DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . This action is in response to an application filed on December 4th, 2023. Claims 1-15 are pending in the current application. The information disclosure agreement has been considered.
Claim Objections
Claim 14 objected to because of the following informalities: References the computer system of claim 15, which was not presented yet, nor does claim 15 present a computer system. Likely meant to be claim 13. Appropriate correction is required.
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.
Claim(s) 1-15 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1, Under Step 1 of the Subject Matter Eligibility Test of Products and Processes, the claim is directed towards a process, which is one of the four statutory categories.
Next, under a Step 2A Prong 1 Analysis, the claim recites the following which are interpreted to be, under the broadest reasonable interpretation, abstracts ideas.
to generate an initial computational graph having trainable variables (mental process)
Generating… a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables (mental process)
generating a new optimizer… based on a weight matrix that fits to the trainable variables of the new computational graph (mental process)
Therefore, we have to examine the claim under Step 2A prong 2, which considers the additional elements within the claim. The claim’s additional elements are:
training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables;
using an initial optimizer on a processor
based on receiving a new data set that extends the data set
on the processor
migrating weights of the trainable variables from the initial optimizer to the new optimizer
training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set.
The limitations, “using an initial optimizer on a processor”, and “on the processor”, are interpreted to be mere instructions to apply a judicial exception, as use the optimizer and processor as tools to perform the abstract idea. (See MPEP 2106.05(f)) The limitations, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are considered to be insignificant extra-solution activity. (See MPEP 2106.05(g)) Therefore, these additional elements do not integrate the abstract idea into a practical application. The claim is directed to an abstract idea.
Under a Step 2B analysis, the claim’s additional elements do not amount to significantly
more than the judicial exception as explained above in Step 2A prong 2. Additionally, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are limitations considered to be well-understood, routine, and conventional, as “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer” is interpreted to be sending or receiving data over a network, (See MPEP 2106.05(d)(ii)) and “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” is WURC as evidenced by Jason Youn and Ilias Tagkopoulos. (“we introduce a new entity/relation embedding layer that learns to differentiate distinctive entity and relation types, therefore allowing the model to learn the structure of the knowledge graph. In this work, we show that further pre training the language models with this additional embedding layer using the triples extracted from the knowledge graph, followed by the standard fine-tuning phase sets a new state-of-the-art performance for the link prediction task on the benchmark datasets.”, pg. 1, Abstract) Therefore, the claim is ineligible.
Regarding claim 13, Under Step 1 of the Subject Matter Eligibility Test of Products and Processes, the claim is directed towards a machine, which is one of the four statutory categories.
Next, under a Step 2A Prong 1 Analysis, the claim recites the following which are interpreted to be, under the broadest reasonable interpretation, abstracts ideas.
to generate an initial computational graph having trainable variables (mental process)
Generating… a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables (mental process)
generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph (mental process)
Therefore, we have to examine the claim under Step 2A prong 2, which considers the additional elements within the claim. The claim’s additional elements are:
training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables;
using an initial optimizer on a processor
based on receiving a new data set that extends the data set
on the processor
migrating weights of the trainable variables from the initial optimizer to the new optimizer
training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set.
The limitations, “using an initial optimizer on a processor”, and “on the processor”, are interpreted to be mere instructions to apply a judicial exception, as use the optimizer and processor as tools to perform the abstract idea. (See MPEP 2106.05(f)) The limitations, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are considered to be insignificant extra-solution activity. (See MPEP 2106.05(g)) Therefore, these additional elements do not integrate the abstract idea into a practical application. The claim is directed to an abstract idea.
Under a Step 2B analysis, the claim’s additional elements do not amount to significantly
more than the judicial exception as explained above in Step 2A prong 2. Additionally, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are limitations considered to be well-understood, routine, and conventional, as “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer” is interpreted to be sending or receiving data over a network, (See MPEP 2106.05(d)(ii)) and “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” is WURC as evidenced by Jason Youn and Ilias Tagkopoulos. (“we introduce a new entity/relation embedding layer that learns to differentiate distinctive entity and relation types, therefore allowing the model to learn the structure of the knowledge graph. In this work, we show that further pre training the language models with this additional embedding layer using the triples extracted from the knowledge graph, followed by the standard fine-tuning phase sets a new state-of-the-art performance for the link prediction task on the benchmark datasets.”, pg. 1, Abstract) Therefore, the claim is ineligible.
Regarding claim 15, Under Step 1 of the Subject Matter Eligibility Test of Products and Processes, the claim is directed towards a manufacture, which is one of the four statutory categories.
Next, under a Step 2A Prong 1 Analysis, the claim recites the following which are interpreted to be, under the broadest reasonable interpretation, abstracts ideas.
to generate an initial computational graph having trainable variables (mental process)
Generating… a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables (mental process)
generating a new optimizer… based on a weight matrix that fits to the trainable variables of the new computational graph (mental process)
Therefore, we have to examine the claim under Step 2A prong 2, which considers the additional elements within the claim. The claim’s additional elements are:
training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables;
using an initial optimizer on a processor
based on receiving a new data set that extends the data set
on the processor
migrating weights of the trainable variables from the initial optimizer to the new optimizer
training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set.
The limitations, “using an initial optimizer on a processor”, and “on the processor”, are interpreted to be mere instructions to apply a judicial exception, as use the optimizer and processor as tools to perform the abstract idea. (See MPEP 2106.05(f)) The limitations, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are considered to be insignificant extra-solution activity. (See MPEP 2106.05(g)) Therefore, these additional elements do not integrate the abstract idea into a practical application. The claim is directed to an abstract idea.
Under a Step 2B analysis, the claim’s additional elements do not amount to significantly
more than the judicial exception as explained above in Step 2A prong 2. Additionally, “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables”, “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer”, and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” are limitations considered to be well-understood, routine, and conventional, as “based on receiving a new data set that extends the data set”, “migrating weights of the trainable variables from the initial optimizer to the new optimizer” is interpreted to be sending or receiving data over a network, (See MPEP 2106.05(d)(ii)) and “training… an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables” , and “training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set” is WURC as evidenced by Jason Youn and Ilias Tagkopoulos. (“we introduce a new entity/relation embedding layer that learns to differentiate distinctive entity and relation types, therefore allowing the model to learn the structure of the knowledge graph. In this work, we show that further pre training the language models with this additional embedding layer using the triples extracted from the knowledge graph, followed by the standard fine-tuning phase sets a new state-of-the-art performance for the link prediction task on the benchmark datasets.”, pg. 1, Abstract) Therefore, the claim is ineligible.
Regarding claims 2 and 14, the claims recite “predicting relational links for the new embeddings such that one or more new entities are connected to each other and/or to one or more existing entities using a link prediction embedding representation function, wherein the embeddings and the new embeddings are iteratively refined based on repeating steps b) and c) for further new data sets.” The limitation, as drafted, is considered to be, under the broadest reasonable interpretation, a “mental process”, which is a grouping of abstract idea. Therefore, the claims are rejected on the same basis as claims 1 and 13.
Regarding claim 3, the claim recites “removing the initial computational graph and the initial optimizer from the processor.” The limitation, as drafted, is considered to be insignificant extra-solution activity, as well as well-understood, routine, and conventional, as shown by Florian Weijers (“selected deletion processes in typical workflows of common electronic data processing systems such as PCs and laptops are to be analyzed and compared.”, pg. 2 of Presentation and evaluation of common methods of deleting user data in common computer file systems.) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 4, the claim recites “applying a lossless autoencoder to the weights of the trainable variables to match the new embedding dimensions, wherein the lossless autoencoder is an unsupervised autoencoder, and wherein applying the lossless autoencoder is based on an embedding dimension of the trainable variables from the initial computational graph being modified.” The limitation, as drafted, merely indicates the field of use and particular technological environment, and “generally links” applying an unsupervised lossless autoencoder to the abstract idea. (See MPEP 2106.05(h)) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 5, the claim recites “a mapping function is defined to migrate the weights of the trainable variables from the initial optimizer to the new optimizer, wherein the mapping function is obtained using a vector-to-vector machine learning model.” The limitation, as drafted, is interpreted to be mere instructions to apply a judicial exception, as it instructs to implement a mapping function to migrate weights, and use a model to obtain the mapping function. (See MPEP 2106.05(f)) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 6, the claim recites “the processor is a graphics processing unit (GPU), and wherein all operations of the initial computational graph and the new computational graph are maintained on the GPU.” The limitation, as drafted, is interpreted to be mere instructions to apply a judicial exception, as it instructs to use a GPU to perform the abstract idea. (See MPEP 2106.05(f)) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 7, the claim recites “instantiating the new embedding dimensions and/or the weights of the trainable variables with at least random values, using a pooling mechanism based on a neighborhood of the new computational graph, or with zeros.” The limitation, as drafted, is considered to be, under the broadest reasonable interpretation, a “mathematical concept”, which is a grouping of abstract idea. Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 8, the claim recites “the embedding- based neural network model is used on the processor to predict potential states from the data set in parallel to training, on the processor, the embedding-based neural network model.” The limitation, as drafted, is interpreted to be mere instructions to apply a judicial exception, as it instructs to use the neural network model to perform the prediction of potential states in parallel to training the model. (See MPEP 2106.05(f)) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 9, the claim recites “identifying a subset of molecules from the new data set using the embedding-based neural network model, the new data set including amino acid sequences, wherein the subset of molecules is distinct from an antibiotic.” The limitation, as drafted, is considered to be, under the broadest reasonable interpretation, a “mental process”, which is a grouping of abstract idea. Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 10, the claim recites “the trainable variables for the initial computational graph include a predefined embedding dimension, relation embeddings, and node embeddings.” The limitation, as drafted, merely indicates the field of use and particular technological environment, and “generally links” embedding dimensions, relation embeddings, and node embeddings to the abstract idea. (See MPEP 2106.05(h)) Therefore, the claim is rejected on the same basis as claim 1.
Regarding claim 11, the claim recites “removing certain evidence types from the new data set using the embedding-based neural network model, the new data set including devices used to obtain evidence or information associated with the evidence, wherein the certain evidence types correspond to a particular device of the devices or a particular piece of the evidence or the information of the evidence or the information.” The limitation, as drafted, is considered to be insignificant extra-solution activity, as well as well-understood, routine, and conventional, as shown by Michele Bellingeri et al. (“In the last few decades, a number of studies investigated the response of real networks to link/node removal (LNR) in what is called “network attack analysis” because it simulates the consequences of an attack on the network [1–8].”, pg. 1, under Introduction, from Link and Node Removal in Real Social Networks: A Review) Therefore, the claim is rejected on the same basis as claim 1. Regarding claim 12, the claim recites “the new optimizer is a re-instantiated version of the initial optimizer, the re-instantiated version of the new optimizer instantiated with new and/or different parameters.” The limitation, as drafted, merely indicates the field of use and particular technological environment, and “generally links” a re-instantiated version of an optimizer to the abstract idea. (See MPEP 2106.05(h)) Therefore, the claim is rejected on the same basis as claim 1.
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.
Claim(s) 1-7, 10, 13, and 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rajarshi Das et al. (Herein referred to as Das) (Probabilistic Case-based Reasoning for Open-World Knowledge Graph Completion) in view of Binon Teji et al. (Herein referred to as Teji) (Graph Embedding Techniques for Predicting Missing Links in Biological Networks: An Empirical Evaluation) and in further view of ZHONGQIN BI et al. (Herein referred to as Bi) (Knowledge Transfer for Out-of-Knowledge-Base Entities: Improving Graph-Neural-Network-Based Embedding Using Convolutional Layers)
Regarding claim 1, Das teaches a computer-implemented method for a dynamic embedding-based machine learning training mechanism, the computer-implemented method comprising: a) training, using an initial optimizer on a processor, an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables; (“We report results on the RotatE model with randomly initialized embeddings for new entities (RotatE) and the model with systematic initialization of new entity embeddings (RotatE+)… All models were trained till the validation set (containing both new and old triples) performance stopped improving.”, pg. 8, left column, under “3.4 Open-World KBC results”) (The original triples are using to train an embedding-based neural network (RotatE). The implicit component responsible for the model training corresponds to an initial optimizer on a processor, as one would need a processor to perform Das’ method.) b) based on receiving a new data set that extends the data set: generating, on the processor, a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables; (“For every new entity arriving in a batch, we initialize a new entity embedding for it… Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained.”, pg. 6, right column, bottom paragraph; pg. 7, left column, second paragraph)) (Based on new data input into the GNN, (corresponding to receiving a new data set that extends the data set) a new graph is reconstructed from an original graph with new embedding dimensions based on the input network data.) and c) training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set. (“Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained… we consider the more challenging setting, where new facts and entities are arriving in a streaming fashion and we give an efficient way of updating parameters using online hierarchical clustering. This allows our method to be applicable in settings where the initial KG is small and it grows continuously.”, pg. 7, left column, second paragraph; pg. 8, right column, under “Inductive representation learning on KGs”)
However, Das does not teach generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; nor migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Teji teaches generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; (“Similarly, a new development in the field involves a non-negative MF [Matrix Factorization] technique [24] that primarily constructs the correlation between the base matrix and weight matrix through the projection of graph nodes into low-dimensional space. The model uses the weighted matrix’s column vectors as the scoring matrix to find the link’s existence.”, pg. 4, left column, first paragraph) (The model of Teji performs Matrix Factorization which corresponds to a new optimizer.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the embedding-based neural network of Das with the weight migration of Teji. One would be motivated to combine the teachings, prior to the filing date of the current application, as matrix factorization helps to measure the strength of association of graph nodes with each other, as disclosed in Teji. (“the MF techniques stretch the idea of constructing a graph-based neighborhood similarity matrix to measure the strength of association of graph nodes with each other… Further, these embeddings help evaluate the relation between nodes based on the dot-product between their embeddings.”, pg. 3, right column, third paragraph)
However, the combination still does not teach migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Bi teaches migrating weights of the trainable variables from the initial optimizer to the new optimizer. (“The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient… We develop a new method to transfer knowledge for OOKB entities. In contrast to using a vector or weight matrix to represent relation embeddings in the KBC model, we use a convolution kernel to learn expressive features from the auxiliary knowledge of OOKB entities.”, pg. 3, left column, first, second and third paragraphs) (The transfer of knowledge, specifically learnable weight corresponds to migrating weight from one model (or optimizer) to another.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by teji, with the transfer learning of Bi. One would be motivated to combine the teachings, prior to the filing date of the current application, as Bi’s method combines the benefits of both a GNN and a CNN, as described by Bi. (“we propose a parameter-efficient embedding model that combines the benefits of GNN and CNN by replacing the transition weight matrix in GNN, which represents the relations, with a multilayer convolutional network. The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient.”, pg. 3, first and second paragraph)
Regarding claim 13, Das teaches a computer system for interpretable domain adaptation for a dynamic embedding- based machine learning training mechanism, the computer system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of the following steps: (While one or more hardware processors are not explicitly disclosed in Das, one would implicitly require those components to perform the method of Das.) a) training, using an initial optimizer on a processor, an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables; (“We report results on the RotatE model with randomly initialized embeddings for new entities (RotatE) and the model with systematic initialization of new entity embeddings (RotatE+)… All models were trained till the validation set (containing both new and old triples) performance stopped improving.”, pg. 8, left column, under “3.4 Open-World KBC results”) (The original triples are using to train an embedding-based neural network (RotatE). The implicit component responsible for the model training corresponds to an initial optimizer on a processor, as one would need a processor to perform Das’ method.) b) based on receiving a new data set that extends the data set: generating, on the processor, a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables; (“For every new entity arriving in a batch, we initialize a new entity embedding for it… Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained.”, pg. 6, right column, bottom paragraph; pg. 7, left column, second paragraph)) (Based on new data input into the GNN, (corresponding to receiving a new data set that extends the data set) a new graph is reconstructed from an original graph with new embedding dimensions based on the input network data.) and c) training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set. (“Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained… we consider the more challenging setting, where new facts and entities are arriving in a streaming fashion and we give an efficient way of updating parameters using online hierarchical clustering. This allows our method to be applicable in settings where the initial KG is small and it grows continuously.”, pg. 7, left column, second paragraph; pg. 8, right column, under “Inductive representation learning on KGs”)
However, Das does not teach generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; nor migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Teji teaches generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; (“Similarly, a new development in the field involves a non-negative MF [Matrix Factorization] technique [24] that primarily constructs the correlation between the base matrix and weight matrix through the projection of graph nodes into low-dimensional space. The model uses the weighted matrix’s column vectors as the scoring matrix to find the link’s existence.”, pg. 4, left column, first paragraph) (The model of Teji performs Matrix Factorization which corresponds to a new optimizer.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the embedding-based neural network of Das with the weight migration of Teji. One would be motivated to combine the teachings, prior to the filing date of the current application, as matrix factorization helps to measure the strength of association of graph nodes with each other, as disclosed in Teji. (“the MF techniques stretch the idea of constructing a graph-based neighborhood similarity matrix to measure the strength of association of graph nodes with each other… Further, these embeddings help evaluate the relation between nodes based on the dot-product between their embeddings.”, pg. 3, right column, third paragraph)
However, the combination still does not teach migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Bi teaches migrating weights of the trainable variables from the initial optimizer to the new optimizer. (“The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient… We develop a new method to transfer knowledge for OOKB entities. In contrast to using a vector or weight matrix to represent relation embeddings in the KBC model, we use a convolution kernel to learn expressive features from the auxiliary knowledge of OOKB entities.”, pg. 3, left column, first, second and third paragraphs) (The transfer of knowledge, specifically learnable weight corresponds to migrating weight from one model (or optimizer) to another.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by teji, with the transfer learning of Bi. One would be motivated to combine the teachings, prior to the filing date of the current application, as Bi’s method combines the benefits of both a GNN and a CNN, as described by Bi. (“we propose a parameter-efficient embedding model that combines the benefits of GNN and CNN by replacing the transition weight matrix in GNN, which represents the relations, with a multilayer convolutional network. The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient.”, pg. 3, first and second paragraph)
Regarding claim 15, Das teaches tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more processors, provide for interpretable domain adaptation for a dynamic embedding-based machine learning training mechanism (While the tangible, non-transitory computer-readable medium is not explicitly disclosed in Das, one would implicitly require those components to distribute and perform the method of Das.) by execution of the following steps: a) training, using an initial optimizer on a processor, an embedding-based neural network model based on a data set to generate an initial computational graph having trainable variables; (“We report results on the RotatE model with randomly initialized embeddings for new entities (RotatE) and the model with systematic initialization of new entity embeddings (RotatE+)… All models were trained till the validation set (containing both new and old triples) performance stopped improving.”, pg. 8, left column, under “3.4 Open-World KBC results”) (The original triples are using to train an embedding-based neural network (RotatE). The implicit component responsible for the model training corresponds to an initial optimizer on a processor, as one would need a processor to perform Das’ method.) b) based on receiving a new data set that extends the data set: generating, on the processor, a new computational graph of the embedding- based neural network model instantiated with new embedding dimensions migrated from the initial computational graph, the new computation graph having the trainable variables; (“For every new entity arriving in a batch, we initialize a new entity embedding for it… Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained.”, pg. 6, right column, bottom paragraph; pg. 7, left column, second paragraph)) (Based on new data input into the GNN, (corresponding to receiving a new data set that extends the data set) a new graph is reconstructed from an original graph with new embedding dimensions based on the input network data.) and c) training the embedding-based neural network model with the new data set by updating embeddings of the embedding-based neural network model and learning new embeddings of the new data set. (“Next, the model is further trained on the new batch of triples so that the new entity embeddings get trained… we consider the more challenging setting, where new facts and entities are arriving in a streaming fashion and we give an efficient way of updating parameters using online hierarchical clustering. This allows our method to be applicable in settings where the initial KG is small and it grows continuously.”, pg. 7, left column, second paragraph; pg. 8, right column, under “Inductive representation learning on KGs”)
However, Das does not teach generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; nor migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Teji teaches generating a new optimizer on the processor based on a weight matrix that fits to the trainable variables of the new computational graph; (“Similarly, a new development in the field involves a non-negative MF [Matrix Factorization] technique [24] that primarily constructs the correlation between the base matrix and weight matrix through the projection of graph nodes into low-dimensional space. The model uses the weighted matrix’s column vectors as the scoring matrix to find the link’s existence.”, pg. 4, left column, first paragraph) (The model of Teji performs Matrix Factorization which corresponds to a new optimizer.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the embedding-based neural network of Das with the weight migration of Teji. One would be motivated to combine the teachings, prior to the filing date of the current application, as matrix factorization helps to measure the strength of association of graph nodes with each other, as disclosed in Teji. (“the MF techniques stretch the idea of constructing a graph-based neighborhood similarity matrix to measure the strength of association of graph nodes with each other… Further, these embeddings help evaluate the relation between nodes based on the dot-product between their embeddings.”, pg. 3, right column, third paragraph)
However, the combination still does not teach migrating weights of the trainable variables from the initial optimizer to the new optimizer.
Bi teaches migrating weights of the trainable variables from the initial optimizer to the new optimizer. (“The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient… We develop a new method to transfer knowledge for OOKB entities. In contrast to using a vector or weight matrix to represent relation embeddings in the KBC model, we use a convolution kernel to learn expressive features from the auxiliary knowledge of OOKB entities.”, pg. 3, left column, first, second and third paragraphs) (The transfer of knowledge, specifically learnable weight corresponds to migrating weight from one model (or optimizer) to another.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by teji, with the transfer learning of Bi. One would be motivated to combine the teachings, prior to the filing date of the current application, as Bi’s method combines the benefits of both a GNN and a CNN, as described by Bi. (“we propose a parameter-efficient embedding model that combines the benefits of GNN and CNN by replacing the transition weight matrix in GNN, which represents the relations, with a multilayer convolutional network. The model has learnable weights that adapt to the amount of information from neighbors and can exploit auxiliary knowledge for OOKB entities to compute their embeddings while remaining parameter efficient.”, pg. 3, first and second paragraph)
Regarding claims 2 and 14, Das, as modified by Teji and Bi, teaches the method of claim 1 and computer system of claim 13/15 respectively, as well as predicting relational links for the new embeddings such that one or more new entities are connected to each other and/or to one or more existing entities using a link prediction embedding representation function, wherein the embeddings and the new embeddings are iteratively refined based on repeating steps b) and c) for further new data sets. (“GCN uses an iterative approach and aggregates node features of the neighboring nodes… Graph embedding is a new graph-learning paradigm for understanding underlying intricate interactions among actors (nodes) of a complex network. We used an encoder-decoder pipeline to infer links/networks from the latent vector. We assessed the performance of ten (10) graph embedding techniques to predict missing links in various homogeneous and heterogeneous biological networks. We ranked the best embedding/encoding models concerning the best decoding similarity and kernel functions.”, pg. 5, left column, second paragraph; pg. 10, right column, last paragraph (Teji)) (The graph-embedding techniques correspond to predicting relational links for the new embeddings using a link prediction embedding representation function. The process is done for a number of iterations as outlined on pg. 5, and on pg. 6’s Table I.)
Regarding claim 3, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, further comprising removing the initial computational graph and the initial optimizer from the processor. (“Two main limitations of the methods in the above-introduced categories are that they are inherently transductive (i.e., they require processing the whole graph in case of node or edge insertion or removal), and they require too many parameters (a different vector of parameter for each node)”, pg. 4, right column, second to last paragraph (Teji)) (Node or edge removal over a whole graph corresponds to a removal for the initial computational graph. If the nodes are removed, then the optimization techniques applies to them would also be removed, corresponding to a removal of an initial optimizer.)
Regarding claim 4, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, further comprising applying a lossless autoencoder to the weights of the trainable variables to match the new embedding dimensions, wherein the lossless autoencoder is an unsupervised autoencoder (“Utilize an encoding-decoding pipeline used in the Graph Auto Encoder (GAE) model for the network reconstruction.”, pg. 2, right column, second bullet (Teji))
Regarding claim 5, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, wherein a mapping function is defined to migrate the weights of the trainable variables from the initial optimizer to the new optimizer, wherein the mapping function is obtained using a vector-to-vector machine learning model. (“Given an embedded matrix Z, link prediction can be represented as a mapping function, LP : Z→G′(V,E′), such that E⊆E′”, pg. 3, left column, Definition 3 (Link Prediction) (Teji); “P is a pooling function that maps a set of vectors into a vector, i.e., P : 2Rd → Rd.”, pgs. 4-5 (Bi)) (Given the matrices as a mapping function of Teji with the pooling function, found with GNN’s of Bi, the configured combination teaches the limitation fully.)
Regarding claim 6, Teji, as modified by Bi teaches the computer-implemented method according to claim 1, wherein the processor is a graphics processing unit (GPU), and wherein all operations of the initial computational graph and the new computational graph are maintained on the GPU. (“The entire experimental assessment is conducted on the Linux-Ubuntu (x86_64) based Intel(R) Xeon(R) Silver 4216 CPU@ 2.10 GHz on Python/C++ environment with 256 GB RAM and NVIDIA Quadro RTX 5000 Graphics Card.”, pg. 7, right column, right above “C. Candidate Embedding Models and Similarity Functions” (Teji))
Regarding claim 7, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, further comprising instantiating the new embedding dimensions and/or the weights of the trainable variables with at least random values, using a pooling mechanism based on a neighborhood of the new computational graph, or with zeros. (“We explore two ways of initializing the new entity embeddings—(a) random initialization, and (b) average of element-wise rotation of entity embeddings w.r.t the relation that this new entity is connected to... Embeddings for new relations are initialized randomly... We ensure that triples in the neighborhood of the newly added entities are ten times likely to be sampled more than other triples.”, pgs. 6 and 7 (Das)) (“P is a pooling function that maps a set of vectors into a vector…“, pgs. 6-7 (Bi)) (With the pooling mechanism of Bi, configured to work with the neighborhood, randomization, and dimensions of Das’ variables the limitation is fully taught.)
Regarding claim 10, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, wherein the trainable variables for the initial computational graph include a predefined embedding dimension, relation embeddings, and node embeddings. (“An illustration of OOKB knowledge base completion. Here, the different colors of the nodes and edges represent different entities and relations, and the blocks are their embeddings. Notably, a relation in a different direction has a different embedding, such as the purple relation in the figure, which has a different embedding when it goes in or goes out blue entity. These embedding are used as convolution kernels in the model. The figure is only a sketch; more details can be found in Algorithm 1.”, pg. 5, Figure 2 (Teji))
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Das, in view of Teji, in further view of Bi, and in further view of Gamal Crichton et al. (Herein referred to as Crichton) (Neural networks for link prediction in realistic biomedical graphs: a multi-dimensional evaluation of graph embedding-based approaches)
Regarding claim 8, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, but does not explicitly teach the embedding- based neural network model is used on the processor to predict potential states from the data set in parallel to training, on the processor, the embedding-based neural network model.
Crichton teaches the embedding- based neural network model is used on the processor to predict potential states from the data set in parallel to training, on the processor, the embedding-based neural network model. (“The state of the graph before t is given to the link predictor and its aim is to predict links formed at a later time. The first setting is applicable when the current knowledge represented by the graph is incomplete and link prediction aims to complete it as well as when the temporal data for the graph is unknown or irrelevant. The second can be used to predict the future state of the graph and so can suggest feasible links to investigate.”, pg. 4, right column, under “Link prediction setting”) (With the prediction of potential states of Crichton, configured to work in parallel with training the embedding-based neural network model of Das, as modified by Teji and Bi, the limitation is fully taught.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by Teji and Bi, with the state prediction method of Crichton. One would be motivated to combine the teachings, prior to the filing date of the current application, as Crichton’s method allows for predictions to be made with unknown or irrelevant data and suggest links to investigate, as described by Crichton. (“The first setting is applicable when the current knowledge represented by the graph is incomplete and link prediction aims to complete it as well as when the temporal data for the graph is unknown or irrelevant. The second can be used to predict the future state of the graph and so can suggest feasible links to investigate.”, pg. 4, right column, under “Link prediction setting”)
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Das, in view of Teji, in further view of Bi, and in further view of Sarah G. Elnaggar et al. (Herein referred to as Elnaggar) (Embedding-Based Deep Neural Network and Convolutional Neural Network Graph Classifiers)
Regarding claim 9, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, but does not explicitly teach identifying a subset of molecules from the new data set using the embedding-based neural network model, the new data set including amino acid sequences, wherein the subset of molecules is distinct from an antibiotic.
Elnaggar teaches identifying a subset of molecules from the new data set using the embedding-based neural network model, the new data set including amino acid sequences, wherein the subset of molecules is distinct from an antibiotic. (“PROTEINS[46] is a collection of 1113 protein structures represented as graphs, where the nodes are secondary structural elements (SSEs), and the edges are neighborhoods in either a 3D space or an amino acid sequence. The graph’s nodes have three different types of labels. Protein structures can be divided into two classes: enzymes and non-enzymes.”, pg. 11, number 3.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by Teji and Bi, with the node corresponding to amino acid sequences of Elnaggar. One would be motivated to combine the teachings, prior to the filing date of the current application, as this allows for the graph classification of different bioinformatic labels, as disclosed in Elnaggar. (“…the graph classification algorithm presented in this paper is compared with graph classification methods based on traditional classifiers and methods based on deep learning in terms of classification accuracy… each dataset are transformed into graphs, where nodes stand in for atoms and edges for chemical bonds.”, pgs. 10 and 11)
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Das, in view of Teji, in further view of Bi, and in further view of Yanfei Lu et al. (Herein referred to as Lu) (Graph Embedding-Based Sensitive Link Protection in IoT Systems)
Regarding claim 11, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, but does not explicitly teach removing certain evidence types from the new data set using the embedding-based neural network model, the new data set including devices used to obtain evidence or information associated with the evidence, wherein the certain evidence types correspond to a particular device of the devices or a particular piece of the evidence or the information of the evidence or the information.
Lu teaches removing certain evidence types from the new data set using the embedding-based neural network model, the new data set including devices used to obtain evidence or information associated with the evidence, wherein the certain evidence types correspond to a particular device of the devices or a particular piece of the evidence or the information of the evidence or the information. (“the networks in IoT can also be regarded as graphs with terminal devices as nodes and communication links as edges… Algorithm 2: Generate privacy graph by deleting edges.”, pg. 1, left column, under “Introduction”; pg. 7, right column, Algorithm 2) (“Evidence types” is interpreted to mean devices, communication links and information related to devices. Lu teaches an algorithm that removes the evidence types from a graph, and so can be easily configured to work with the embedding-based GNN of Das, as modified by Teji and Bi.)
Therefore, it would have been considered obvious to one of ordinary skill in the art,
prior to the current application’s filing date, to combine the neural network of Das, as modified by Teji and Bi, with the removal of difference evidence as described in Lu. One would be motivated to combine the teachings, prior to the filing date of the current application, as this allows for the generation of a privacy graph, as well as link protection in the Internet-of-Things (IoT), as disclosed in Lu. (“The research on link protection against link prediction in IoT is of great significance for entity privacy. Through the simulation of the datasets, the feasibility of our SLPGE is preliminarily verified.”, pg. 13, left column, under “7. Conclusion”)
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Das, in view of Teji, in further view of Bi, and in further view of Anum Talpur and Mohan Gurusamy. (Herein referred to as Talpur) (Optimizing Vehicle-to-Edge Mapping with Load Balancing for Attack-Resilience in IoV)
Regarding claim 12, Das, as modified by Teji and Bi, teaches the computer-implemented method according to claim 1, but does not explicitly teach the new optimizer is a re-instantiated version of the initial optimizer, the re-instantiated version of the new optimizer instantiated with new and/or different parameters.
Talpur teaches the new optimizer is a re-instantiated version of the initial optimizer, the re-instantiated version of the new optimizer instantiated with new and/or different parameters. (“it solves the SRP model to find the optimal instance recovery placements to re-instantiate the affected service instances on the attack free server, in the given state of the environment. Once the SRP is solved, the attacked SIs are re-instantiated and the network becomes resilient to attack with a negligible loss in the network performance”, pg. 3, left column, first paragraph) (Talpur teaches to re-instantiate a model with new or different parameters. Applying Talpur’s re-instantiation technique to the optimizer of Das, with the modification of Teji and Bi, fully teaches this limitation.)
Therefore, it would have been considered obvious to one of ordinary skill in the art, prior to the current application’s filing date, to combine the neural network of Das, as modified by Teji and Bi, with the re-instantiation of Talpur. One would be motivated to combine the teachings, prior to the filing date of the current application, as this allows for the network to have increased resilience, as disclosed in Talpur. (“In the state of attack, our framework promptly maps attack-affected vehicles to the existing unaffected service instances on the attack-free servers using PSVM to ensure service availability. In the post-attack scenario, it solves the SRP model to find the optimal instance recovery placements to re-instantiate the affected service instances on the attack free server, in the given state of the environment. Once the SRP is solved, the attacked SIs are re-instantiated and the network becomes resilient to attack with a negligible loss in the network performance.”, pg. 3, left column, first paragraph)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Tyler E Iles whose telephone number is (571)272-5442. The examiner can normally be reached 9:00am - 5:00pm.
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, Kakali Chaki can be reached at (571) 272-3719. 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.
/T.E.I./Patent Examiner, Art Unit 2122
/KAKALI CHAKI/Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights