there exists an edge eij connecting a vertex vi and tis neighbor vertex vj, we get an edge feature gij by applying a nonlinear function h{•,•} with learnable parameters Θ on vertex vi and edge eij. As a result, each vertex having k nearest neighbors will generate a P-dimensional feature as follows:
PNG
media_image2.png
70
348
media_image2.png
Greyscale
Where hΘ denotes a MLP mapping and N(i) is a set of local neighbors’ indexes around vertex vi” (Wang, page 4, 1st column, last paragraph). Examiner notes that the initial graph is N-dimensional 2D fixed grid points, and the modified graph is graph G in C dimensions. Examiner notes that the N-dimensional and C dimensions could be 2 dimensions for the initial graph and the modified graph.);
the input data representation is a point cloud (Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the input data representation is a point cloud.);
Wang does not teach, but Yang does teach
and the determining of the preliminary reconstruction of the input data representation includes performing a deforming operation based on the codeword and the 2D point set that is initialized with a pre-determined sampling in a plane (Yang, page 2, 2nd column, last paragraph – page 3, 1st column, 1st paragraph, “First, the proposed folding architecture is designed with a clear geometric interpretation: we want to impose a “virtual force” to deform/cut/stretch a 2D grid lattice onto a 3D object surface, while such a deformation force should be influenced or regulated by interconnections induced by the lattice neighborhood” where “We are going to show that a 2D grid structure is not only a sampling structure for imaging, but can indeed be used to construct a point cloud through the proposed folding operation. This is based on the observation that the 3D point clouds of our interest are obtained from object surfaces: either discretized from boundary representations in CAD and computer graphics, or sampled from line-of-sight sensors like LIDAR. Intuitively, any 3D object surface can be transformed to a 2D plane through certain operations like cutting, squeezing, and stretching. The inverse procedure is to glue those 2D point samples back onto an object surface via certain folding operations which are initialized as 2D grid samples” (Yang, page 2, 1st column, 1st paragraph). and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the intermediate point cloud is the preliminary reconstruction of the input data representation, the deforming operation is the 1st folding which uses the codeword and 2D grid points or 2D point set. Examiner further notes that the 2D point set is initialized as 2D grid samples which are in a plane and created from a predetermined 3D point cloud.).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to use a deforming operation. Doing so is advantageous because “folding can achieve higher classification accuracy than other unsupervised methods” (Yang, page 3, 1st column, 3rd bullet point).
Regarding claim 14, Wang teaches the method of claim 1. Wang does not teach, but Yang does teach
wherein the determining of the modified graph includes: performing a tearing operation, based on the preliminary reconstruction of the input data representation, the codeword and the initial graph to generate the modified graph (Yang, page 2, 2nd column, last paragraph – page 3, 1st column, 1st paragraph, “First, the proposed folding architecture is designed with a clear geometric interpretation: we want to impose a “virtual force” to deform/cut/stretch a 2D grid lattice onto a 3D object surface, while such a deformation force should be influenced or regulated by interconnections induced by the lattice neighborhood” where “We are going to show that a 2D grid structure is not only a sampling structure for imaging, but can indeed be used to construct a point cloud through the proposed folding operation. This is based on the observation that the 3D point clouds of our interest are obtained from object surfaces: either discretized from boundary representations in CAD and computer graphics, or sampled from line-of-sight sensors like LIDAR. Intuitively, any 3D object surface can be transformed to a 2D plane through certain operations like cutting, squeezing, and stretching. The inverse procedure is to glue those 2D point samples back onto an object surface via certain folding operations which are initialized as 2D grid samples” (Yang, page 2, 1st column, 1st paragraph). and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the intermediate point cloud is the preliminary reconstruction of the input data representation and modified graph, the tearing operation is the 1st folding which uses the codeword and 2D grid points or initial graph. Examiner further notes that the tearing operation is interpreted similarly to paragraph 0124 of the instant specification, “The final output matrix N x 2 665 of the Tearing/T-Net module 556 may represent a modification/evolvement of the 2Dgrid 540(e.g., 2D grid x).”).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to perform a tearing operation. Doing so is advantageous because “folding can achieve higher classification accuracy than other unsupervised methods” (Yang, page 3, 1st column, 3rd bullet point).
Regarding claim 21, Wang teaches the method of claim 1. Wang does not teach, but Yang does teach
wherein the determining of the modified graph includes:
replicating the received or obtained codeword K times to generate a KxD codeword matrix, wherein K is a number of nodes in the initial graph and D is a length of the codeword (Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the codeword is replicated m times and generates a mx512 codeword matrix. M maps to K and 512 maps to D.);
concatenating, the KxD codeword matrix and the initial graph, as a KxN matrix, to generate a Kx(D+N) concatenated matrix; (Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx512 codeword matrix and mx2 2D grid points or initial graph are concatenated to form the mx514 concatenated matrix. M maps to K, N maps to 2 and 512 maps to D);
inputting, the concatenated matrix to the second neural network corresponding one or more Convolutional Neural Networks (CNNs) or Multi-layer Perceptrons (MLPs) (Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx514 concatenated matrix is input into the 1st folding or MLPs which is part of the autoencoder, or second neural network.)
generating, by the second neural network from the concatenated matrix, the modified graph (Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the intermediate point cloud is the modified graph and the autoencoder that is shown is the second neural network.);
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to have a concatenation matrix. Doing so is advantageous because “this design makes the proposed decoder much smaller in parameter size” (Yang, page 2, 1st column, last paragraph).
Regarding claim 22, Wang in view of Yang teach the method of claim 21. Yang further teaches
concatenating the codeword matrix to an output of a first set of CNN or MLP layers, as a concatenated intermediary matrix (Yang, page 4, 2nd column, 3rd – 4th paragraph, “Before we feed the codeword into the decoder, we replicate it m times and concatenate the m-by-512 matrix with an m-by-2 matrix that contains the m grid points on a square centered at the origin. The result of the concatenation is a matrix of size m-by-514. The matrix is processed row-wise by a 3-layer perceptron and the output is a matrix of size m-by-3. After that, we again concatenate the replicated codewords to the m-by-3 output and feed it into a 3-layer perceptron. This output is the reconstructed point cloud. The parameter n is set as per the input point cloud size, e.g. n = 2048 in our experiments, which is the same as [1].We choose m grid points in a square, so m is chosen as 2025 which is the closest square number to 2048. [4th paragraph] Notice that the concatenation of the codewords to the 2-dimensional grids, followed by a 3-layer perceptron essentially implements a folding operation” and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx512 is the codeword matrix, and the mx3 intermediate point cloud is the output of a first set of MLP layers. The mx512 codeword matrix and mx3 output are concatenated to form the mx515 matrix or the intermediary matrix.);
and inputting, the concatenated intermediary matrix to a next set of CNN or MLP layers following the first set of CNN or MLP layers (Yang, page 4, 2nd column, 3rd – 4th paragraph, “Before we feed the codeword into the decoder, we replicate it m times and concatenate the m-by-512 matrix with an m-by-2 matrix that contains the m grid points on a square centered at the origin. The result of the concatenation is a matrix of size m-by-514. The matrix is processed row-wise by a 3-layer perceptron and the output is a matrix of size m-by-3. After that, we again concatenate the replicated codewords to the m-by-3 output and feed it into a 3-layer perceptron. This output is the reconstructed point cloud. The parameter n is set as per the input point cloud size, e.g. n = 2048 in our experiments, which is the same as [1].We choose m grid points in a square, so m is chosen as 2025 which is the closest square number to 2048. [4th paragraph] Notice that the concatenation of the codewords to the 2-dimensional grids, followed by a 3-layer perceptron essentially implements a folding operation” and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx515 matrix or the intermediary matrix is input into the 2nd folding or next set of MLP layers following the first MLP layers.).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to have a concatenated intermediary matrix. Doing so is advantageous because “this design makes the proposed decoder much smaller in parameter size” (Yang, page 2, 1st column, last paragraph).
Regarding claim 25, Wang teaches the NNBD of claim 23. Wang further teaches
the second NN includes one or more Convolutional Neural Networks (CNNs) (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally. We follow the settings in [Groueix et al., 2018] for point set generation and shape autoencoding, i.e., ResNet-18 [He et al., 2016] and PointNet [Qi et al., 2017a] for feature encoding on images or point clouds respectively. Specifically, the ResNet-18 contains four residual blocks followed by one fully connected layer, and each block consists of five 2D convolution layers, while the PointNet has four layers with three 1D convolution layers and on fully-connected layer. Inspired by the FoldingNet [Yang et al., 2018], we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input.” Examiner notes that the MLPs with convolutional layers are CNNs. Examiner further notes that the CNNs process 2D grids or matrices. Additionally, the autoencoder that holds the MLPs or CNNs is the second NN)
the NNBD is configured to generate a… matrix (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally. We follow the settings in [Groueix et al., 2018] for point set generation and shape autoencoding, i.e., ResNet-18 [He et al., 2016] and PointNet [Qi et al., 2017a] for feature encoding on images or point clouds respectively. Specifically, the ResNet-18 contains four residual blocks followed by one fully connected layer, and each block consists of five 2D convolution layers, while the PointNet has four layers with three 1D convolution layers and on fully-connected layer. Inspired by the FoldingNet [Yang et al., 2018], we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input.” Examiner notes that the MLPs with convolutional layers are CNNs. Examiner further notes that the CNNs process 2D grids or matrices. )
and the one or more CNNs are configured to process the … matrix and to generate the modified graph or a refined modified graph (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally. We follow the settings in [Groueix et al., 2018] for point set generation and shape autoencoding, i.e., ResNet-18 [He et al., 2016] and PointNet [Qi et al., 2017a] for feature encoding on images or point clouds respectively. Specifically, the ResNet-18 contains four residual blocks followed by one fully connected layer, and each block consists of five 2D convolution layers, while the PointNet has four layers with three 1D convolution layers and on fully-connected layer. Inspired by the FoldingNet [Yang et al., 2018], we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input” and Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the MLPs with convolutional layers are CNNs. Examiner further notes that the CNNs process 2D grids or matrices to generate the course point sets for the graph encoder in which the graph encoder generates the modified graph.)
Wang does not teach, but Yang does teach
generate a concatenation matrix using at least (1) a replicated codeword, (2) the initial graph and (3) the preliminary reconstruction of the input data representation (Yang, page 4, 2nd column, 3rd – 4th paragraph, “Before we feed the codeword into the decoder, we replicate it m times and concatenate the m-by-512 matrix with an m-by-2 matrix that contains the m grid points on a square centered at the origin. The result of the concatenation is a matrix of size m-by-514. The matrix is processed row-wise by a 3-layer perceptron and the output is a matrix of size m-by-3. After that, we again concatenate the replicated codewords to the m-by-3 output and feed it into a 3-layer perceptron. This output is the reconstructed point cloud. The parameter n is set as per the input point cloud size, e.g. n = 2048 in our experiments, which is the same as [1].We choose m grid points in a square, so m is chosen as 2025 which is the closest square number to 2048. [4th paragraph] Notice that the concatenation of the codewords to the 2-dimensional grids, followed by a 3-layer perceptron essentially implements a folding operation” and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx512 matrix is the replicated codewords, the mx2 matrix is the initial graph. Examiner further notes that those two matrices are concatenated to form a mx514 matrix. Examiner further notes that the mx3 matrix or the reconstructed data representation is also concatenated to the codewords. By feeding the mx514 matrix into a 3 layer perceptron and concatenating the codewords to the mx3 matrix output, the replicated codewords, initial graph and reconstructed data representation are concatenated to form an mx515 matrix or concatenation matrix.)
the concatenation matrix (Yang, page 4, 2nd column, 3rd – 4th paragraph, “Before we feed the codeword into the decoder, we replicate it m times and concatenate the m-by-512 matrix with an m-by-2 matrix that contains the m grid points on a square centered at the origin. The result of the concatenation is a matrix of size m-by-514. The matrix is processed row-wise by a 3-layer perceptron and the output is a matrix of size m-by-3. After that, we again concatenate the replicated codewords to the m-by-3 output and feed it into a 3-layer perceptron. This output is the reconstructed point cloud. The parameter n is set as per the input point cloud size, e.g. n = 2048 in our experiments, which is the same as [1].We choose m grid points in a square, so m is chosen as 2025 which is the closest square number to 2048. [4th paragraph] Notice that the concatenation of the codewords to the 2-dimensional grids, followed by a 3-layer perceptron essentially implements a folding operation” and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the mx512 matrix is the replicated codewords, the mx2 matrix is the initial graph. Examiner further notes that those two matrices are concatenated to form a mx514 matrix. Examiner further notes that the mx3 matrix or the reconstructed data representation is also concatenated to the codewords. By feeding the mx514 matrix into a 3 layer perceptron and concatenating the codewords to the mx3 matrix output, the replicated codewords, initial graph and reconstructed data representation are concatenated to form an mx515 matrix or concatenation matrix.)
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified the matrix in Wang to be a concatenation matrix. Doing so is advantageous because “this design makes the proposed decoder much smaller in parameter size” (Yang, page 2, 1st column, last paragraph).
Regarding claim 33, teaches the NNBD of claim 23. Wang further teaches
the initial graph and the modified graph are 2 dimensional (2D) point sets ((Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape” and “A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set. We employ the edge convolution [Wang et al., 2018b] on such a k-NN graph. If there exists an edge eij connecting a vertex vi and tis neighbor vertex vj, we get an edge feature gij by applying a nonlinear function h{•,•} with learnable parameters Θ on vertex vi and edge eij. As a result, each vertex having k nearest neighbors will generate a P-dimensional feature as follows:
PNG
media_image2.png
70
348
media_image2.png
Greyscale
Where hΘ denotes a MLP mapping and N(i) is a set of local neighbors’ indexes around vertex vi” (Wang, page 4, 1st column, last paragraph). Examiner notes that the initial graph is N-dimensional 2D fixed grid points, and the modified graph is graph G in C dimensions. Examiner notes that the N-dimensional and C dimensions could be 2 dimensions for the initial graph and the modified graph.);
the input data representation is a point cloud (Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the input data representation is a point cloud.);
Wang does not teach, but Yang does teach
and the first NN is configured to perform a deforming operation based on the codeword and the 2D point set that is initialized with a pre-determined sampling in a plane (Yang, page 2, 2nd column, last paragraph – page 3, 1st column, 1st paragraph, “First, the proposed folding architecture is designed with a clear geometric interpretation: we want to impose a “virtual force” to deform/cut/stretch a 2D grid lattice onto a 3D object surface, while such a deformation force should be influenced or regulated by interconnections induced by the lattice neighborhood” where “We are going to show that a 2D grid structure is not only a sampling structure for imaging, but can indeed be used to construct a point cloud through the proposed folding operation. This is based on the observation that the 3D point clouds of our interest are obtained from object surfaces: either discretized from boundary representations in CAD and computer graphics, or sampled from line-of-sight sensors like LIDAR. Intuitively, any 3D object surface can be transformed to a 2D plane through certain operations like cutting, squeezing, and stretching. The inverse procedure is to glue those 2D point samples back onto an object surface via certain folding operations which are initialized as 2D grid samples” (Yang, page 2, 1st column, 1st paragraph). and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the intermediate point cloud is the preliminary reconstruction of the input data representation, the deforming operation is the 1st folding which uses the codeword and 2D grid points or 2D point set. Examiner further notes that the 2D point set is initialized as 2D grid samples which are in a plane and created from a predetermined 3D point cloud. Additionally, the first NN is the autoencoder.).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to use a deforming operation. Doing so is advantageous because “folding can achieve higher classification accuracy than other unsupervised methods” (Yang, page 3, 1st column, 3rd bullet point).
Regarding claim 35, Wang teaches the NNBD of claim 33. Wang does not teach, but Yang does teach
wherein the second NN is configured to perform a tearing operation, based on the preliminary reconstruction of the input data representation, the codeword and the initial graph to generate the modified graph (Yang, page 2, 2nd column, last paragraph – page 3, 1st column, 1st paragraph, “First, the proposed folding architecture is designed with a clear geometric interpretation: we want to impose a “virtual force” to deform/cut/stretch a 2D grid lattice onto a 3D object surface, while such a deformation force should be influenced or regulated by interconnections induced by the lattice neighborhood” where “We are going to show that a 2D grid structure is not only a sampling structure for imaging, but can indeed be used to construct a point cloud through the proposed folding operation. This is based on the observation that the 3D point clouds of our interest are obtained from object surfaces: either discretized from boundary representations in CAD and computer graphics, or sampled from line-of-sight sensors like LIDAR. Intuitively, any 3D object surface can be transformed to a 2D plane through certain operations like cutting, squeezing, and stretching. The inverse procedure is to glue those 2D point samples back onto an object surface via certain folding operations which are initialized as 2D grid samples” (Yang, page 2, 1st column, 1st paragraph). and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the intermediate point cloud is the preliminary reconstruction of the input data representation and modified graph, the tearing operation is the 1st folding which uses the codeword and 2D grid points or initial graph. Examiner further notes that the tearing operation is interpreted similarly to paragraph 0124 of the instant specification, “The final output matrix N x 2 665 of the Tearing/T-Net module 556 may represent a modification/evolvement of the 2Dgrid 540(e.g., 2D grid x).” Additionally, the second NN is the autoencoder.).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to perform a tearing operation. Doing so is advantageous because “folding can achieve higher classification accuracy than other unsupervised methods” (Yang, page 3, 1st column, 3rd bullet point).
Regarding claim 41, Wang teaches the NNBD of claim 23. Wang does not explicitly disclose, but Yang does teach
the initial graph is a 2D grid that includes a matrix of points, each point indicating a 2D position (Yang, page 5, 1st column, 4th paragraph, “The main idea is to show that in the worst case, the points in the 2D grid function as a selective logic gate to map the 2D points in the 2D grid to the corresponding 3D points in the point cloud.” Examiner notes that the 2D grid is the initial graph. Examiner further notes that since the 2D points in the grid correspond to 3D points in the point cloud, each point in the grid indicates a 2D position. );
the 2D grid is associated with a manifold, each point indicating a fixed position on the manifold (Yang, page 7, 1st column, 2nd paragraph, “The proposed decoder relies on the folding of an inherently 2D manifold corresponding to the point cloud inside the 3D space” and Yang, page 1, Table 1
PNG
media_image5.png
850
559
media_image5.png
Greyscale
Examiner notes that the 2D grid is associated with a manifold and each point indicates a fixed position on the manifold since the points in the 2D grid from Table 1 must fold the grid in order to change the position of the points. );
and the 2D grid is a fixed set of sampled points from a 2D plane (Yang, page 2, 2nd column, last paragraph – page 3, 1st column, 1st paragraph, “First, the proposed folding architecture is designed with a clear geometric interpretation: we want to impose a “virtual force” to deform/cut/stretch a 2D grid lattice onto a 3D object surface, while such a deformation force should be influenced or regulated by interconnections induced by the lattice neighborhood” where “We are going to show that a 2D grid structure is not only a sampling structure for imaging, but can indeed be used to construct a point cloud through the proposed folding operation. This is based on the observation that the 3D point clouds of our interest are obtained from object surfaces: either discretized from boundary representations in CAD and computer graphics, or sampled from line-of-sight sensors like LIDAR. Intuitively, any 3D object surface can be transformed to a 2D plane through certain operations like cutting, squeezing, and stretching. The inverse procedure is to glue those 2D point samples back onto an object surface via certain folding operations which are initialized as 2D grid samples” (Yang, page 2, 1st column, 1st paragraph). and Yang, page 2, Figure 1,
PNG
media_image4.png
606
1401
media_image4.png
Greyscale
Examiner notes that the 2D point set is initialized as 2D grid samples which are in a plane and created from a predetermined 3D point cloud. By having the 3D point cloud predetermined and sampling grid samples from it, the 2D grid is a fixed set of sampled points from a 2D plane.).
Wang and Yang are considered analogous to the claimed invention because they both use autoencoders to modify a point cloud. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to use a 2D grid associated with a manifold. Doing so is advantageous because “we can see that the folding decoder almost always has a higher accuracy and lower reconstruction loss. Compared to the fully-connected decoder that relies on the unnatural “1D order” of the reconstructed 3D points in 3D space, the proposed decoder relies on the folding of an inherently 2D manifold corresponding to the point cloud inside the 3D space. As we mentioned earlier, this folding operation is more natural than the fully-connected decoder. Moreover, the number of parameters in the fully-connected decoder is 1.52 x107, while the number of parameters in our folding decoder is 1.05x106, which is about 7% of the fully-connected decoder” (Yang, page 7, 1st column, 2nd paragraph).
Regarding claim 42, claim 42 recites substantially similar limitations to claim 21, and is therefore rejected under the same analysis.
Regarding claim 43, claim 43 recites substantially similar limitations to claim 22, and is therefore rejected under the same analysis.
Claim(s) 9 and 30 is/are rejected under 35 U.S.C. 103 as being unpatentable over Wang in view of Windrim et al. (“Unsupervised Feature-Learning for Hyperspectral Data with Autoencoders”) (hereafter referred to as Windrim).
Regarding claim 9, Wang teaches
The method of claim 1, wherein: the NNBD includes one or more Multi-layer Perceptrons (MLPs) (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally.” Examiner notes that AtlasNet is part of the NNBD.);
and the modified graph and the refined reconstruction of the input data representation are further based on …information generated by the one or more MLPs (Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the MLPs generate information for the modified graph created in the graph encoder and the refined reconstruction or fine point set.)
Wang does not explicitly disclose, but Windrim does disclose
… gradient information generated by the one or more MLPs (Windrim, page 3, Background section, 1st paragraph, “An autoencoder is a special case of a Multi-layer Perceptron (MLP), which regresses the input data (or some variant of the input) in its output layer [26]. In doing so, it requires no labelled data, making it unsupervised, and can be used to learn non-linear features from the data. It comprises an encoder stage, a code layer, and a decoder stage. With one or more hidden layers, the encoder stage maps the input data to the code layer, which usually has fewer neurons than the input layer. The code must capture all the vital information needed to reconstruct the input in the output layer, via the decoder stage. The encoder and decoder stages are typically symmetric. The autoencoder, as with the MLP, comprises a network of trainable weights. The weights are optimized using a gradient descent process which aims to minimize an objective function, which usually includes a reconstruction error term and a regularization term.” Examiner notes that the autoencoder, an MLP, provides gradient data through a gradient descent process.).
Wang and Windrim are considered analogous to the claimed invention because they both use autoencoders and MLPs. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to generate gradient information. Doing so is advantageous because “The weights are optimized using a gradient descent process which aims to minimize an objective function, which usually includes a reconstruction error term and a regularization term” (Windrim, page 3, Background section, 1st paragraph).
Regarding claim 30, Wang teaches
The NNBD of claim 23, wherein: the first NN includes one or more Multi-layer Perceptrons (MLPs) configured to generate…information (Wang, page 3, 2nd column, last paragraph – page 4, 1st column, 1st paragraph, “Figure 3 illustrates the deep structure of the AtlasNet, which contains an encoder and n (e.g., five in our experiments) multi-layer perceptron (MLP) decoders, each of which with four fully connected layers aims to predict a parametric surface patch locally.” Examiner notes that AtlasNet is part of the autoencoder or the first NN.);
and the second NN is configured to output the modified graph and the refined reconstruction of the data representation is further based on …information generated by the one or more MLPs (Wang, page 3, Figure 2
PNG
media_image1.png
401
1169
media_image1.png
Greyscale
Examiner notes that the MLPs generate information for the modified graph created in the graph encoder and the refined reconstruction or fine point set. Examiner further notes that the graph encoder is part of the autoencoder or second NN.)
Wang does not explicitly disclose, but Windrim does disclose
… one or more Multi-layer Perceptrons (MLPs) configured to generate gradient information (Windrim, page 3, Background section, 1st paragraph, “An autoencoder is a special case of a Multi-layer Perceptron (MLP), which regresses the input data (or some variant of the input) in its output layer [26]. In doing so, it requires no labelled data, making it unsupervised, and can be used to learn non-linear features from the data. It comprises an encoder stage, a code layer, and a decoder stage. With one or more hidden layers, the encoder stage maps the input data to the code layer, which usually has fewer neurons than the input layer. The code must capture all the vital information needed to reconstruct the input in the output layer, via the decoder stage. The encoder and decoder stages are typically symmetric. The autoencoder, as with the MLP, comprises a network of trainable weights. The weights are optimized using a gradient descent process which aims to minimize an objective function, which usually includes a reconstruction error term and a regularization term.” Examiner notes that the autoencoder, an MLP, provides gradient data through a gradient descent process.).
… gradient information generated by the one or more MLPs (Windrim, page 3, Background section, 1st paragraph, “An autoencoder is a special case of a Multi-layer Perceptron (MLP), which regresses the input data (or some variant of the input) in its output layer [26]. In doing so, it requires no labelled data, making it unsupervised, and can be used to learn non-linear features from the data. It comprises an encoder stage, a code layer, and a decoder stage. With one or more hidden layers, the encoder stage maps the input data to the code layer, which usually has fewer neurons than the input layer. The code must capture all the vital information needed to reconstruct the input in the output layer, via the decoder stage. The encoder and decoder stages are typically symmetric. The autoencoder, as with the MLP, comprises a network of trainable weights. The weights are optimized using a gradient descent process which aims to minimize an objective function, which usually includes a reconstruction error term and a regularization term.” Examiner notes that the autoencoder, an MLP, provides gradient data through a gradient descent process.).
Wang and Windrim are considered analogous to the claimed invention because they both use autoencoders and MLPs. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Wang to generate gradient information. Doing so is advantageous because “The weights are optimized using a gradient descent process which aims to minimize an objective function, which usually includes a reconstruction error term and a regularization term” (Windrim, page 3, Background section, 1st paragraph).
Response to Arguments
The IDS filed 07/23/2025 complies with CFR 1.97.
The previous claim objections have been overcome in light of the of the instant amendments.
The previous 112(b) and 112(a) rejections have been overcome in light of the instant amendments.
On page 9, Applicant argues:
Amended claim 1 recites a method including the steps of determining a preliminary reconstruction of a data input representation and determining a refined reconstruction of the input data representation. Thus, claim 1 is directed to an invention that has a specific application and improves a technology in the marketplace, that is, data processing.
The current application identifies a technical problem with existing autoencoders: the inability to handle complex 3D topologies or a scene with multiple objects (see paragraphs [0085]). The claimed invention provides a technical solution: a novel autoencoder architecture (GCAE with a TearingNet), which may be implemented and may promote a locally-connected graph as a better approximation of the 3D PC topology (see paragraph [0098]).
Thus, Applicant respectfully submits that claim 1 satisfies the requirements of 35 U.S.C. § 101 and recites patentable subject matter. Claim 23 recites limitations similar to claim 1 in an apparatus form, and thus also recites patentable subject matter. The dependent claims also satisfy the requirements of 35 U.S.C. § 101 at least due to their dependencies from the base claims.
Regarding the Applicant’s argument that the claims are directed to an invention that has a specific application and improves a technology in the marketplace, the Examiner respectfully disagrees. Specifically, Examiner respectfully notes that any improvements to the technology come from additional elements (MPEP 2106.04(d)(II)). With this in mind, Examiner respectfully notes that the current additional elements of claim 1 include “implemented by a neural network-based decoder (NNBD)”, “by a first neural network”, and “by a second neural network”, all of which amounts to mere “apply it on a computer” and cannot provide significantly more (MPEP 2106.05(f)). Examiner further notes that another additional element of claim 1 is “obtaining or receiving, by the NNBD, a codeword, as a descriptor of an input data representation” is also within claim 1. This additional element recites insignificant extra-solution activity of data gathering (2106.05(g)) and is further a well understood, routine, and conventional activity of “transmitting or receiving data over a network” (see MPEP 2106.05(d)(II); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network)). Claim 23 is rejected for similar reasons as claim 1 (see 101 rejection above).
Regarding the Applicant’s argument that the dependent claims are allowable at least due in part to their dependency on the independent claims, the Examiner respectfully disagrees and notes the instant rejections and response to arguments regarding the independent claims above.
On page 10, Applicant argues:
Wang proposes a two-stage cascade model to generate 3D geometry of an object on a point cloud. Specifically, Wang's method adopts the point set autoencoder to generate a sparsely coarse shape first, and then locally refines it by encoding neighborhood connectivity on a graph representation, as illustrated in the reproduced figure below. (See Wang, Abstract.)
…
Specifically, Wang uses AtlasNet for coarse shape generation (first stage) and adopts graph convolution based encoding and an ensemble of decoders for shape refinement (second stage). That is, Wang does not teach using the same neural network (the "first neural network") to reconstruct both the preliminary reconstruction and the refined reconstruction of the input data representation as recited in claim 1.
In particular, as illustrated in FIG. 5 (reproduced below), the decoder can include two Folding modules and a Tearing module. The Folding module 550-1 takes the codeword 530 and an initial 2D grid 540 as inputs and outputs a preliminary reconstruction of the point cloud. The Tearing module 556 takes the initial 2D grid 540, the codeword 530 and the preliminary reconstruction of the point cloud as inputs and outputs the modified 2D grid 558 (tom 2D grid). The Folding module 550-2 takes the tom 2D grid 558 and the codeword 530 as inputs and outputs a refined reconstruction of the point cloud. The Folding module 550-1 and 550-2 may share the same neural network architecture and the same neural network parameters.
…
Clearly, the claimed invention uses the same first neural network (Folding module) for obtaining the preliminary reconstruction of the point cloud and the refined reconstruction of the point cloud. However, Wang uses different networks to obtain the preliminary (coarser) reconstruction and the refined reconstruction of the point cloud. That is, Wang uses a different architecture from the claimed invention.
Regarding the Applicant’s argument that the prior art does not teach using the same neural network to reconstruct both the preliminary reconstruction and the refined reconstruction of the input representation, Examiner respectfully disagrees. Specifically, Examiner respectfully notes that Wang uses the autoencoder (or first neural network) from the first stage to create a collection of Nxn 3D points (or a preliminary reconstruction of the input representation) (Wang, page 2, 1st column, 1st paragraph,). Examiner further notes that Wang uses the same MLPs from the first stage autoencoder (or first neural network) to generate the fine point sets (or refined reconstruction of the input representation) (Wang, page 4, 2nd column, section 3.3 An Ensemble of Point Decoders; Wang, page 3, Figure 2). Examiner additionally notes that the folding modules and tearing module are not claimed in claim 1. Rather, a first and second neural network are claimed in which by broadest reasonable interpretation, is interpreted to be the same neural network.
On pages 10-11, Applicant argues:
In addition, the claimed invention teaches using a second neural network to modify the initial graph (2D grid) to a modified graph (torn 2D grid), which will be used as input to the first neural network to refine the point cloud. Nowhere in Wang teaches or suggests modifying the initial graph (e.g., 2D grid) to refine the reconstructed point cloud.
In the Office action, it appears that the Examiner corresponds the Graph Encoder of Wang to the second neural network (Tearing module) in claim 1. However, as clearly illustrated in Figure 2 of Wang, the Graph Encoder accepts coarse point sets as inputs and outputs an (N x n) x (3P + 2C) output vector. This is completely different from the second neural network (Tearing module) which takes the initial graph, the codeword and the preliminary reconstruction of the point cloud as inputs and outputs the modified graph recited in amended claim 1.
Thus, claim 1 is not anticipated by the cited reference. Accordingly, claim 1 is patentably distinguishable over the cited reference for at least the reasons set forth above.
Amended claim 23 recites limitations similar to those recited in amended claim 1. Thus, Applicant respectfully submits that claim 23 is also patentably distinguishable for at least the same reasons as discussed above for claim 1. The dependent claims are also patentably distinguishable over the cited reference at least by virtue of their ultimate dependency from their base claims.
Regarding the Applicant’s argument that the prior art of reference does not teach a modified graph, Examiner respectfully disagrees. Specifically, Examiner notes that Wang teaches modifying the initial graph, by a second neural network based on at least the preliminary reconstruction and the codeword, to generate a modified graph representation (Wang, page 2, 1st column, 1st paragraph, “we use tiled N-dimensional 2D fixed grid points as 2D primitives during reconstruction rather than 2D points via uniformly random sampling, which together with the latent vector encoded by either ResNet-18 or PointNet are fed into the decoder as input. The output dimension of hidden layers in each MLP based decoder is fixed to [1024,512,256,3] followed by ReLU non-linearity operation. Finally, the output of the AtlasNet is a collection of N x n 3D points to represent a coarse surface of 3D object shape” where “we first extract a C-dimensional (C=24) feature for each 3D point via one MLP layer, which contains one 1D convolution layer. A k nearest neighbor (k-NN) graph G = {V, E} in ℝC containing N x n vertices V = {v1,…, vNxn} and E
⊆
VxV is constructed from an unstructured point feature set….the output layer of the encoder is a graph max pooling layer to take the maximum among the k vertex neighbors” (Wang, page 4, 1st column, 2nd paragraph -2nd column, 1st paragraph). Examiner notes that the graph G is the modified graph. The modified graph was created by using extracted features from the 3D point set, in which the 3D points, or preliminary reconstruction were outputted by the first neural network or autoencoder by modifying the initial graph, or N-dimensional 2D fixed grid points. Examiner further notes that the codeword or latent vector was used as input into the autoencoder or first neural network. Additionally, the second neural network is the autoencoder.) Examiner further notes that the second neural network is the entirety of the autoencoder, not just the Graph Encoder. Once again, Examiner respectfully notes that the Tearing module is not claimed in claim 1.
Examiner further notes that claim 23 is rejected similarly to claim 1 (see 102 rejection above).
On page 10, Applicant argues:
Claims 4, 9, 12, 14, 21 and 22 directly or indirectly depend from claim 1 and, thus, include all limitations of claim 1. Similarly, claims 25, 30 33, 35 and 41-43 directly or indirectly depend from claim 23 and, thus, include all limitations of claim 23. Applicant submits that the alleged additional teachings of the above-mentioned references, assuming for argument sake that they are valid, fail to overcome the defect of Wang as applied to the base claims. Accordingly, claims 4, 9, 12, 14, 21, 22, 25, 30, 33, 35 and 41-43 are patentably distinct and non-obvious over the cited references for at least the reasons set forth above with respect to independent claims 1 and 23, respectively. Thus, Applicant respectfully requests reconsideration and withdrawal of the 35 U.S.C. § 103 rejection of these claims.
Regarding the Applicant’s argument that the dependent claims are allowable at least due in part to their dependency on the independent claims, the Examiner respectfully disagrees and notes the instant rejections and response to arguments regarding the independent claims above.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Groueix et al. (“AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation”) also describes using MLPs to process point clouds.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to KAITLYN R LAU whose telephone number is (571)272-1429. The examiner can normally be reached Monday - Thursday: 7:15 am - 5:15 pm EST.
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, Michelle Bechtold can be reached at (571) 431-0762. 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.
/K.R.L./Examiner, Art Unit 2148 /MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights