Prosecution Insights
Last updated: November 04, 2026
Application No. 17/554,097

System and Method For Multi-Task Learning Through Spatial Variable Embeddings

Final Rejection §101§103
Filed
Dec 17, 2021
Priority
Dec 31, 2020 — provisional 63/132,591
Examiner
LAU, KAITLYN RENEE
Art Unit
2148
Tech Center
2100 — Computer Architecture & Software
Assignee
Cognizant Technology Solutions U S Corporation
OA Round
4 (Final)
60%
Grant Probability
Moderate
5-6
OA Rounds
0m
Est. Remaining
99%
With Interview

Examiner Intelligence

Grants 60% of resolved cases
60%
Career Allowance Rate
6 granted / 10 resolved
+5.0% vs TC avg
Strong +67% interview lift
Without
With
+66.7%
Interview Lift
resolved cases with interview
Typical timeline
3y 11m
Avg Prosecution
27 currently pending
Career history
48
Total Applications
across all art units

Statute-Specific Performance

§101
31.2%
-8.8% vs TC avg
§103
30.9%
-9.1% vs TC avg
§102
14.5%
-25.5% vs TC avg
§112
22.4%
-17.6% vs TC avg
Black line = Tech Center average estimate • Based on career data from 10 resolved cases

Office Action

§101 §103
DETAILED ACTION This action is in response to the amendment filed 04/21/2026. Claims 1-4, 6-11, 13-17, and 19-20 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 . Information Disclosure Statement The information disclosure statement filed 01/12/2022 fails to comply with 37 CFR 1.98(a)(2), which requires a legible copy of each cited foreign patent document; each non-patent literature publication or that portion which caused it to be listed; and all other information or that portion which caused it to be listed. The information disclosure statement filed 01/12/2022 listed non patent literature documents that were not in the application file. It has been placed in the application file, but the information referred to therein has not been considered. 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, 6-11, 13-17, and 19-20 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 process, one of the four statutory categories of patentable subject matter. Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 1 recites Measuring the diverse tasks with disjoint input and output variable sets in a shared space z (This limitation is a mental process as it encompasses a human mentally measuring tasks.) for each diverse task, encoding by a function f a value of each observed input variable xi given its shared space location vector zi within shared space z,(This limitation is a mental process as it encompasses a human mentally encoding each task.) aggregating encodings by elementwise addition (This limitation is a mental process as it encompasses a human mentally aggregating encodings.) decoding by a function g the aggregated encodings to predict output variable yj at its location vector zj, wherein zi and zj are variable embeddings (This limitation is a mental process as it encompasses a human mentally decoding a task.) and learning the variable embeddings zi and zj by gradient descent jointly with trainable parameters θf of the neural network encoder and θg of the single neural network decoder (This limitation is a mental process as it encompasses a human mentally learning embeddings and parameters by gradient descent.) Therefore, claim 1 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 1 further recites additional elements of a process, implemented in a computing environment, for training a single model comprised of multiple neural networks across diverse tasks from multiple domains (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)).) wherein the encoding is implemented by a first neural network encoder, wherein a same neural network encoder is used for all diverse tasks (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)).) and further wherein the decoding is implemented by a single neural network decoder(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 a process, implemented in a computing environment, for training a single model comprised of multiple neural networks across diverse tasks from multiple domains uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). wherein the encoding is implemented by a first neural network encoder, wherein a same neural network encoder is used for all diverse tasks uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). and further wherein the decoding is implemented by a single neural network decoder 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 the same abstract ideas as claim 1. Therefore, claim 2 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 2 further recites additional elements of wherein the encoding and decoding are conditioned on the variable embeddings via Feature-wise Linear Modulation (FiLM) layers. (This element does not integrate the abstract idea into a practical application because it recites a technological environment in which to apply a judicial exception (see MPEP 2106.05(h)).) Therefore, claim 2 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 2 do not provide significantly more than the abstract idea itself, taken alone and in combination because wherein the encoding and decoding are conditioned on the variable embeddings via Feature-wise Linear Modulation (FiLM) layers. specifies a particular technological environment to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(h)). Therefore, claim 2 is subject-matter ineligible. Regarding Claim 3: Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 3 recites wherein function g is decomposed (This limitation is a mental process as it encompasses a human mentally re-decomposing a function.) Therefore, claim 3 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 3 further recites additional elements of a core, g1, which is independent of output variable yj(This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) a decoder, g2, which is conditioned on output variable yj (This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) Therefore, claim 3 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 3 do not provide significantly more than the abstract idea itself, taken alone and in combination because a core, g1, which is independent of output variable yj uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). a decoder, g2, which is conditioned on output variable yj uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). Therefore, claim 3 is subject-matter ineligible. Regarding Claim 4: Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 4 recites wherein the single model is in the form of: E [ y j | x ] =   g 2 ( g 1 ( ∑ i = 1 n f ( x , ,   z i ) ) ,   z j ). (This limitation is an abstract idea because it recites a mathematical equation) Therefore, claim 4 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 4 does not recites any additional elements to integrate the abstract idea into a practical application. Therefore, claim 4 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: Since there are no additional elements to provide significantly more than the abstract idea itself, taken alone and in combination, claim 4 is subject-matter ineligible. Regarding Claim 6: Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 6 recites the same abstract ideas as claim 1. Therefore, claim 6 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 6 further recites additional elements of wherein the neural networks are residual block networks (This element does not integrate the abstract idea into a practical application because it recites generic computer components on which to perform the abstract idea (see MPEP 2106.05(f)).) Therefore, claim 6 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 6 do not provide significantly more than the abstract idea itself, taken alone and in combination because wherein the neural networks are residual block networks uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). Therefore, claim 6 is subject-matter ineligible. Regarding Claim 7: Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 7 recites wherein the shared space z is 2-Dimensional (This limitation is an abstract idea because it further describes the mental process of measuring tasks as recited in claim 1.) Therefore, claim 7 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 7 does not recites any additional elements to integrate the abstract idea into a practical application. Therefore, claim 7 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: Since there are no additional elements to provide significantly more than the abstract idea itself, taken alone and in combination, claim 7 is subject-matter ineligible. Regarding Claim 8: Subject Matter Eligibility Analysis Step 1: Claim 8 recites a non-transitory computer readable medium and is thus a product, one of the four statutory categories of patentable subject matter Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 8 recites measuring the diverse tasks with disjoint input and output variable sets in a shared space z (This limitation is a mental process as it encompasses a human mentally measuring tasks.) for each diverse task, encoding by a function f a value of each observed variable xi given its shared space location zi within shared space z (This limitation is a mental process as it encompasses a human mentally encoding each task.) aggregating encodings by elementwise addition (This limitation is a mental process as it encompasses a human mentally aggregating encodings.) decoding by a function g the aggregated encodings to predict yj at its location zj, wherein zi and zj are variable embeddings (This limitation is a mental process as it encompasses a human mentally decoding a task.) and learning the variable embeddings zi and zj by gradient descent jointly with trainable parameters θf of the neural network encoder and θg of the single neural network decoder (This limitation is a mental process as it encompasses a human mentally learning embeddings and parameters by gradient descent.) Therefore, claim 8 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 8 further recites additional elements of At least one computer-readable medium storing instructions that, when executed by a computer, perform a process for training a single model across diverse tasks (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)).) Wherein the encoding is implemented by a neural network encoder, wherein a same neural network encoder is used for all diverse tasks (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)).) And further wherein the decoding is implemented by a single neural network decoder (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 8 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 8 do not provide significantly more than the abstract idea itself, taken alone and in combination because At least one computer-readable medium storing instructions that, when executed by a computer, perform a process for training a single model across diverse tasks uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). Wherein the encoding is implemented by a neural network encoder, wherein a same neural network encoder is used for all diverse tasks uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). And further wherein the decoding is implemented by a single neural network decoder uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). Therefore, claim 8 is subject-matter ineligible. Regarding claim 9, claim 9 recites substantially similar limitations as claim 2, and is therefore rejected under the same analysis. Regarding claim 10, claim 10 recites substantially similar limitations as claim 3, and is therefore rejected under the same analysis. Regarding claim 11, claim 11 recites substantially similar limitations as claim 4, and is therefore rejected under the same analysis. Regarding claim 13, claim 13 recites substantially similar limitations as claim 6, and is therefore rejected under the same analysis. Regarding claim 14, claim 14 recites substantially similar limitations as claim 7, and is therefore rejected under the same analysis. Regarding Claim 15: Subject Matter Eligibility Analysis Step 1: Claim 15 recites a single universal prediction model, thus a product, one of the four statutory categories of patentable subject matter. Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 15 recites aggregating the encoder outputs (This limitation is a mental process as it encompasses a human mentally aggregating outputs.) generating… a prediction yj given its location in the shared space z. (This limitation is a mental process as it encompasses a human mentally generating a prediction.) wherein the shared space vectors zi and zj are variable embeddings learned by gradient descent jointly with trainable parameters of the encoder and the decoder (This limitation is a mental process as it encompasses a human mentally learning embeddings and parameters by gradient descent.) Therefore, claim 15 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 15 further recites additional elements of A single universal prediction model comprised of multiple neural networks trained across diverse tasks from multiple domains in a shared space with disjoint input and output variable sets (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)).) An encoder, f, which is conditioned on a shared space vector zi (This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) generating by a neural network encoder an encoder output for a task variable xi given its location in the shared space z, wherein a same neural network encoder is used for all diverse tasks (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)).) an aggregator (This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) a core, g1, which is independent of output (This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) a decoder, g2, which is conditioned on shared space vector zj (This element does not integrate the abstract idea into a practical application because it recites a generic computing component on which to perform the abstract idea (see MPEP 2106.05(f)).) generating by a single neural network decoder (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 15 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 15 do not provide significantly more than the abstract idea itself, taken alone and in combination because A single universal prediction model comprised of multiple neural networks trained across diverse tasks from multiple domains in a shared space with disjoint input and output variable sets uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). An encoder, f, which is conditioned on shared space vector zi sets uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). generating by a neural network encoder an encoder output for a task variable xi given its location in the shared space z, wherein a same neural network encoder is used for all diverse tasks 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)). an aggregator uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). a core, g1, which is independent of output uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). a decoder, g2, which is conditioned on shared space vector zj uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). generating by a single neural network decoder uses a computer as a tool to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(f)). Therefore, claim 15 is subject-matter ineligible. Regarding claim 16, claim 16 recites substantially similar limitations as claim 4, and is therefore rejected under the same analysis. Regarding Claim 17: Subject Matter Eligibility Analysis Step 2A Prong 1: Claim 17 recites the same abstract ideas as claim 15. Therefore, claim 17 recites an abstract idea. Subject Matter Eligibility Analysis Step 2A Prong 2: Claim 17 further recites additional elements of wherein vector zi and zj are variable embeddings. (This element does not integrate the abstract idea into a practical application because it recites a technological environment in which to apply a judicial exception (see MPEP 2106.05(h)).) Therefore, claim 17 is not integrated into a practical application. Subject Matter Eligibility Analysis Step 2B: The additional elements of claim 17 do not provide significantly more than the abstract idea itself, taken alone and in combination because wherein vector zi and zj are variable embeddings specifies a particular technological environment to perform the abstract idea and cannot provide significantly more (see MPEP 2106.05(h)). Therefore, claim 17 is subject-matter ineligible. Regarding claim 19, claim 19 recites substantially similar limitations as claim 6, and is therefore rejected under the same analysis. Regarding claim 20, claim 20 recites substantially similar limitations as claim 2, and is therefore rejected under the same analysis. Claim Rejections - 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows: 1. Determining the scope and contents of the prior art. 2. Ascertaining the differences between the prior art and the claims at issue. 3. Resolving the level of ordinary skill in the pertinent art. 4. Considering objective evidence present in the application indicating obviousness or nonobviousness. This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention. Claim(s) 1, 7, 8, and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kuga et al. (“Multi-task Learning using Multi-modal Encoder-Decoder Networks with Shared Skip Connections”) (hereafter referred to as Kuga) in view of Rabideau et al. (US 2022/0383994 A1) (hereafter referred to as Rabideau), Malone et al. (US 2019/0325995 A1) (hereafter referred to as Malone), and Sawruk et al. (US 2018/0341702 A1) (hereafter referred to as Sawruk). Regarding claim 1, Kuga teaches A process, implemented in a computing environment, for training a single model comprised of multiple neural networks across diverse tasks from multiple domains (Kuga, page 403, abstract, “We propose a multi-modal encoder-decoder networks to harness the multi-modal nature of multi-task scene recognition. In additional to the shared latent representation among encoder-decoder pairs, our model also has shared skip connections from different encoders” where “we describe details of the cost functions defined for semantic label decoder, image decoder, and depth decoder” (Kuga, page 406, 1st column, 3rd paragraph) and where “we discuss details of the proposed network by taking the task of depth and semantic label estimation from RGB images assuming a training dataset consisting with three modalities: image, depth and semantic labels” (Kuga, page 405, 1st column last paragraph). Examiner notes that the diverse tasks from multiple domains are the label decoding, image decoding, and depth decoding.) , comprising: Measuring the diverse tasks with disjoint input and output variable sets in a shared space (Kuga, page 406, 1st column, 3rd paragraph, “we describe details of the cost functions defined for semantic label decoder, image decoder, and depth decoder” where “we discuss details of the proposed network by taking the task of depth and semantic label estimation from RGB images assuming a training dataset consisting with three modalities: image, depth and semantic labels” (Kuga, page 405, 1st column last paragraph) and “to exploit the commonality among different tasks, all encoder/decoder pairs are connected with each other via the shared latent representation” (Kuga, page 404, 2nd column, last paragraph). Examiner notes that the diverse tasks with disjoint input and output are the tasks, where the diverse tasks are the image, depth, and semantic labels. Examiner further notes that measuring the tasks is the cost functions and the shared space is latent representation.); for each diverse task, encoding by a function f a value of each observed input variable xi given its shared space location vector zi within shared space z, wherein the encoding is implemented by a neural network encoder (Kuga, page 405, 2nd column, last paragraph, “all encoder/decoder pairs are connected with the shared latent representation. Letting x ∈{xi, xs, xd } be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei,Es,Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi),Es(xs),Ed(xd)}.” Examiner notes that for each modalities, or task, a function E is encoding an input value x where the input value x is the value of each observed variable. Examiner further notes that the shared space location is the shared latent representation and the encoder is the first neural network.); wherein a same neural network encoder is used for all diverse tasks (Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that Figure 1 shows the encoder in the left box. ) and decoding by a function g the aggregated encodings to predict output variable yj at its location vector zj within shared space z, wherein zi and zj are variable embeddings, and further wherein the decoding is implemented by a single neural network decoder(Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph) and Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that the aggregated encodings are the max pooled feature maps, the prediction is the class probability, and the location is the feature map or latent representation. Examiner further notes that the feature map is created by embedding values, and so the feature map is made of the variable embeddings. Additionally, the decoder is the second neural network shown in Figure 1 in the right circle.). Kuga teaches encoding by a function f a value of each observed input variable xi given its shared space location …zi within shared space z and decoding by a function g the aggregated encodings to predict output variable yj at its location…zj within shared space z. Kuga does not teach that the latent representation is a vector. Rabideau teaches encoding … given its shared space location vector zi (Rabideau, page 10, paragraph 0054, “In the Variational AutoEncoder (VAE) described herein, the learned output of the encoder is a latent vector in a latent representation space 107.”) decoding … to predict output variable yj at its location vector zj (Rabideau, page 10, paragraph 0054, “decoder 108 similarly constructed as a multilayered neural network, has an input layer 119 (in this the input layer receives a latent vector representation from the latent representation space 107) a hidden layer 120 and an output layer 121.”) Kuga and Rabideau are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date to have modified Kuga to use vectors as latent representations like in Rabideau. Doing so allows for “the loss for this objective function is found such that backpropagation can be performed in order to adjust the weights on the Variational AutoEncoder’s computational graph such that this objective function is maximized” (Rabideau, page 12, paragraph 0067). Kuga in view of Rabideau does not teach, but Malone does teach aggregating encodings by elementwise addition (Malone, page 18, paragraph 0134, “For each input, the output vectors of the encoding functions are added element-wise and the resulting vector used in the margin-based contrastive loss function”. Examiner notes that the elementwise addition is summing the encodings.); Kuga, Rabideau and Malone are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga and Rabideau to aggregate encodings by elementwise addition like in Malone. Doing so “allows to learn embeddings for the missing data values” (Malone, page 14, paragraph 0099). Kuga in view of Rabideau and Malone does not teach, but Sawruk does teach and learning the variable embeddings zi and zj by gradient descent jointly with trainable parameters θf of the neural network encoder and θg of the single neural network decoder (Sawruk, page 41, paragraph 0091, The model is updated using gradient descent, an optimization algorithm that numerically tries to find the minimum of a function. The function in this case is the neural network itself. The neural network is then trained for several iterations. The inner matrix W1 now holds the embedding vectors. This is called the embedding matrix” where “This model is known as a neural machine translation (NMT). The first neural network is called the encoder and transforms each input phrase into a numerical representation. The second neural network is called the decoder and transforms the numerical representation into output text, such as keywords describing the phrase” (Sawruk, page 44, paragraph 0101) and where “After the neural network is trained, a weight matrix is extracted where each row (or column) of the weight matrix corresponds to each element in the vocabulary 558. These are called the embeddings” (Sawruk, page 36, paragraph 0076). Examiner notes that the embeddings are within weight matrix. Examiner further notes that the embeddings are the parameters of the encoder and the decoder. Since the embeddings are the parameters, the embeddings and parameters are learned jointly by gradient descent.) Kuga, Rabideau, Malone, and Sawruk are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga, Rabideau, and Malone to use gradient descent like in Sawruk. Doing so “causes similar items to cluster together in the 5-dimensional space. Accordingly, the embedding vectors encode contextual information and the vector shown above represents a semantically meaningful encoding” (Sawruk, page 41, paragraph 0093). Regarding claim 7, Kuga in view of Rabideau, Malone and Sawruk teaches the process according to claim 1. Kuga further teaches wherein the shared space z is 2-Dimensional (Kuga, page 405, 2nd column, 1st paragraph, “Two max-pooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input” where “the input image size is fixed to 96 x 96” (Kuga, page 406, 2nd column, last paragraph). Examiner notes that since the latent representation is created from pooling a 2D input image, the latent representation, or shared space is also 2 dimensional. ). Regarding claim 8, Kuga teaches perform a process for training a single model comprised of multiple neural networks across diverse tasks from multiple (Kuga, page 403, abstract, “We propose a multi-modal encoder-decoder networks to harness the multi-modal nature of multi-task scene recognition. In additional to the shared latent representation among encoder-decoder pairs, our model also has shared skip connections from different encoders” where “we describe details of the cost functions defined for semantic label decoder, image decoder, and depth decoder” (Kuga, page 406, 1st column, 3rd paragraph) and where “we discuss details of the proposed network by taking the task of depth and semantic label estimation from RGB images assuming a training dataset consisting with three modalities: image, depth and semantic labels” (Kuga, page 405, 1st column last paragraph). Examiner notes that the diverse tasks from multiple domains are the label decoding, image decoding, and depth decoding.), comprising: measuring the diverse tasks with disjoint input and output variable sets in a shared space z (Kuga, page 406, 1st column, 3rd paragraph, “we describe details of the cost functions defined for semantic label decoder, image decoder, and depth decoder” where “we discuss details of the proposed network by taking the task of depth and semantic label estimation from RGB images assuming a training dataset consisting with three modalities: image, depth and semantic labels” (Kuga, page 405, 1st column last paragraph) and “to exploit the commonality among different tasks, all encoder/decoder pairs are connected with each other via the shared latent representation” (Kuga, page 404, 2nd column, last paragraph). Examiner notes that the tasks with disjoint input and output are the tasks, where the tasks are the image, depth, and semantic labels. Examiner further notes that measuring the tasks is the cost functions and the shared space is latent representation.); for each diverse task, encoding by a function f a value of each observed input variable xi given its shared space location vector zi within shared space z, wherein the encoding is implemented by a neural network encoder(Kuga, page 405, 2nd column, last paragraph, “all encoder/decoder pairs are connected with the shared latent representation. Letting x ∈{xi, xs, xd } be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei,Es,Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi),Es(xs),Ed(xd)}.” Examiner notes that for each modalities, or task, a function E is encoding an input value x where the input value x is the value of each observed variable. Examiner further notes that the shared space location is the shared latent representation and the encoder is the first neural network.); wherein a same neural network encoder is used for all diverse tasks (Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that Figure 1 shows the encoder in the left box. ) and decoding by a function g the aggregated encodings to predict output variable yj at its location vector zj within shared space z, wherein zi and zj are variable embeddings, and further wherein the decoding is implemented by a single neural network decoder(Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph) and Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that the aggregated encodings are the max pooled feature maps, the prediction is the class probability, and the location is the feature map or latent representation. Examiner further notes that the feature map is created by embedding values, and so the feature map is made of the variable embeddings. Additionally, the decoder is the second neural network shown in Figure 1 in the right circle.). Kuga teaches encoding by a function f a value of each observed input variable xi given its shared space location …zi within shared space z and decoding by a function g the aggregated encodings to predict output variable yj at its location…zj within shared space z. Kuga does not teach that the latent representation is a vector. Rabideau teaches encoding … given its shared space location vector zi (Rabideau, page 10, paragraph 0054, “In the Variational AutoEncoder (VAE) described herein, the learned output of the encoder is a latent vector in a latent representation space 107.”) decoding … to predict output variable yj at its location vector zj (Rabideau, page 10, paragraph 0054, “decoder 108 similarly constructed as a multilayered neural network, has an input layer 119 (in this the input layer receives a latent vector representation from the latent representation space 107) a hidden layer 120 and an output layer 121.”) Kuga and Rabideau are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date to have modified Kuga to use vectors as latent representations like in Rabideau. Doing so allows for “the loss for this objective function is found such that backpropagation can be performed in order to adjust the weights on the Variational AutoEncoder’s computational graph such that this objective function is maximized” (Rabideau, page 12, paragraph 0067). Kuga in view of Rabideau does not teach, but Malone does teach at least one non-transitory computer-readable medium storing instructions that when executed by a computer perform a process (Malone, page 11, paragraph 0035, “The present invention provides a tangible, non-transitory computer-readable medium having instructions thereon which, upon execution by one or more processors with memory, provide for execution of a method”.) aggregating encodings by elementwise addition (Malone, page 18, paragraph 0134, “For each input, the output vectors of the encoding functions are added element-wise and the resulting vector used in the margin-based contrastive loss function”. Examiner notes that the elementwise addition is summing the encodings.); Kuga, Rabideau and Malone are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga and Rabideau to aggregate encodings by elementwise addition like in Malone. Doing so “allows to learn embeddings for the missing data values” (Malone, page 14, paragraph 0099). Additionally, it would have been obvious to one having ordinary skill in the art prior to the effective filing date of the claimed invention to have implemented Kuga and Rabideau on a computer-readable medium storing instructions executed by a computer. Thus, this would be applying a known technique (multi-modal encoder-decoder networks) to a known device (computing device such as a processor) ready for improvement to yield predictable results (scene recognition outputs) (MPEP 2143 I. (C) Use of known technique to improve similar devices (methods, or products) in the same way). Kuga in view of Rabideau and Malone does not teach, but Sawruk does teach and learning the variable embeddings zi and zj by gradient descent jointly with trainable parameters θf of the neural network encoder and θg of the single neural network decoder (Sawruk, page 41, paragraph 0091, The model is updated using gradient descent, an optimization algorithm that numerically tries to find the minimum of a function. The function in this case is the neural network itself. The neural network is then trained for several iterations. The inner matrix W1 now holds the embedding vectors. This is called the embedding matrix” where “This model is known as a neural machine translation (NMT). The first neural network is called the encoder and transforms each input phrase into a numerical representation. The second neural network is called the decoder and transforms the numerical representation into output text, such as keywords describing the phrase” (Sawruk, page 44, paragraph 0101) and where “After the neural network is trained, a weight matrix is extracted where each row (or column) of the weight matrix corresponds to each element in the vocabulary 558. These are called the embeddings” (Sawruk, page 36, paragraph 0076). Examiner notes that the embeddings are within weight matrix. Examiner further notes that the embeddings are the parameters of the encoder and the decoder. Since the embeddings are the parameters, the embeddings and parameters are learned jointly by gradient descent.) Kuga, Rabideau, Malone, and Sawruk are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga, Rabideau, and Malone to use gradient descent like in Sawruk. Doing so “causes similar items to cluster together in the 5-dimensional space. Accordingly, the embedding vectors encode contextual information and the vector shown above represents a semantically meaningful encoding” (Sawruk, page 41, paragraph 0093). Regarding claim 14, claim 14 recites substantially similar limitations as claim 7, and is therefore rejected under the same analysis. Claim(s) 2 and 9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kuga in view of Rabideau, Malone, and Sawruk in further view of Meseguer-Brocal et al. (“Conditioned-U-Net: Introducing a Control Mechanism in the U-Net for Multiple Source Separations”) (hereafter referred to as Brocal). Regarding claim 2, Kuga in view of Rabideau, Malone, and Sawruk teach the process according to claim 1. Kuga in view of Rabideau, Malone and Sawruk do not teach, but Brocal does teach wherein the encoding and decoding are conditioned on the variable embeddings via Feature-wise Linear Modulation (FiLM) layers (Brocal, page 3, 1st column, 5th paragraph, “our conditioned network is a standard U-Net that computes a set of generic source separation filters that we use to separate the various instruments. It adapts itself through the control mechanism (the condition generator) with FiLM layers inserted at different depths. Our external data is a condition vector…(a one-hot-encoding) which specify the instrument to be separated….The vector…is the input to the control mechanism/condition generator.” where “a FiLM layer is inserted inside each encoding block” (Brocal, page 4, 1st column, 2nd paragraph) and where “the decoder maps the encoder, with 6 decoders blocks” (Brocal, page 4, 1st column, 1st paragraph). Examiner notes that the variable embeddings is the condition vector.). Kuga in view of Rabideau, Malone, Sawruk and Brocal are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, Malone and Sawruk to use FiLM layers like in Brocal. Doing so, “can do several source separation tasks without losing performance as it does not introduce any limitation and makes use of the commonalities of the distinct instruments” (Brocal, page 6, 2nd column, 3rd paragraph). Regarding claim 9, claim 9 recites substantially similar limitations as claim 2, and is therefore rejected under the same analysis. Claim(s) 3-4, 6, 10-11, and 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kuga in view of Rabideau, Malone, and Sawruk in further view of Li et al. (“Autoencoder Based Residual Deep Networks for Robust Regression Prediction and Spatiotemporal Estimation”) (hereafter referred to as Li). Regarding claim 3, Kuga in view of Rabideau, Malone, and Sawruk teaches the process according to claim 1. Kuga in view of Rabideau, Malone, and Sawruk does not teach, but Li does teach wherein function g is decomposed into a core, g1, which is independent of output variable yj, and a decoder ,g2, which is conditioned on output variable yj (Li, page 2, Fig. 1 (see below), PNG media_image2.png 229 410 media_image2.png Greyscale Examiner notes that the function g is the decoder, code y is g1, the rest of the decoder is the g2, and output x’ is the output variable.) Kuga in view of Rabideau, Malone Sawruk and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, Malone, and Sawruk to use a decomposed decoder like in Li. Doing so “aims to learn an efficient data representation… typically for dimensionality reduction” (Li, page 2, 2nd column, first paragraph). Regarding claim 4, Kuga in view of Rabideau, Malone, Sawruk and Li teaches the process according to claim 3. Kuga further teaches f (xi, zi) (Kuga, page 405, 2nd column, last paragraph, “all encoder/decoder pairs are connected with the shared latent representation. Letting x ∈{xi, xs, xd } be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei,Es,Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi),Es(xs),Ed(xd)}.” Examiner notes that xi is x, zi is the latent representation, and f is the encoder that takes x as input and creates the latent representation.) E [ y j | x ] = g 1 ( g 2 ( ∑ i = 1 n f ( x i ,   z i )   ) ,   z j ) (Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph). Examiner notes that ∑ i = 1 n f ( x i ,   z i )   is the max pooled feature maps, E [ y j | x ] is the class probability, g2 is the decoder, and zj is the feature map or latent representation.) While Kuga in view of Rabideau, Malone, and Sawruk disclose the decoding of ∑ i = 1 n f ( x i ,   z i )   , Kuga in view of Rabideau, Malone and Sawruk do not explicitly disclose the decomposition part of the formula. Li, however does disclose the decomposition portion of the formula. In combination with Kuga, Rabideau, Malone, and Sawruk Li teaches g 1 ( g 2 ( ∑ i = 1 n f ( x i ,   z i )   ) ,   z j ) (Li, page 2, Fig. 1 (see below), PNG media_image2.png 229 410 media_image2.png Greyscale Examiner notes that the combination of Kuga, Rabideau, Malone, and Li disclose the entire formula with Li teaching g1 which is the code and g2 which are the decoder layers.) Kuga in view of Rabideau, Malone, Sawruk and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, Malone, and Sawruk to use a decomposed decoder like in Li. Doing so “aims to learn an efficient data representation… typically for dimensionality reduction” (Li, page 2, 2nd column, first paragraph). Regarding claim 6, Kuga in view of Rabideau, Malone, and Sawruk in further view of Li teaches the process according to claim 1. Li further teaches wherein the neural networks are residual block networks (Li, page 3, 2nd column, 4th paragraph, “Shortcut connections have been added to neural network to address vanishing/exploding gradients and degradation in accuracy in residual CNN. We use the shortcut connection of identity mapping from the encoding layer to its corresponding mirrored decoding layer to implement residual networks in a nested way from the outermost to the innermost layers.”). Kuga in view of Rabideau, Malone, Sawruk and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, Malone, and Sawruk to use residual block networks for the neural networks as in Li because “such residual connections can compensate the loss in error backpropagation through the long path of regular deep layers and collaborate with regular connections to reduce the known problem of degradation in accuracy” (Li, page 10, last paragraph 1st column – first paragraph 2nd column). Regarding claim 10, claim 10 recites substantially similar limitations as claim 3, and is therefore rejected under the same analysis. Regarding claim 11, claim 11 recites substantially similar limitations as claim 4, and is therefore rejected under the same analysis. Regarding claim 13, claim 13 recites substantially similar limitations as claim 6, and is therefore rejected under the same analysis. Claim(s) 15-17 and 19-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kuga in view of Rabideau in further view of Malone, Li and Sawruk. Regarding claim 15, Kuga teaches a single universal prediction model comprised of multiple neural networks trained across diverse tasks from multiple domains in a shared space with disjoint input and output variable sets (Kuga, page 403, abstract, “We propose a multi-modal encoder-decoder networks to harness the multi-modal nature of multi-task scene recognition. In additional to the shared latent representation among encoder-decoder pairs, our model also has shared skip connections from different encoders” where “we describe details of the cost functions defined for semantic label decoder, image decoder, and depth decoder” (Kuga, page 406, 1st column, 3rd paragraph) and where “we discuss details of the proposed network by taking the task of depth and semantic label estimation from RGB images assuming a training dataset consisting with three modalities: image, depth and semantic labels” (Kuga, page 405, 1st column last paragraph) and “to exploit the commonality among different tasks, all encoder/decoder pairs are connected with each other via the shared latent representation” (Kuga, page 404, 2nd column, last paragraph). Examiner notes that the tasks with disjoint input and output are the tasks, where the tasks are the image, depth, and semantic labels. Examiner further notes that measuring the tasks is the cost functions and the shared space is latent representation.); an encoder, f, which is conditioned on a shared spaced vector zi, for generating by a neural network encoder an encoder output for a task variable xi given its location in the shared space z (Kuga, page 405, 2nd column, last paragraph, “all encoder/decoder pairs are connected with the shared latent representation. Letting x ∈{xi, xs, xd } be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei,Es,Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi),Es(xs),Ed(xd)}.” Examiner notes that for each modalities, or task, a function E is encoding an input value x where the input value x is the value of each observed variable. Examiner further notes that the shared space location and zi is the shared latent representation and the encoder is the first neural network.); wherein a same neural network encoder is used for all diverse tasks (Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that Figure 1 shows the encoder in the left box. ) a decoder, g2, which is conditioned on shared space vector zj, for generating by a single neural network decoder a prediction yj given its location in the shared space z (Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph) and Kuga, page 405, Figure 1, PNG media_image1.png 580 835 media_image1.png Greyscale Examiner notes that the aggregated encodings are the max pooled feature maps, the prediction is the class probability, and the location is the feature map or latent representation. Examiner further notes that the feature map is created by embedding values, and so the feature map is made of the variable embeddings. Additionally, the decoder is the second neural network shown in Figure 1 in the right circle.). Kuga teaches an encoder, f, which is conditioned on … zi and a decoder, g2, which is conditioned on …zj. Kuga does not teach that the latent representation is a vector. Rabideau teaches an encoder, f, which is conditioned on shared space vector zi (Rabideau, page 10, paragraph 0054, “In the Variational AutoEncoder (VAE) described herein, the learned output of the encoder is a latent vector in a latent representation space 107.”) a decoder, g2, which is conditioned on shared space vector zj (Rabideau, page 10, paragraph 0054, “decoder 108 similarly constructed as a multilayered neural network, has an input layer 119 (in this the input layer receives a latent vector representation from the latent representation space 107) a hidden layer 120 and an output layer 121.”) Kuga and Rabideau are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date to have modified Kuga to use vectors as latent representations like in Rabideau. Doing so allows for “the loss for this objective function is found such that backpropagation can be performed in order to adjust the weights on the Variational AutoEncoder’s computational graph such that this objective function is maximized” (Rabideau, page 12, paragraph 0067). Kuga in view of Rabideau does not teach, but Malone does teach an aggregator for aggregating the encoder outputs (Malone, page 14, paragraph 0100, “FIG. 3 shows an example of a submodel 50 which learns embeddings 52 for combined heart rate and blood pressure data as a data modality 44, including when some of the raw patient data 46 is missing….The submodel 50 consists of two function, fd- and fm…..According to an embodiment, element-wise addition is used to combine the embeddings 52 of the episode snapshot 42a.” Examiner notes that the submodel is the aggregator and the embeddings are the encoder outputs.); Kuga, Rabideau and Malone are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga and Rabideau to aggregate encodings by elementwise addition like in Malone. Doing so “allows to learn embeddings for the missing data values” (Malone, page 14, paragraph 0099). Kuga in view of Rabideau and Malone do not teach, but Li does teach a core, g1, which is independent of output (Li, page 2, Fig. 1 (see below), PNG media_image2.png 229 410 media_image2.png Greyscale Examiner notes that code y is g1, and the output variable is output x).; Kuga in view Rabideau, Malone and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau and Malone to use a decomposed decoder like in Li. Doing so “aims to learn an efficient data representation… typically for dimensionality reduction” (Li, page 2, 2nd column, first paragraph). Kuga in view of Rabideau, Malone, and Li does not teach, but Sawruk does teach wherein the shared space vectors zi and zj are variable embeddings learned by gradient descent jointly with trainable parameters of the encoder and the decoder (Sawruk, page 41, paragraph 0091, The model is updated using gradient descent, an optimization algorithm that numerically tries to find the minimum of a function. The function in this case is the neural network itself. The neural network is then trained for several iterations. The inner matrix W1 now holds the embedding vectors. This is called the embedding matrix” where “This model is known as a neural machine translation (NMT). The first neural network is called the encoder and transforms each input phrase into a numerical representation. The second neural network is called the decoder and transforms the numerical representation into output text, such as keywords describing the phrase” (Sawruk, page 44, paragraph 0101) and where “After the neural network is trained, a weight matrix is extracted where each row (or column) of the weight matrix corresponds to each element in the vocabulary 558. These are called the embeddings” (Sawruk, page 36, paragraph 0076). Examiner notes that the embeddings are within weight matrix. Examiner further notes that the embeddings are the parameters of the encoder and the decoder. Since the embeddings are the parameters, the embeddings and parameters are learned jointly by gradient descent.) Kuga, Rabideau, Malone, Li and Sawruk are analogous to the claimed invention because they use an encoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga, Rabideau, Malone, and Li to use gradient descent like in Sawruk. Doing so “causes similar items to cluster together in the 5-dimensional space. Accordingly, the embedding vectors encode contextual information and the vector shown above represents a semantically meaningful encoding” (Sawruk, page 41, paragraph 0093). Regarding claim 16, Kuga in view of Rabideau, Malone, Li and Sawruk teach the single universal prediction model of claim 15. Kuga further teaches f (xi, zi) (Kuga, page 405, 2nd column, last paragraph, “all encoder/decoder pairs are connected with the shared latent representation. Letting x ∈{xi, xs, xd } be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei,Es,Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi),Es(xs),Ed(xd)}.” Examiner notes that xi is x, zi is the latent representation, and f is the encoder that takes x as input and creates the latent representation.) E [ y j | x ] = g 1 ( g 2 ( ∑ i = 1 n f ( x i ,   z i )   ) ,   z j ) (Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph). Examiner notes that ∑ i = 1 n f ( x i ,   z i )   is the max pooled feature maps, E [ y j | x ] is the class probability, g2 is the decoder, and zj is the feature map or latent representation.) Kuga does not teach that the latent representations are vectors. Kuga in view of Rabideau, however teaches vector zi and zj (Rabideau, page 10, paragraph 0054, “In the Variational AutoEncoder (VAE) described herein, the learned output of the encoder is a latent vector in a latent representation space 107. A decoder 108 similarly constructed as a multilayered neural network, has an input layer 119 (in this the input layer receives a latent vector representation from the latent representation space 107) a hidden layer 120 and an output layer 121.”) While Kuga in view of Rabideau and Malone disclose the decoding of ∑ i = 1 n f ( x i ,   z i )   , Kuga in view of Rabideau and Malone do not explicitly disclose the decomposition part of the formula. Li, however does disclose the decomposition portion of the formula. In combination with Kuga, Rabideau, and Malone, Li teaches g 1 ( g 2 ( ∑ i = 1 n f ( x i ,   z i )   ) ,   z j ) (Li, page 2, Fig. 1 (see below), PNG media_image2.png 229 410 media_image2.png Greyscale Examiner notes that the combination of Kuga, Rabideau, Malone, and Li disclose the entire formula with Li teaching g1 which is the code and g2 which are the decoder layers.) Kuga in view of Rabideau, Malone and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, and Malone to use a decomposed decoder like in Li. Doing so “aims to learn an efficient data representation… typically for dimensionality reduction” (Li, page 2, 2nd column, first paragraph). Regarding claim 17, Kuga in view of Rabideau, Malone, Li, and Sawruk teaches the single universal prediction model of claim 15. Kuga further teaches wherein vector zi and zj are variable embeddings (Kuga, page 405, 2nd column, 1st paragraph, “each convolution layer (Conv) in the encoder is followed by a batch-normalization layer (Norm) and activation function (ReLU). Two maxpooling operations are placed in the middle of seven Conv+Norm+ReLU components, which makes the dimension of the latent representation 1/16 of the input. Similarly, the decoder network consists of seven Conv+Norm+ReLU components except for the final layer, while max-pooling operations are replaced by un-pooling operations for expanding a feature map. The max-pooling operation pools a feature map by taking maximum values, and the un-pooling operation restores the pooled feature into un-pooled feature map by embedding the same values. The final output of the decoder is then rescaled to the original input size. The rescaled output is further fed into a softmax layer to produce the class probability distribution at the label decoder” where “we can obtain the outputs y ∈ {Di(r),Ds(r),Dd(r)} from any r, where D ∈ {Di,Ds,Dd} are decoder functions” (Kuga, page 406, 1st column, 1st paragraph). Examiner notes that the aggregated encodings are the max pooled feature maps, the prediction is the class probability, and the location is the feature map or latent representation. Examiner further notes that the feature map is created by embedding values, and so the feature map is made of the variable embeddings.). Kuga does not teach that the latent representations zi and zj are vectors. However Rabideau teaches vector zi and zj (Rabideau, page 10, paragraph 0054, “In the Variational AutoEncoder (VAE) described herein, the learned output of the encoder is a latent vector in a latent representation space 107. A decoder 108 similarly constructed as a multilayered neural network, has an input layer 119 (in this the input layer receives a latent vector representation from the latent representation space 107) a hidden layer 120 and an output layer 121.”) Kuga and Rabideau are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to a person having ordinary skill in the art prior to the effective filing date to have modified Kuga to use vectors as latent representations like in Rabideau. Doing so allows for “the loss for this objective function is found such that backpropagation can be performed in order to adjust the weights on the Variational AutoEncoder’s computational graph such that this objective function is maximized” (Rabideau, page 12, paragraph 0067). Regarding claim 19, Kuga in view Rabideau, Malone and Li teaches the single universal prediction of claim 15. Li further teaches wherein the neural networks are residual block networks (Li, page 3, 2nd column, 4th paragraph, “Shortcut connections have been added to neural network to address vanishing/exploding gradients and degradation in accuracy in residual CNN. We use the shortcut connection of identity mapping from the encoding layer to its corresponding mirrored decoding layer to implement residual networks in a nested way from the outermost to the innermost layers.”). Kuga in view Rabideau, Malone and Li are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view Rabideau, Malone and Li to use residual block networks for the neural networks as in Li because “such residual connections can compensate the loss in error backpropagation through the long path of regular deep layers and collaborate with regular connections to reduce the known problem of degradation in accuracy” (Li, page 10, last paragraph 1st column – first paragraph 2nd column). Regarding claim 20, Kuga in view of Rabideau, Malone, and Li teach the single universal prediction of claim 15. Kuga in view of Rabideau, Malone, and Li do not teach wherein the encoding and decoding are conditioned on the variable embeddings via Feature-wise Linear Modulation (FiLM) layers (Brocal, page 3, 1st column, 5th paragraph, “our conditioned network is a standard U-Net that computes a set of generic source separation filters that we use to separate the various instruments. It adapts itself through the control mechanism (the condition generator) with FiLM layers inserted at different depths. Our external data is a condition vector…(a one-hot-encoding) which specify the instrument to be separated….The vector…is the input to the control mechanism/condition generator.” where “a FiLM layer is inserted inside each encoding block” (Brocal, page 4, 1st column, 2nd paragraph) and where “the decoder maps the encoder, with 6 decoders blocks” (Brocal, page 4, 1st column, 1st paragraph). Examiner notes that the variable embeddings is the condition vector.). Kuga in view of Rabideau, Malone, Li and Brocal are analogous to the claimed invention because they use an encoder-decoder structure to produce an output. It would have been obvious to one having ordinary skill in the art prior to the effective filing date to have modified Kuga in view of Rabideau, Malone, and Li to use FiLM layers like in Brocal. Doing so, “can do several source separation tasks without losing performance as it does not introduce any limitation and makes use of the commonalities of the distinct instruments” (Brocal, page 6, 2nd column, 3rd paragraph). Response to Arguments On page 6, Applicant argues: The Applicant appreciates the Examiner's confirmation that the Yang and Hospedales NPL document submitted with the Applicant's April 21, 2026 response has been considered. The Examiner further notes that "a significant number of NPL documents are still required and are missing based on the IDS filed 01/12/2022." The Applicant has re-reviewed the Information Disclosure Statement filed January 12, 2022 against the electronic application file and submits herewith copies of the non-patent literature documents that were lined through on the returned PTO-1449. Consideration of these documents and return of an initialed PTO-1449 are respectfully requested. Regarding the Applicant’s argument that the copies of non-patent literature documents were submitted, Examiner respectfully disagrees. Specifically, Examiner notes that many NPL documents were provided, however the documents that are still lined through either have not been provided or are illegible. Regarding the Applicant’s 112(a) arguments on pages 7-8, Examiner has found the arguments persuasive and has withdrawn the 112(a) rejection accordingly. On page 9, Applicant argues: First, regarding the mental process characterization, the Examiner reasons that each of the four identified limitations of claim 11 "encompasses" a human mentally performing it, concluding that "[s]ince these elements encompass evaluations, under MPEP 2106.04(a), these elements are mental processes." Office Action, pp. 50-51. Respectfully, this reasoning is only possible by first stripping each limitation of its express implementation. The claims do not recite "encoding" and "decoding" in the abstract; they recite encoding "implemented by a neural network encoder, wherein a same neural network encoder is used for all diverse tasks," conditioned on learned, multi-dimensional variable embeddings within a shared space, and decoding "implemented by a single neural network decoder." A human mind cannot practically apply a trained neural network encoder to each observed variable of diverse tasks, condition that encoder on learned variable embeddings, aggregate the resulting latent encodings by elementwise addition in a latent space of size M, and decode the aggregate through a second neural network- much less learn the embeddings "by gradient descent alongside θf and θg." Specification, ~ [0038]. As Kim Memo I instructs: The mental process grouping is not without limits. Examiners are reminded not to expand this grouping in a manner that encompasses claim limitations that cannot practically be performed in the human mind .... Claim limitations that encompass AI in a way that cannot be practically performed in the human mind do not fall within this grouping. The Applicant accordingly maintains its challenge to the finding of any abstract idea in the first place. Regarding the Applicant’s argument that encoding, aggregating, and decoding are not mental processes, Examiner respectfully disagrees. Specifically, a human can mentally encode and decode data. Examiner additionally notes that a human can aggregate data mentally as well. Examiner further notes that a claim that requires a computer may still recite a mental process (MPEP 2106.04(a)(2)(III)(C)). Examiner further notes that the encoder and decoder that have been classified as additional elements are classified as such because they are computers that are used as a tool to perform the mental process of encoding, aggregating and decoding. On page 10, Applicant argues: Second, divorcing the recited functions from the implementing neural network encoders and decoders, and then dismissing those same neural networks as "generic" or as mere "apply it on a computer" instructions, remains contrary to the requirement that the claims be analyzed as a whole. Kim Memo I could not be clearer: "the additional limitations should not be evaluated in a vacuum, completely separate from the recited judicial exception." This is precisely the analytical error the Appeals Review Panel corrected in Desjardins: the panel there had "essentially equated any machine learning with an unpatentable 'algorithm' and the remaining additional elements as 'generic computer components,' without adequate explanation," and was reversed, with the express warning that "[e]xaminers and panels should not evaluate claims at such a high level of generality." Regarding the Applicant’s argument that the Examiner has not analyzed the claims as a whole, Examiner respectfully disagrees. Specifically, by analyzing each limitation in the claim as shown above in the 101 rejections, the claims have been analyzed as a whole. Additionally, Examiner followed the guidance provided by the MPEP and memos to analyze the abstract idea and additional elements. On pages 10-11, Applicant argues: Third, regarding the Examiner's new assertion that paragraphs [0012]-[0017] "provide a problem, but not the solution and thus cannot provide an improvement" (Office Action, p. 50), and that paragraph [0017] "merely states a need for a solution" (Office Action, p. 51), the Applicant respectfully disagrees with the premise. The BACKGROUND passages were cited to establish the technical problem: cross-domain multi-task learning with disjoint input and output variable sets previously required "separate decoders and encoders for each domain," which "absorbs too much functionality to share optimally" and "makes it difficult to analyze the relationships between tasks." Specification, [0016]. The solution is described in the balance of the specification and is recited in the claims themselves: a single model in which a same neural network encoder and a single neural network decoder serve all diverse tasks, conditioned on learned variable embeddings Zi and ZJ and in a shared space, such that "functionality is otherwise completely shared across the models." Specification, [0030], [0034], [0037], [0038], [0040], [0041]; Figures 1-2; Eqs. (2) and (4). The improvement is thus supplied by the very additional elements the Examiner discounts - the claimed architecture - which enables joint training of "a single architecture across seemingly unrelated tasks with disjoint input and output spaces"[0032]) and is shown in the specification to "outperform single-task[] models tuned to each task[], as well as current Deep MTL alternatives" [0033]). That is an improvement to the functioning of artificial neural network technology itself, exactly as in Desjardins, Carmody and Baker, and exactly as contemplated by MPEP §§ 2106.04(d)(l) and 2106.05(a) and Example 39 of the 2019 PEG. In view of Desjardins, Carmody and Baker, the Board has already determined that claims directed to improving deep neural networks (DNN) are patent-eligible. See also, e.g., Ex Parte Mixer, Appeal No. 2023-003543 (Nov. 6, 2024) (reversing§ 101 rejection because "combination of elements recited by the claims result in an improvement to existing artificial neural network technology"); Ex Parte Helenius et. al., Appeal No. 2024-000079 (Aug. 13, 2024); Ex Parte Nair, et. al., Appeal No. 2024-000911 (Aug. 7, 2024); Ex Parte Holtmann-Rice et al., Appeal 2024-000046 (Mar. 27, 2024); Ex Parte Adachi, Appeal No. 2023-003642 (July 9, 2024) (reversing§ 101 rejection because the claim "recites an improvement to a neural network that includes an integrated layer that combines a dropout layer and a fully connected layer"). Moreover, the specification does not merely assert an improvement - it demonstrates one, quantitatively. Learned variable embeddings outperform fixed zero and random embeddings and achieve performance on par with hand-specified oracle embeddings (Table 2: binary cross-entropy of 0.591 for learned VEs versus 0.662 and 0.660 for zero and random VEs on CIFAR; RMSE of 3.32 versus 4.29 and 4.27 on Daily Temperature). Specification, [0046]. Across the multi-task testbeds, TOM outperforms all competing methods (Table 3: 92.90% accuracy on Concentric Hyperspheres versus 64.52% for the next-best alternative). Specification, [0051]. And on the UCI-121 suite of 121 real-world classification tasks, TOM outperforms every alternative approach across every metric (Tables 4a-4b). Specification, [0054]. This is a specification-supported improvement in the functioning of artificial neural network technology itself, of exactly the kind contemplated by MPEP §§ 2106.04(d)(l) and 2106.05(a). To place these dispositive facts beyond argument, claims 1, 8 and 15 have been amended herein to expressly recite that the variable embeddings Zi and Zj and are learned by gradient descent jointly with the trainable parameters of the neural network encoder and the single neural network decoder. Support is found at least at [0030] ("the embedded locations z are not known a priori . . . but they can be learned alongside f and g by gradient descent"), [003 8] ("these z can be learned by gradient descent alongside Brand 0g"), and [0035]. A process that learns, by gradient descent, both the parameters of a neural network encoder and decoder and the multi-dimensional location embedding of every input and output variable cannot practically be performed in the human mind, consistent with Example 39 of the 2019 PEG and the express instruction of Kim Memo I quoted above. The same limitation ties the claims directly to the demonstrated improvement of Table 2, in which the learned variable embeddings are shown to outperform fixed alternatives. Regarding the Applicant’s argument that these elements provide an improvement, Examiner respectfully disagrees. Specifically, Examiner notes that the Applicant provides a bare assertion of an improvement without the detail necessary to be apparent to one of ordinary skill in the art and, thus, cannot provide an improvement (MPEP 2106.04(d)(1)). Specifically, Examiner respectfully notes that “outperforming single-tasks models tuned to each tasks, as well as current Deep MTL alternatives” does not provide details how this specific architecture outperforms other models. Examiner further notes that learning variable embeddings is a mental process and thus cannot provide an improvement. On pages 12-13, Applicant argues: In the Response to Arguments, the Examiner asserts that "claim I does not recite a single encoder. Instead, claim I recites 'a single model comprised of multiple neural networks'." Office Action, p. 52. Respectfully, this misreads the claim. Claim I (and claims 8 and 15) expressly recites, as amended in the Applicant's prior response, "wherein a same neural network encoder is used for all diverse tasks" and "wherein the decoding is implemented by a single neural network decoder." The "multiple neural networks" of which the single model is comprised are the recited encoder and the recited decoder - they are not multiple encoders. The Examiner further asserts that Kuga nonetheless teaches these limitations "through Figure I," annotating that "the left box is the encoder and the right circle is the decoder of the model in Kuga." Office Action, p. 20 et seq, 53. Respectfully, that annotation is factually incorrect, as Kuga' s own text makes explicit. Kuga' s left box depicts three separate encoder networks, one per input modality, and the right circle depicts three separate decoder networks, one per output modality: In this work, we propose a multi-modal encoder-decoder networks for multi-task scene recognition. Our model consists of encoders and decoders for each modality, and the whole network is trained in an end-to-end manner taking into account all conversion paths - both cross-modal encoder-decoder pairs and within-modal self-encoders. Kuga, p. 404, 1st column (emphasis added). Letting x ∈ { Xi, Xs, Xd} be the input modalities for each encoder; image, semantic label and depth map, and E ∈ {Ei, Es, Ed} be the corresponding encoder functions, then the outputs from each encoder r, which means latent representation, is defined by r ∈ {Ei(xi), Es(xs), Ed(xd)}. Kuga, pp. 405-406 (emphasis added). we can obtain the outputs y ∈ {Di(r), Ds(r), Dd(r)} from any r, where D ∈ {Di, Ds, Dd} are decoder functions. Kuga, p. 406, 1st column (emphasis added). Kuga thus discloses three distinct encoder functions Ei, Es and Ed - an image encoder, a semantic label encoder, and a depth encoder, so labeled in Kuga's Figure 2 (reproduced below)- and three distinct decoder functions Di, Ds and Dd. What is "shared" in Kuga is the latent representation and skip connections "from different encoders" (Kuga, Abstract) - not the encoder and not the decoder: … A same neural network encoder used for all diverse tasks, and a single neural network decoder decoding the aggregated encodings for all diverse tasks, is precisely what Kuga does not have and does not suggest. Neither Rabideau nor Malone cures this deficiency, and the Examiner does not rely on either for it. Regarding the Applicant’s argument that Kuga does not disclose a same neural network encoder used for all diverse tasks and a single neural network decoder, Examiner respectfully disagrees. Specifically, Examiner notes that under broadest reasonable interpretation, the claimed “same neural network encoder” includes multiple sub-neural networks as portrayed in Figure 1 of Kuga. Thus, the left box surrounding the encoders make up the larger encoding function, or rather the same neural network encoder. This is further supported by Kuga by using a single shared latent representation to connect the encoders and decoders, making the encoders part of a larger encoder or encoding function and the decoders part of a larger decoder or decoding function. On pages 13-14, Applicant argues: Independently, the combination fails for a second reason. The claims require encoding a value of each observed input variable xi “given its shared space location vector and decoding to predict output variable yj "at its location vector zj wherein zi and zj are variable embeddings. The location vectors are inputs that condition the encoder and decoder. In Kuga and Rabideau, by contrast, the "latent representation" upon which the Examiner relies is the output of the encoder - Kuga: "the outputs from each encoder r, which means latent representation"; Rabideau: "the learned output of the encoder is a latent vector in a latent representation space 107" ( [0054]). The Office Action maps this same latent representation simultaneously to the shared space z, to the location vectors zi and zj and to the encodings themselves. An encoder output cannot be the conditioning input "given" to that encoder. Neither Kuga, Rabideau nor Malone teaches or suggests per-variable embeddings supplied to a same encoder and single decoder as claimed. Regarding the Applicant’s argument that Kuga does not teach “given its shared space location” and “at its location”, Examiner respectfully disagrees. Specifically, Examiner notes that under broadest reasonable interpretation, the shared space location is given with the architecture of the model. As noted above, Kuga does not teach the vector portion of these limitations. Instead, Rabideau teaches the shared space location vector. Additionally, Examiner notes that the claim does not explicitly cite encodings. Instead, the claim cites an action of encoding input variables. Thus, the alleged encodings are not the same as the location vectors. On page 14, Applicant argues: With respect to claims 7 and 14, the Examiner reasons that "since the laten[t] representation is created from pooling a 2D input image, the latent representation, or shared space is also 2 dimensional." Office Action, p. 23, 29. Kuga expressly defines otherwise: "r ∈ Rcxnxw where C, H, and W are the number of channels, height and width, respectively." Kuga, p. 406, 1st column. Kuga's latent representation is a three-dimensional tensor, not a 2- Dimensional shared space z in which input and output variables are located by variable embeddings. Regarding the Applicant’s argument that Kuga does not teach a 2D shared space, Examiner respectfully disagrees. Specifically, Kuga describes the input image size as a 2D size. Kuga also describes the latent representation as 1/16 of that size, which results in a shared latent representation that is also 2 dimensional (see 103 rejection above). Thus, Kuga describes a shared location that is two dimensional. On page 14, Applicant argues: As to Malone, the cited element-wise addition combines the outputs of two functions fd (observed data) and fm (missing data) within a single data modality of a patient episode snapshot in order to compensate for missing values (Malone, [0099]-[0100], [0134]). It is not the claimed aggregation, by elementwise addition, of the encodings of each observed input variable produced by a same encoder conditioned on variable embeddings, and the proffered motivation - "allows to learn embeddings for the missing data values" - is inapposite to claims that aggregate encodings of observed variables. Regarding the Applicant’s argument that Malone does not teach the element-wise addition, Examiner respectfully disagrees. Specifically, Kuga, Rabideau, Malone, and Sawruk in combination teach claim 1. Further Malone specifically teaches aggregating encodings by elementwise addition (Malone, page 18, paragraph 0134, “For each input, the output vectors of the encoding functions are added element-wise and the resulting vector used in the margin-based contrastive loss function”. Examiner notes that the elementwise addition is summing the encodings.). On page 14, Applicant argues: Finally, the amendment presented herein - reciting that the variable embeddings Zi and Zj are learned by gradient descent jointly with the trainable parameters of the neural network encoder and the single neural network decoder - further distinguishes the cited art. None of Kuga, Rabideau or Malone teaches or suggests learning per-variable location embeddings, supplied as conditioning inputs to a same encoder and a single decoder, jointly with the parameters of that encoder and decoder. Applicant’s arguments with respect to claim(s) 1 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. On page 15, Applicant argues: Meseguer-Brocal does not remedy the deficiencies of the art as discussed above. The Applicant additionally notes that the condition vector of Meseguer-Brocal is "a one-hot encoding" that "specif[ies] the instrument to be separated" - at most a task selector -Not variable embeddings Zi and encoding the location of each individual input and output variable zi in a shared space. Accordingly, claims 2 and 9 are patentable over the combination of Kuga, Rabideau, Malone and Meseguer-Brocal at least by virtue of their dependence on claims 1 and 8. Regarding the Applicant’s argument that Brocal doesn’t disclose variable embeddings, Examiner respectfully disagrees. Specifically, Brocal teaches wherein the encoding and decoding are conditioned on the variable embeddings via Feature-wise Linear Modulation (FiLM) layers (Brocal, page 3, 1st column, 5th paragraph, “our conditioned network is a standard U-Net that computes a set of generic source separation filters that we use to separate the various instruments. It adapts itself through the control mechanism (the condition generator) with FiLM layers inserted at different depths. Our external data is a condition vector…(a one-hot-encoding) which specify the instrument to be separated….The vector…is the input to the control mechanism/condition generator.” where “a FiLM layer is inserted inside each encoding block” (Brocal, page 4, 1st column, 2nd paragraph) and where “the decoder maps the encoder, with 6 decoders blocks” (Brocal, page 4, 1st column, 1st paragraph). Examiner notes that the variable embeddings is the condition vector.). On page 15, Applicant argues: Li et al. does not remedy the deficiencies of the art as discussed above. The Applicant additionally notes that the Examiner maps the claimed core g1 to the "code y" of Li's Figure 1. The "code" in Li is a latent data representation - the output of Li's encoder- not a function at all; it cannot be the claimed core g1, which is a function that transforms the aggregated encodings and is independent of output variable Yi· Nor is the decoder of Li's symmetric autoencoder "conditioned on output variable y/' - it merely reconstructs the input. Accordingly, claims 3-4, 6, 10-11 and 13 are patentable over the combination of Kuga, Rabideau, Malone and Li et al. at least by virtue of their dependence on claims 1 and 8. Regarding the Applicant’s argument that Li does not disclose core g1, Examiner respectfully disagrees. Specifically, Examiner notes that claims 3, 4, 6, 10, 11, and 13 do not recite that g1 must be a function. Instead they recite the function g is decomposed into g1 and g2. As such, the code y of Li maps to the core g1. On pages 15-16, Applicant argues: Claim 15 recites "wherein a same neural network encoder is used for all diverse tasks" and "a decoder, g2, ... for generating by a single neural network decoder a prediction Yi given its location in the shared space z." For the reasons set forth above with respect to claims 1, 7, 8 and 14, which apply with equal force here, Kuga discloses separate per-modality encoders (Ei, Es, Ed) and separate per-modality decoders (Di, Ds, Dd), sharing only the latent representation between them, and neither Kuga nor Rabideau teaches variable embeddings zi, zj as conditioning inputs to the encoder and decoder rather than as encoder outputs. Neither Rabideau, Malone nor Li remedies these deficiencies. The amendment to claim 15 reciting that the shared space vectors are variable embeddings learned by gradient descent jointly with the trainable parameters of the encoder and the decoder further distinguishes the combination, for the reasons given above with respect to claim 1. Regarding Applicant’s argument that the prior art of record does not teach claim 15 for similar reasons as claim 1, Examiner respectfully disagrees. Specifically, claim 1 recites similar limitations to claim 15 and as such, claim 15 has been rejected in a similar fashion to claim 1 with the exception of Li which discloses the core g1. On page 16, Applicant argues: To the extent not addressed, the undersigned submits that it is not necessary to address the rejections of the dependent claims with respect to the elements therein in this response since the dependent claims are believed to be allowable in view of the allowability of the independent claims. The undersigned does not admit to the propriety of the Office's rejection. 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. Thathachar et al. (US 2020/0090035 A1) also discusses an encoder-decoder structure using multi-task learning. Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to KAITLYN R LAU whose telephone number is (571)272-1429. The examiner can normally be reached Monday - Thursday: 8:00 am - 6:00 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
Read full office action

Prosecution Timeline

Show 2 earlier events
Oct 28, 2025
Response Filed
Dec 29, 2025
Final Rejection mailed — §101, §103
Mar 23, 2026
Response after Non-Final Action
Apr 21, 2026
Request for Continued Examination
Apr 27, 2026
Response after Non-Final Action
Jun 03, 2026
Non-Final Rejection mailed — §101, §103
Aug 31, 2026
Response Filed
Oct 19, 2026
Final Rejection mailed — §101, §103 (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12782601
USING MACHINE LEARNING FOR MODELING CLIMATE DATA
4y 11m to grant Granted Nov 03, 2026
Patent 12688298
FEATURE SELECTION FOR CYBERSECURITY THREAT DISPOSITION
4y 7m to grant Granted Jul 21, 2026
Patent 12602431
METHODS FOR PERFORMING INPUT-OUTPUT OPERATIONS IN A STORAGE SYSTEM USING ARTIFICIAL INTELLIGENCE AND DEVICES THEREOF
3y 10m to grant Granted Apr 14, 2026
Patent 12572828
METHOD FOR INDUSTRY TEXT INCREMENT AND ELECTRONIC DEVICE
4y 5m to grant Granted Mar 10, 2026
Study what changed to get past this examiner. Based on 4 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

5-6
Expected OA Rounds
60%
Grant Probability
99%
With Interview (+66.7%)
3y 11m (~0m remaining)
Median Time to Grant
High
PTA Risk
Based on 10 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month