DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
This Office Action is in response to claims filed on 01/05/2026.
Claims 1-3, 5-14, and 16-20 are pending.
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 1-3, 5-14, and 16-20 are rejected under 35 U.S.C. 101 because the claimed invention recites a judicial exception, is directed to that judicial exception, an abstract idea, as it has not been integrated into practical application and the claims further do not recite significantly more than the judicial exception. Examiner has evaluated the claims under the framework provided in the 2019 Patent Eligibility Guidance published in the Federal Register 01/07/2019 and has provided such analysis below.
Step 1: Claims 1-3 and 5-11 are directed to a method and fall within the statutory category of process. Claims 12-14 and 16-19 are directed to a system and fall within the statutory category of machine. Claim 20 is directed to a non-transitory computer readable medium and falls within the statutory category of machine. Therefore, “Are the claims to a process, machine, manufacture or composition of matter?” Yes.
In order to evaluate the Step 2A inquiry “Is the claim directed to a law of nature, a natural phenomenon or an abstract idea?” we must determine, at Step 2A Prong 1, whether the claim recites a law of nature, a natural phenomenon or an abstract idea and further whether the claim recites additional elements that integrate the judicial exception into a practical application.
Step 2A Prong 1:
Claims 1, 12, and 20: The limitations of “determine(ing) a size of the input training dataset and a content type of an entry in the input training dataset;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe an input dataset and, based on these observations, can mentally determine a size of the input dataset and a content type of an entry in the input training dataset. Further, the claims recite additional abstract idea recitations of “identify(ing) respective processing and networking limits of each of the plurality of computing resources”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a plurality of computing resources and, based on these observations, can mentally identify respective processing and networking limits of each of the plurality of computing resources. This may also be done with pencil and paper. Further, the limitations of “identify(ing), from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a plurality of computing resources and, based on these observations, can mentally identify at least one computing resource to accommodate the size and the content type associated with an input training dataset. Further, a person can observe processing and network limits and a time limit of a task and, based on these observations, can mentally select at least one computing resource by determining, through mental evaluation, that the processing and network limits enable successful completion of the task within the time limit. This may also be done with pencil and paper. Further, the limitations of “identify(ing) attributes of the input training dataset;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe an input dataset and, based on these observations, can mentally identify attributes of the input training dataset. This may also be done with pencil and paper. Further, the limitations of “select(ing), from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a task, an input training dataset, and a plurality of artificial intelligence models and, based on these observations, can mentally select an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset.
Therefore, Yes, claims 1, 12, and 20 recite a judicial exception.
Step 2A Prong 2:
Claims 1, 12, and 20: The judicial exception is not integrated into a practical application. In particular, the claims recite additional element recitations of “a memory; and a hardware processor communicatively coupled with the memory and configured to:”, “A non-transitory computer readable medium storing thereon computer executable instructions for provisioning artificial intelligence resources, including instructions for:”, “train(ing), on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset;” and “and execute(ing), on the at least one computing resource, the trained artificial intelligence model to perform the task”, which are merely recitations of generically using a computer as a tool to implement the abstract idea (see MPEP § 2106.05(f)) which does not integrate a judicial exception into practical application. Further, the claims recite additional element recitations of “receive(ing) an input training dataset and an indication of a task to perform using the input training dataset, wherein the indication of the task includes a time limit for performing the task;”, which are merely recitations of data reception which is insignificant extra solution activity (see MPEP §2106.05(g)) which does not integrate a judicial exception into practical application.
Therefore, “Do the claims recite additional elements that integrate the judicial exception into a practical application? No, these additional elements do not integrate the abstract idea into a practical application and they do not impose any meaningful limits on practicing the abstract idea. The claims are directed to an abstract idea.
After having evaluated the inquires set forth in Steps 2A Prong 1 and 2, it has been concluded that claims 1, 12, and 20 not only recite a judicial exception but that the claims are directed to the judicial exception as the judicial exception has not been integrated into practical application.
Step 2B:
Claims 1, 12, and 20: The claims do not include additional elements, alone or in combination, that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements amount to no more than generic computing components and insignificant extra solution activity which do not amount to significantly more than the abstract idea. Further, the insignificant extra solution activity is well-understood, routine, and conventional in the art. “The courts have recognized the following computer functions as well‐understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity. i. Receiving or transmitting data over a network…iv. Storing and retrieving information in memory” [MPEP§ 2106.05(d)(II)].
Therefore, “Do the claims recite additional elements that amount to significantly more than the judicial exception? No, these additional elements, alone or in combination, do not amount to significantly more than the judicial exception.
Having concluded analysis within the provided framework, Claims 1, 12, and 20 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 2 and 13, the claims recite additional element recitations of “wherein the plurality of computing resources are computing devices each with different memory, storage, networking, and processing capabilities”, which are merely recitations of technological environment/field of use (see MPEP § 2106.05(h)) which does not integrate a judicial exception into practical application. Further, claims 2 and 13 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 2 and 13 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 2 and 13 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 3 and 14, the claims recite additional abstract idea recitations of “prior to receiving the input training dataset: identify(ing) respective storage limits of each of the plurality of computing resources;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a plurality of computing resources and, based on these observations, can mentally identify respective storage limits of each of the plurality of computing resources. This may also be done with pencil and paper. Further, the claims recite additional abstract idea recitations of “and identify(ing) respective processing and networking limits of each of the plurality of computing resources”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a plurality of computing resources and, based on these observations, can mentally identify respective processing and networking limits of each of the plurality of computing resources. This may also be done with pencil and paper. Further, claims 3 and 14 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 3 and 14 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 3 and 14 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 5 and 16, the claims recite additional abstract idea recitations of “wherein identifying the at least one computing resource to accommodate the size and the content type is in response to determining that a storage limit of the at least one computing resource exceeds the size” and “wherein the hardware processor is further configured to identify the at least one computing resource to accommodate the size and the content type in response to determining that a storage limit of the at least one computing resource exceeds the size”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a storage limit of at least one computing resource and a size of an input dataset and, based on these observations, can determine, through mental comparison, that the storage limit exceeds the size. This may also be done with pencil and paper. Further, claims 5 and 16 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 5 and 16 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 5 and 16 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 6 and 17, the claims recite additional abstract idea recitations of “wherein identifying the at least one computing resource to accommodate the size and the content type is in response to determining that the at least one computing resource can store the input training dataset” and “wherein the hardware processor is further configured to identify the at least one computing resource to accommodate the size and the content type in response to determining that the at least one computing resource can store the input training dataset”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe at least one computing resource and an input training dataset and, based on these observations, can determine, through mental comparison, that the at least one computing resource can store the input training dataset. This may also be done with pencil and paper. Further, the claims recite additional abstract idea recitations of “and is compatible with the content type”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe at least one computing resource and an input training dataset and, based on these observations, can determine, through mental evaluation, that the at least one computing resource is compatible with the content type of an entry in the input training dataset. Further, claims 6 and 17 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 6 and 17 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 6 and 17 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 7 and 18, the claims recite additional abstract idea recitations of “wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe an input dataset and, based on these observations, can mentally identify one or more of: (1) dimensions of a entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values. This may also be done with pencil and paper. Further, claims 7 and 18 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 7 and 18 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 7 and 18 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claims 8 and 19, the claims recite additional element recitations of “amend code associated with the artificial intelligence model to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model”, which are merely recitations of generically using a computer as a tool to implement the abstract idea (see MPEP § 2106.05(f)) which does not integrate a judicial exception into practical application. Further, claims 8 and 19 do not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claims 8 and 19 also fail both Step 2A prong 2, thus the claims are directed to the judicial exception as it has not been integrated into practical application, and fail Step 2B as not amounting to significantly more. Therefore, Claims 8 and 19 do not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claim 9, the claim recites additional abstract idea recitations of “further comprising: determining whether an error rate of the trained artificial intelligence model is greater than the target error rate;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe an error rate and a target error rate and, based on these observations, can determine, through mental comparison, whether the error rate is greater than the target error rate. This may also be done with pencil and paper. Further, the claim recites additional element recitations of “wherein the indication of the task further includes a target error rate,”, which are merely recitations of data reception which is insignificant extra solution activity (see MPEP §2106.05(g)) which does not integrate a judicial exception into practical application. Further, the claim recites additional element recitations of “and in response to determining that the error rate of the trained artificial intelligence model is greater than the target error rate, re-training the trained artificial intelligence model”, which are merely recitations of generically using a computer as a tool to implement the abstract idea (see MPEP § 2106.05(f)) which does not integrate a judicial exception into practical application. Further, claim 9 does not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claim 9 also fails both Step 2A prong 2, thus the claim is directed to the judicial exception as it has not been integrated into practical application, and fails Step 2B as not amounting to significantly more. Therefore, Claim 9 does not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claim 10, the claim recites additional abstract idea recitations of “further comprising: in response to determining that an error rate of the re-trained artificial intelligence model is greater than the target error rate, selecting, from the plurality of artificial intelligence models, a different artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset;”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a task, an input training dataset, and a plurality of artificial intelligence models and, based on these observations, can mentally select a different artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset. This may also be done with pencil and paper. Further, the claim recites additional element recitations of “and training, using the input training dataset, the different artificial intelligence model to perform the task at an error rate not greater than the target error rate”, which are merely recitations of generically using a computer as a tool to implement the abstract idea (see MPEP § 2106.05(f)) which does not integrate a judicial exception into practical application. Further, claim 10 does not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claim 10 also fails both Step 2A prong 2, thus the claim is directed to the judicial exception as it has not been integrated into practical application, and fails Step 2B as not amounting to significantly more. Therefore, Claim 10 does not recite patent eligible subject matter under 35 U.S.C. § 101.
With regard to claim 11, the claim recites additional abstract idea recitations of “further comprising: generating a report that indicates errors made by the trained artificial intelligence model”, as drafted, is a process that, but for the recitation of generic computing components, under its broadest reasonable interpretation, covers performance of the limitation in the mind. For example, a person can observe a trained artificial intelligence model and, based on these observations, can mentally generate a report that indicates errors made by the trained artificial intelligence model. This may also be done with pencil and paper. Further, claim 11 does not recite any further additional elements and for the same reasons as above with regard to integration into practical application and whether additional elements amount to significantly more, claim 11 also fails both Step 2A prong 2, thus the claim is directed to the judicial exception as it has not been integrated into practical application, and fails Step 2B as not amounting to significantly more. Therefore, Claim 11 does not recite patent eligible subject matter under 35 U.S.C. § 101.
Therefore, Claims 1-3, 5-14, and 16-20 do not recite patent eligible subject matter under U.S.C. §101.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-3, 5-6, 11-14, 16-17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11,698,863 B1) in view of Dubeyko (US 2020/0257562 A1) in view of Arora (US 2022/0206774 A1) in view of Rakshit (US 11,553,038 B1).
With regard to claim 1, Wang teaches:
A method for provisioning artificial intelligence resources, the method comprising: receiving an input training dataset “Generally, when initiating a training task, the algorithm personnel usually need to manually download these data to a node to start the training task; however, with regard to the AI resource management platform, a manual download data set is usually optimized as an automatic download data set; and when starting a training task, the AI resource management platform will automatically download the required data set for the training task. As the AI resource management platform, a variety of data sets will be provided for the algorithm personnel, and these data sets will be cached to a computing node according to the requirements of training tasks” [Wang Col. 1 Lines 31-42].
and an indication of a task to perform using the input training dataset, “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
determining a size of the input training dataset “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
identifying, from the plurality of computing resources, at least one computing resource to accommodate the size “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19]. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
and executing, on the at least one computing resource, the trained artificial intelligence model to perform the task. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “FIG. 3 shows a schematic block diagram of an embodiment of a data set and node cache-based scheduling device according to the present disclosure, as shown in FIG. 3, the device 101 includes: … a training task execution module 15 configured to obtain and delete an obsolete data set cache in the host node to be executed, and execute the training task in the host node to be executed” [Wang Col. 8 Lines 9-12, 29-32].
Wang fails to teach determining a size of the input training dataset and a content type of an entry in the input training dataset; identifying, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identifying attributes of the input training dataset.
However, Dubeyko teaches:
determining a size of the input training dataset and a content type of an entry in the input training dataset; “In other embodiments, the data type may include other or additional data types. "User data" is any data that is created or owned by a user. The user data may be of different types. For example, some user data (e.g., database user data) may be frequently updated, while other types of user data (e.g., video or image data) may be infrequently updated ("cold" data). Thus, in some embodiments, the user data may be categorized based upon the frequency of updates. Other categorizations of user data may be used in other embodiments” [Dubeyko ¶ 50]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
identifying, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; “The operating system 150 may determine the data type from the extended attributes of files in the file system (e.g., extended attributes of the applications 130). In other embodiments, the operating system 150 may be configured to determine the data type in other ways, as discussed above. Further, each data type (e.g., user data, metadata, executable code) may have certain characteristics known to the operating system 150 based on which the operating system may allocate memory to the supersets” [Dubeyko ¶ 51]. “Therefore, upon determining the data type at the operation 805, the operating system 150 determines which of the memory characteristics to prioritize over other memory properties for identifying the most suitable memory category for that data type. For example, if the operating system 150 determines the data type to be metadata, the operating system may conclude that metadata is frequently updated data, and therefore, a fast volatile memory and/or volatile memory having good memory endurance is more critical than a memory with low power consumption” [Dubeyko ¶ 77, Fig. 8]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc. Based on these superset features, the operating system may determine whether volatile or non-volatile memory is more suitable for the data/code being stored in the superset. If the selected memory category (volatile or non-volatile) has multiple memory types, the operating system may also determine which of the various memory types may be most suitable for the data/code in the superset” [Dubeyko ¶ 24].
identifying attributes of the input training dataset; “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
Dubeyko is considered to be analogous to the claimed invention because it is in the same field of computing resource provisioning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang to incorporate the teachings of Dubeyko and include: determining a size of the input training dataset and a content type of an entry in the input training dataset; identifying, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identifying attributes of the input training dataset. Doing so would allow for further efficiency in task execution through the allocation of memory. “Thus, each superset is an abstraction that is allocated memory from the physical memory based on one or more features that facilitates efficient execution of the associated application” [Dubeyko ¶ 24].
Wang in view of Dubeyko fails to teach selecting, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset.
However, Arora teaches:
selecting, from a plurality of artificial intelligence models, an artificial intelligence model “In some embodiments, the system may automatically select a machine learning model for a user based on user's data, including consideration based on the type of data (e.g., image, text, etc.) and/or based on how various machine learning models are capable of handling any particular type of data (e.g., based on one or more performance metrics)” [Arora ¶ 24].
that is configured to perform the task “For example, a facial recognition model may be associated with images. In this case, the system may identify one or more machine learning models associated with images. For each of the one or more machine learning models associated with the data type, the system may generate a respective metric value indicating a performance of the machine learning model over the input data; and select a machine learning model from among the one or more machine learning models by comparing metric values generated for each of the one or more machine learning models” [Arora ¶ 32].
and is compatible with the attributes of the input training dataset; “As discussed above, in some embodiments the system may identify one or more machine learning models (e.g., based on matching the data type associated with the machine learning models to user's data), and automatically select a suitable machine learning model” [Arora ¶ 31].
training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset; “For example, to deploy a facial recognition system, a user may select several components in the workflow, including preparing training data containing facial recognition training images, selecting a machine learning model for facial recognition, training the machine learning model using the training data, evaluating the trained machine learning model, configuring business logics such as notification upon detection of certain subjects, and deploying the machine learning model” [Arora ¶ 53].
Arora is considered to be analogous to the claimed invention because it is in the same field of machine learning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko to incorporate the teachings of Arora and include: selecting, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset. Doing so would allow for the automatic selection of a suitable machine learning algorithm for user needs. “These techniques may be executed by a system that provides a graphical user interface which allows a user to visually define a workflow for a machine learning application, without requiring the user to be an expert in machine learning or programming. The system may automatically represent the workflow as a specification that may be used to build and deploy a machine learning application” [Arora ¶ 24].
Wang in view of Dubeyko in view of Arora fails to teach wherein the indication of the task includes a time limit for performing the task; identifying respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit.
However, Rakshit teaches:
wherein the indication of the task includes a time limit for performing the task; “User or processing activities on an edge device sometimes have a performance requirement and a defined priority within a defined service level agreement (SLA) within which the activity must be completed” [Rakshit Col. 3 Lines 15-18]. “In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 9-11].
identifying respective processing and networking limits of each of a plurality of computing resources; “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit; “In this manner, aspects of the invention utilize AI to select an optimal subset of edge computing devices and communication protocols to complete computational tasks within a defined performance metric (e.g., a time limit defined in an SLA). In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 3-11]. “While performing any computational task in edge computing, data transportation time and data processing time are factors that affect the ability to meet the SLA or performance metrics for the task. For an edge computing computational task that has a bound completion time, e.g., as defined in an SLA, there is a need for a system to identify which edge computing devices to use for data processing for completing the task and how data is transported to such devices optimally” [Rakshit Col. 3 Lines 56-64].
Rakshit is considered to be analogous to the claimed invention because it is in the same field of computing resource provisioning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora to incorporate the teachings of Rakshit and include: wherein the indication of the task includes a time limit for performing the task; identifying respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit. Doing so would allow for optimization of bandwidth usage through the allocation of computing resources to perform tasks. “With continued reference to FIG. 4, at step 404 the AI-enabled system optimizes the processing power and bandwidth usage for data transportation when performing the assigned computational task” [Rakshit Col. 12 Lines 15-18].
With regard to claim 2, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang further teaches:
wherein the plurality of computing resources are computing devices each with different memory, “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19].
storage, “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51].
and processing capabilities. “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19].
Wang in view of Dubeyko in view of Arora fails to teach wherein the plurality of computing resources are computing devices each with different … networking … capabilities.
However, Rakshit teaches wherein the plurality of computing resources are computing devices each with different … networking … capabilities. “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
With regard to claim 3, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang further teaches further comprising prior to receiving the input training dataset: identifying respective storage limits of each of the plurality of computing resources; “In the embodiment shown in FIG. 1, the method includes at least the following steps: S100, obtaining storage resource information of each host node;” [Wang Col. 4 Lines 34-37]. “The present disclosure has at least the following advantageous technical effects: the present disclosure is a scheduling strategy for selecting a node based on the sizes of a node storage and a data set required by training task in a cluster environment” [Wang Col. 3 Lines 44-48].
Wang in view of Dubeyko in view of Arora fails to teach and identifying respective processing and networking limits of each of the plurality of computing resources.
However, Rakshit teaches and identifying respective processing and networking limits of each of the plurality of computing resources. “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
With regard to claim 5, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 3, as referenced above. Wang further teaches wherein identifying the at least one computing resource to accommodate the size and the content type is in response to determining that a storage limit of the at least one computing resource exceeds the size. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
With regard to claim 6, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang further teaches:
wherein identifying the at least one computing resource to accommodate the size and the content type is in response to determining that the at least one computing resource can store the input training dataset “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
Wang fails to teach determining that the at least one computing resource … is compatible with the content type.
However, Dubeyko teaches and is compatible with the content type. “Further, each data type (e.g., user data, metadata, executable code) may have certain characteristics known to the operating system 150 based on which the operating system may allocate memory to the supersets” [Dubeyko ¶ 51]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc. Based on these superset features, the operating system may determine whether volatile or non-volatile memory is more suitable for the data/code being stored in the superset. If the selected memory category (volatile or non-volatile) has multiple memory types, the operating system may also determine which of the various memory types may be most suitable for the data/code in the superset” [Dubeyko ¶ 24].
With regard to claim 11, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang in view of Dubeyko fails to teach further comprising: generating a report that indicates errors made by the trained artificial intelligence model.
However, Arora teaches further comprising: generating a report that indicates errors made by the trained artificial intelligence model. “Area 1204 may be configured to display a comparison of performance metric values for multiple machine learning models to be selected. As shown in areas 1204, ROC curves for four machine learning models for facial recognition are displayed and compared. User interface 1200 may also include area 1206 configured to display the accuracy of the machine learning models being evaluated” [Arora ¶ 101].
With regard to claim 12, Wang teaches:
A system for provisioning artificial intelligence resources, comprising: a memory; “The storage medium of the program may be a magnetic disk, an optical disk, a ROM, read-Only Memory or RAM, random Access Memory, etc.” [Wang Col. 8 Lines 65-67].
and a hardware processor communicatively coupled with the memory and configured to: “It should be noted that a person skilled in the art would understand that the implementation of all or part of the flows in the methods of the above-mentioned embodiments may be performed by a computer program instructing relevant hardware, and a program of a data set and node cache-based scheduling method may be stored in a computer-readable storage medium, and when executed, the program may include the flows of the embodiments of the methods as described above” [Wang Col. 8 Lines 57-65].
receive an input training dataset “Generally, when initiating a training task, the algorithm personnel usually need to manually download these data to a node to start the training task; however, with regard to the AI resource management platform, a manual download data set is usually optimized as an automatic download data set; and when starting a training task, the AI resource management platform will automatically download the required data set for the training task. As the AI resource management platform, a variety of data sets will be provided for the algorithm personnel, and these data sets will be cached to a computing node according to the requirements of training tasks” [Wang Col. 1 Lines 31-42].
and an indication of a task to perform using the input training dataset, “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
determine a size of the input training dataset “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
identify, from the plurality of computing resources, at least one computing resource to accommodate the size “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19]. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
and execute, on the at least one computing resource, the trained artificial intelligence model to perform the task. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “FIG. 3 shows a schematic block diagram of an embodiment of a data set and node cache-based scheduling device according to the present disclosure, as shown in FIG. 3, the device 101 includes: … a training task execution module 15 configured to obtain and delete an obsolete data set cache in the host node to be executed, and execute the training task in the host node to be executed” [Wang Col. 8 Lines 9-12, 29-32].
Wang fails to teach determine a size of the input training dataset and a content type of an entry in the input training dataset; identify, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identify attributes of the input training dataset.
However, Dubeyko teaches:
determine a size of the input training dataset and a content type of an entry in the input training dataset; “In other embodiments, the data type may include other or additional data types. "User data" is any data that is created or owned by a user. The user data may be of different types. For example, some user data (e.g., database user data) may be frequently updated, while other types of user data (e.g., video or image data) may be infrequently updated ("cold" data). Thus, in some embodiments, the user data may be categorized based upon the frequency of updates. Other categorizations of user data may be used in other embodiments” [Dubeyko ¶ 50]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
identify, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; “The operating system 150 may determine the data type from the extended attributes of files in the file system (e.g., extended attributes of the applications 130). In other embodiments, the operating system 150 may be configured to determine the data type in other ways, as discussed above. Further, each data type (e.g., user data, metadata, executable code) may have certain characteristics known to the operating system 150 based on which the operating system may allocate memory to the supersets” [Dubeyko ¶ 51]. “Therefore, upon determining the data type at the operation 805, the operating system 150 determines which of the memory characteristics to prioritize over other memory properties for identifying the most suitable memory category for that data type. For example, if the operating system 150 determines the data type to be metadata, the operating system may conclude that metadata is frequently updated data, and therefore, a fast volatile memory and/or volatile memory having good memory endurance is more critical than a memory with low power consumption” [Dubeyko ¶ 77, Fig. 8]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc. Based on these superset features, the operating system may determine whether volatile or non-volatile memory is more suitable for the data/code being stored in the superset. If the selected memory category (volatile or non-volatile) has multiple memory types, the operating system may also determine which of the various memory types may be most suitable for the data/code in the superset” [Dubeyko ¶ 24].
identify attributes of the input training dataset; “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
It would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang to incorporate the teachings of Dubeyko and include: determine a size of the input training dataset and a content type of an entry in the input training dataset; identify, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identify attributes of the input training dataset. Doing so would allow for further efficiency in task execution through the allocation of memory. “Thus, each superset is an abstraction that is allocated memory from the physical memory based on one or more features that facilitates efficient execution of the associated application” [Dubeyko ¶ 24].
Wang in view of Dubeyko fails to teach select, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; train, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset.
However, Arora teaches:
select, from a plurality of artificial intelligence models, an artificial intelligence model “In some embodiments, the system may automatically select a machine learning model for a user based on user's data, including consideration based on the type of data (e.g., image, text, etc.) and/or based on how various machine learning models are capable of handling any particular type of data (e.g., based on one or more performance metrics)” [Arora ¶ 24].
that is configured to perform the task “For example, a facial recognition model may be associated with images. In this case, the system may identify one or more machine learning models associated with images. For each of the one or more machine learning models associated with the data type, the system may generate a respective metric value indicating a performance of the machine learning model over the input data; and select a machine learning model from among the one or more machine learning models by comparing metric values generated for each of the one or more machine learning models” [Arora ¶ 32].
and is compatible with the attributes of the input training dataset; “As discussed above, in some embodiments the system may identify one or more machine learning models (e.g., based on matching the data type associated with the machine learning models to user's data), and automatically select a suitable machine learning model” [Arora ¶ 31].
train, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset; “For example, to deploy a facial recognition system, a user may select several components in the workflow, including preparing training data containing facial recognition training images, selecting a machine learning model for facial recognition, training the machine learning model using the training data, evaluating the trained machine learning model, configuring business logics such as notification upon detection of certain subjects, and deploying the machine learning model” [Arora ¶ 53].
It would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko to incorporate the teachings of Arora and include: select, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; train, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset. Doing so would allow for the automatic selection of a suitable machine learning algorithm for user needs. “These techniques may be executed by a system that provides a graphical user interface which allows a user to visually define a workflow for a machine learning application, without requiring the user to be an expert in machine learning or programming. The system may automatically represent the workflow as a specification that may be used to build and deploy a machine learning application” [Arora ¶ 24].
Wang in view of Dubeyko in view of Arora fails to teach wherein the indication of the task includes a time limit for performing the task; identify respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit.
However, Rakshit teaches:
wherein the indication of the task includes a time limit for performing the task; “User or processing activities on an edge device sometimes have a performance requirement and a defined priority within a defined service level agreement (SLA) within which the activity must be completed” [Rakshit Col. 3 Lines 15-18]. “In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 9-11].
identify respective processing and networking limits of each of a plurality of computing resources; “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit; “In this manner, aspects of the invention utilize AI to select an optimal subset of edge computing devices and communication protocols to complete computational tasks within a defined performance metric (e.g., a time limit defined in an SLA). In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 3-11]. “While performing any computational task in edge computing, data transportation time and data processing time are factors that affect the ability to meet the SLA or performance metrics for the task. For an edge computing computational task that has a bound completion time, e.g., as defined in an SLA, there is a need for a system to identify which edge computing devices to use for data processing for completing the task and how data is transported to such devices optimally” [Rakshit Col. 3 Lines 56-64].
Rakshit is considered to be analogous to the claimed invention because it is in the same field of computing resource provisioning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora to incorporate the teachings of Rakshit and include: wherein the indication of the task includes a time limit for performing the task; identify respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit. Doing so would allow for optimization of bandwidth usage through the allocation of computing resources to perform tasks. “With continued reference to FIG. 4, at step 404 the AI-enabled system optimizes the processing power and bandwidth usage for data transportation when performing the assigned computational task” [Rakshit Col. 12 Lines 15-18].
With regard to claim 13, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 12, as referenced above. Wang further teaches:
wherein the plurality of computing resources are computing devices each with different memory, “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19].
storage, “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51].
and processing capabilities. “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19].
Wang in view of Dubeyko in view of Arora fails to teach wherein the plurality of computing resources are computing devices each with different … networking … capabilities.
However, Rakshit teaches wherein the plurality of computing resources are computing devices each with different … networking … capabilities. “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
With regard to claim 14, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 12, as referenced above. Wang further teaches wherein the hardware processor is further configured to prior to receiving the input training dataset: identify respective storage limits of each of the plurality of computing resources; “In the embodiment shown in FIG. 1, the method includes at least the following steps: S100, obtaining storage resource information of each host node;” [Wang Col. 4 Lines 34-37]. “The present disclosure has at least the following advantageous technical effects: the present disclosure is a scheduling strategy for selecting a node based on the sizes of a node storage and a data set required by training task in a cluster environment” [Wang Col. 3 Lines 44-48].
Wang in view of Dubeyko in view of Arora fails to teach and identify respective processing and networking limits of each of the plurality of computing resources.
However, Rakshit teaches and identify respective processing and networking limits of each of the plurality of computing resources. “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
With regard to claim 16, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 14, as referenced above. Wang further teaches wherein the hardware processor is further configured to identify the at least one computing resource to accommodate the size and the content type in response to determining that a storage limit of the at least one computing resource exceeds the size. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
With regard to claim 17, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 12, as referenced above. Wang further teaches:
wherein the hardware processor is further configured to identify the at least one computing resource to accommodate the size and the content type is in response to determining that the at least one computing resource can store the input training dataset “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
Wang fails to teach determining that the at least one computing resource … is compatible with the content type.
However, Dubeyko teaches and is compatible with the content type. “Further, each data type (e.g., user data, metadata, executable code) may have certain characteristics known to the operating system 150 based on which the operating system may allocate memory to the supersets” [Dubeyko ¶ 51]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc. Based on these superset features, the operating system may determine whether volatile or non-volatile memory is more suitable for the data/code being stored in the superset. If the selected memory category (volatile or non-volatile) has multiple memory types, the operating system may also determine which of the various memory types may be most suitable for the data/code in the superset” [Dubeyko ¶ 24].
With regard to claim 20, Wang teaches:
A non-transitory computer readable medium storing thereon computer executable instructions for provisioning artificial intelligence resources, including instructions for: “It should be noted that a person skilled in the art would understand that the implementation of all or part of the flows in the methods of the above-mentioned embodiments may be performed by a computer program instructing relevant hardware, and a program of a data set and node cache-based scheduling method may be stored in a computer-readable storage medium, and when executed, the program may include the flows of the embodiments of the methods as described above” [Wang Col. 8 Lines 57-65].
receiving an input training dataset “Generally, when initiating a training task, the algorithm personnel usually need to manually download these data to a node to start the training task; however, with regard to the AI resource management platform, a manual download data set is usually optimized as an automatic download data set; and when starting a training task, the AI resource management platform will automatically download the required data set for the training task. As the AI resource management platform, a variety of data sets will be provided for the algorithm personnel, and these data sets will be cached to a computing node according to the requirements of training tasks” [Wang Col. 1 Lines 31-42].
and an indication of a task to perform using the input training dataset, “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
determining a size of the input training dataset “According to step S200, the user submits a training task on the resource management platform, operation information of the training task includes the used data set information including a name of the data set, a unique identifier of the data set dataSettaskid used by the task, a size of the data set dataSettaskSize used by the task, and other basic resource information (CPU, memory, graphics processing unit (GPU), etc.) for operating the training task…” [Wang Col. 5 Lines 8-16].
identifying, from the plurality of computing resources, at least one computing resource to accommodate the size “… and after receiving the resource request of the training task, the scheduler firstly uses a kubernetes default algorithm to screen out nodes with sufficient CPU, memory and GPU cards” [Wang Col. 5 Lines 16-19]. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “… selecting a host node that executes the training task from a plurality of pending host nodes based on the scheduling strategy in response to screening out the plurality of pending host nodes that satisfy the space required for the training task from the host nodes” [Wang Col. 6 Lines 9-13].
and executing, on the at least one computing resource, the trained artificial intelligence model to perform the task. “According to the present disclosure, an AI training task may be operated on a host node with a required data set or a host node with sufficient node storage space…” [Wang Col. 3 Lines 49-51]. “FIG. 3 shows a schematic block diagram of an embodiment of a data set and node cache-based scheduling device according to the present disclosure, as shown in FIG. 3, the device 101 includes: … a training task execution module 15 configured to obtain and delete an obsolete data set cache in the host node to be executed, and execute the training task in the host node to be executed” [Wang Col. 8 Lines 9-12, 29-32].
Wang fails to teach determining a size of the input training dataset and a content type of an entry in the input training dataset; identifying, from a plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identifying attributes of the input training dataset.
However, Dubeyko teaches:
determining a size of the input training dataset and a content type of an entry in the input training dataset; “In other embodiments, the data type may include other or additional data types. "User data" is any data that is created or owned by a user. The user data may be of different types. For example, some user data (e.g., database user data) may be frequently updated, while other types of user data (e.g., video or image data) may be infrequently updated ("cold" data). Thus, in some embodiments, the user data may be categorized based upon the frequency of updates. Other categorizations of user data may be used in other embodiments” [Dubeyko ¶ 50]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
identifying, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; “The operating system 150 may determine the data type from the extended attributes of files in the file system (e.g., extended attributes of the applications 130). In other embodiments, the operating system 150 may be configured to determine the data type in other ways, as discussed above. Further, each data type (e.g., user data, metadata, executable code) may have certain characteristics known to the operating system 150 based on which the operating system may allocate memory to the supersets” [Dubeyko ¶ 51]. “Therefore, upon determining the data type at the operation 805, the operating system 150 determines which of the memory characteristics to prioritize over other memory properties for identifying the most suitable memory category for that data type. For example, if the operating system 150 determines the data type to be metadata, the operating system may conclude that metadata is frequently updated data, and therefore, a fast volatile memory and/or volatile memory having good memory endurance is more critical than a memory with low power consumption” [Dubeyko ¶ 77, Fig. 8]. “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc. Based on these superset features, the operating system may determine whether volatile or non-volatile memory is more suitable for the data/code being stored in the superset. If the selected memory category (volatile or non-volatile) has multiple memory types, the operating system may also determine which of the various memory types may be most suitable for the data/code in the superset” [Dubeyko ¶ 24].
identifying attributes of the input training dataset; “Specifically, upon receiving a request for memory allocation, the operating system identifies one or more superset features from the application requesting memory allocation. These superset features may be one or more of data type, workload type, power requirement, latency, locking primitive, hardware acceleration engine, etc.” [Dubeyko ¶ 24].
It would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang to incorporate the teachings of Dubeyko and include: determining a size of the input training dataset and a content type of an entry in the input training dataset; identifying, from the plurality of computing resources, at least one computing resource to accommodate the size and the content type associated with the input training dataset; identifying attributes of the input training dataset. Doing so would allow for further efficiency in task execution through the allocation of memory. “Thus, each superset is an abstraction that is allocated memory from the physical memory based on one or more features that facilitates efficient execution of the associated application” [Dubeyko ¶ 24].
Wang in view of Dubeyko fails to teach selecting, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset.
However, Arora teaches:
selecting, from a plurality of artificial intelligence models, an artificial intelligence model “In some embodiments, the system may automatically select a machine learning model for a user based on user's data, including consideration based on the type of data (e.g., image, text, etc.) and/or based on how various machine learning models are capable of handling any particular type of data (e.g., based on one or more performance metrics)” [Arora ¶ 24].
that is configured to perform the task “For example, a facial recognition model may be associated with images. In this case, the system may identify one or more machine learning models associated with images. For each of the one or more machine learning models associated with the data type, the system may generate a respective metric value indicating a performance of the machine learning model over the input data; and select a machine learning model from among the one or more machine learning models by comparing metric values generated for each of the one or more machine learning models” [Arora ¶ 32].
and is compatible with the attributes of the input training dataset; “As discussed above, in some embodiments the system may identify one or more machine learning models (e.g., based on matching the data type associated with the machine learning models to user's data), and automatically select a suitable machine learning model” [Arora ¶ 31].
training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset; “For example, to deploy a facial recognition system, a user may select several components in the workflow, including preparing training data containing facial recognition training images, selecting a machine learning model for facial recognition, training the machine learning model using the training data, evaluating the trained machine learning model, configuring business logics such as notification upon detection of certain subjects, and deploying the machine learning model” [Arora ¶ 53].
It would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko to incorporate the teachings of Arora and include: selecting, from a plurality of artificial intelligence models, an artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; training, on the at least one computing resource, the artificial intelligence model to perform the task using the input training dataset. Doing so would allow for the automatic selection of a suitable machine learning algorithm for user needs. “These techniques may be executed by a system that provides a graphical user interface which allows a user to visually define a workflow for a machine learning application, without requiring the user to be an expert in machine learning or programming. The system may automatically represent the workflow as a specification that may be used to build and deploy a machine learning application” [Arora ¶ 24].
Wang in view of Dubeyko in view of Arora fails to teach wherein the indication of the task includes a time limit for performing the task; identifying respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit.
However, Rakshit teaches:
wherein the indication of the task includes a time limit for performing the task; “User or processing activities on an edge device sometimes have a performance requirement and a defined priority within a defined service level agreement (SLA) within which the activity must be completed” [Rakshit Col. 3 Lines 15-18]. “In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 9-11].
identifying respective processing and networking limits of each of a plurality of computing resources; “According to aspects of the invention, the polling module 541 is configured to obtain performance data from each of the edge computing device 525a-n. In embodiments, the performance data includes maximum processing capacity of the particular edge computing device, current processing capacity of the particular edge computing device, maximum storage (memory) capacity of the particular edge computing device, current storage (memory) capacity of the particular edge computing device, network speed of the particular edge computing device, and communications protocols supported by the particular edge computing device” [Rakshit Col. 16 Lines 1-11].
by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit; “In this manner, aspects of the invention utilize AI to select an optimal subset of edge computing devices and communication protocols to complete computational tasks within a defined performance metric (e.g., a time limit defined in an SLA). In embodiments, the AI includes a machine learning model that is configured to select the optimal subset of edge computing devices and communication protocols to minimize latency while meeting a time limit defined in an SLA” [Rakshit Col. 4 Lines 3-11]. “While performing any computational task in edge computing, data transportation time and data processing time are factors that affect the ability to meet the SLA or performance metrics for the task. For an edge computing computational task that has a bound completion time, e.g., as defined in an SLA, there is a need for a system to identify which edge computing devices to use for data processing for completing the task and how data is transported to such devices optimally” [Rakshit Col. 3 Lines 56-64].
Rakshit is considered to be analogous to the claimed invention because it is in the same field of computing resource provisioning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora to incorporate the teachings of Rakshit and include: wherein the indication of the task includes a time limit for performing the task; identifying respective processing and networking limits of each of a plurality of computing resources; by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit. Doing so would allow for optimization of bandwidth usage through the allocation of computing resources to perform tasks. “With continued reference to FIG. 4, at step 404 the AI-enabled system optimizes the processing power and bandwidth usage for data transportation when performing the assigned computational task” [Rakshit Col. 12 Lines 15-18].
Claims 7, 8, 18, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11,698,863 B1) in view of Dubeyko (US 2020/0257562 A1) in view of Arora (US 2022/0206774 A1) in view of Rakshit (US 11,553,038 B1) in view of Perumalla (US 12,067,482 B1).
With regard to claim 7, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang in view of Dubeyko in view of Arora in view of Rakshit fails to teach wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values.
However, Perumalla teaches wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values. “The data may be provided by one or more "edge" electronic devices in a first format, and the preprocessing adapter can seamlessly convert the data of the first format into one or more multiple different formats, via preprocessing operations, to be used with one or more ML models” [Perumalla Col. 2 Lines 38-42]. “For example, as shown by circle '2A', embodiments can analyze the first input layer 106 of a ML model 102A (e.g., identified in the configuration data 204) to identify the input characteristic requirements of the model, and thus, the user beneficially doesn't need to know the specifics of the model's architecture. In some embodiments, the PPA 112 can interact with the ML model by looking into the data structures of the first input layer 106 of the model to identify, e.g., what dimensions of an image are required (e.g., a line of data of the input layer may read "input=30x30" and thus, it can be determined that the model requires a 30 pixel by 30 pixel image as input data). By looking into this first input layer, the configuration engine 206 can learn what is required in terms of preprocessing” [Perumalla Col. 9 Lines 4-17].
Perumalla is considered to be analogous to the claimed invention because it is in the same field of machine learning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora in view of Rakshit to incorporate the teachings of Perumalla and include: wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values. Doing so would allow for the input dataset to be changed into a format with the same dimensions as required by any machine learning model selected for use. “The data may be provided by one or more "edge" electronic devices in a first format, and the preprocessing adapter can seamlessly convert the data of the first format into one or more multiple different formats, via preprocessing operations, to be used with one or more ML models” [Perumalla Col. 2 Lines 38-42].
With regard to claim 8, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang in view of Dubeyko fails to teach further comprising: amending code associated with the artificial intelligence model to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model.
However, Arora teaches:
further comprising: amending code associated with the artificial intelligence model “Yet due to the above described process by which the application is built, adapting or otherwise changing the way the application is tuned can require a great deal of time and expertise. For instance, a reconfiguration may require a user to re-build the system and/or even re-write some of the code” [Arora ¶ 23]. “The engineer may also need to identify various data sources for training the machine learning model (also referred to herein as "training data"), and may need to configure the system to perform numerous steps of data preprocessing and/or data cleansing to properly prepare training data for use” [Arora ¶ 21]. “For instance, consider a workflow that includes a data preprocessing context following by a training context that includes the component for which a model is to be selected. The data preprocessing context may first be executed as part of act 404 to produce the data to be evaluated since this data is to be input into the training context” [Arora ¶ 81].
to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model. “For instance, consider a workflow that includes a data preprocessing context following by a training context that includes the component for which a model is to be selected. The data preprocessing context may first be executed as part of act 404 to produce the data to be evaluated since this data is to be input into the training context” [Arora ¶ 81]. “In this example, components 602 and 604 are associated with a data processing context, where component 602 may be configured to perform text preprocessing by cleaning the text and formatting the text data, and component 604 may perform vectorization operation to convert the preprocessed text data into numerical data” [Arora ¶ 96].
Wang in view of Dubeyko in view of Arora in view of Rakshit fails to explicitly teach to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model.
However, Perumalla teaches to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model. “For example, as shown by circle '2A', embodiments can analyze the first input layer 106 (native training dataset) of a ML model 102A (e.g., identified in the configuration data 204) to identify the input characteristic requirements of the model, and thus, the user beneficially doesn't need to know the specifics of the model's architecture. In some embodiments, the PPA 112 can interact with the ML model by looking into the data structures of the first input layer 106 of the model to identify, e.g., what dimensions of an image are required (e.g., a line of data of the input layer may read "input=30x30" and thus, it can be determined that the model requires a 30 pixel by 30 pixel image as input data). By looking into this first input layer, the configuration engine 206 can learn what is required in terms of preprocessing” [Perumalla Col. 9 Lines 4-17 Examiner notes this interpretation is in accordance with the example given in ¶ 55 of the instant specification].
With regard to claim 18, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 12, as referenced above. Wang in view of Dubeyko in view of Arora in view of Rakshit fails to teach wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values.
However, Perumalla teaches wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values. “The data may be provided by one or more "edge" electronic devices in a first format, and the preprocessing adapter can seamlessly convert the data of the first format into one or more multiple different formats, via preprocessing operations, to be used with one or more ML models” [Perumalla Col. 2 Lines 38-42]. “For example, as shown by circle '2A', embodiments can analyze the first input layer 106 of a ML model 102A (e.g., identified in the configuration data 204) to identify the input characteristic requirements of the model, and thus, the user beneficially doesn't need to know the specifics of the model's architecture. In some embodiments, the PPA 112 can interact with the ML model by looking into the data structures of the first input layer 106 of the model to identify, e.g., what dimensions of an image are required (e.g., a line of data of the input layer may read "input=30x30" and thus, it can be determined that the model requires a 30 pixel by 30 pixel image as input data). By looking into this first input layer, the configuration engine 206 can learn what is required in terms of preprocessing” [Perumalla Col. 9 Lines 4-17].
It would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora in view of Rakshit to incorporate the teachings of Perumalla and include: wherein the attributes comprise one or more of: (1) dimensions of the entry in the input training dataset, (2) a number of input values in an entry, (3) a number of output values in an entry, (4) a number of unique output values in the input training dataset, and (5) a balance between the unique output values. Doing so would allow for the input dataset to be changed into a format with the same dimensions as required by any machine learning model selected for use. “The data may be provided by one or more "edge" electronic devices in a first format, and the preprocessing adapter can seamlessly convert the data of the first format into one or more multiple different formats, via preprocessing operations, to be used with one or more ML models” [Perumalla Col. 2 Lines 38-42].
With regard to claim 19, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the system of claim 12, as referenced above. Wang in view of Dubeyko fails to teach wherein the hardware processor is further configured to: amend code associated with the artificial intelligence model to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model.
However, Arora teaches:
wherein the hardware processor is further configured to: amend code associated with the artificial intelligence model “Yet due to the above described process by which the application is built, adapting or otherwise changing the way the application is tuned can require a great deal of time and expertise. For instance, a reconfiguration may require a user to re-build the system and/or even re-write some of the code” [Arora ¶ 23]. “The engineer may also need to identify various data sources for training the machine learning model (also referred to herein as "training data"), and may need to configure the system to perform numerous steps of data preprocessing and/or data cleansing to properly prepare training data for use” [Arora ¶ 21]. “For instance, consider a workflow that includes a data preprocessing context following by a training context that includes the component for which a model is to be selected. The data preprocessing context may first be executed as part of act 404 to produce the data to be evaluated since this data is to be input into the training context” [Arora ¶ 81].
to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model. “For instance, consider a workflow that includes a data preprocessing context following by a training context that includes the component for which a model is to be selected. The data preprocessing context may first be executed as part of act 404 to produce the data to be evaluated since this data is to be input into the training context” [Arora ¶ 81]. “In this example, components 602 and 604 are associated with a data processing context, where component 602 may be configured to perform text preprocessing by cleaning the text and formatting the text data, and component 604 may perform vectorization operation to convert the preprocessed text data into numerical data” [Arora ¶ 96].
Wang in view of Dubeyko in view of Arora in view of Rakshit fails to explicitly teach to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model.
However, Perumalla teaches to accommodate structure differences between the input training dataset and a native training dataset used to train the artificial intelligence model. “For example, as shown by circle '2A', embodiments can analyze the first input layer 106 (native training dataset) of a ML model 102A (e.g., identified in the configuration data 204) to identify the input characteristic requirements of the model, and thus, the user beneficially doesn't need to know the specifics of the model's architecture. In some embodiments, the PPA 112 can interact with the ML model by looking into the data structures of the first input layer 106 of the model to identify, e.g., what dimensions of an image are required (e.g., a line of data of the input layer may read "input=30x30" and thus, it can be determined that the model requires a 30 pixel by 30 pixel image as input data). By looking into this first input layer, the configuration engine 206 can learn what is required in terms of preprocessing” [Perumalla Col. 9 Lines 4-17 Examiner notes this interpretation is in accordance with the example given in ¶ 55 of the instant specification].
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11,698,863 B1) in view of Dubeyko (US 2020/0257562 A1) in view of Arora (US 2022/0206774 A1) in view of Rakshit (US 11,553,038 B1) in view of Natarajan (US 2021/0064418 A1).
With regard to claim 9, Wang in view of Dubeyko in view of Arora in view of Rakshit teaches the method of claim 1, as referenced above. Wang in view of Dubeyko fails to teach wherein the indication of the task further includes a target error rate, further comprising: determining whether an error rate of the trained artificial intelligence model is greater than the target error rate;
However, Arora teaches:
wherein the indication of the task further includes a target error rate, “As further shown in FIG. 11, the user interface 1100 may allow the user to enter one or more parameters associated with selecting the machine learning model. For example, the user interface may include a widget 1104 (e.g., a slider bar) to allow the user to define an accuracy threshold (target error rate) for evaluating the machine learning model” [Arora ¶ 99].
further comprising: (evaluating) determining whether an error rate of the trained artificial intelligence model (using) is greater than the target error rate; “Alternatively, if the data type is image data and/or if the machine learning model class is identified to be an image model, other performance metrics may be evaluated as relating to the performance of different image models on the same image data. For example, for a classification model, the performance metrics for evaluating the classification model may include a true positive, a false positive, a false negative, a true negative, an accuracy rate, a precision rate, a recall rate, a F1 score, or a combination thereof” [Arora ¶ 83]. “For example, for a regression application, the system may select the model that has the lowest error among the different machine learning models” [Arora ¶ 88].
Wang in view of Dubeyko in view of Arora in view of Rakshit fails to explicitly teach further comprising: determining whether an error rate of the trained artificial intelligence model is greater than the target error rate; and in response to determining that the error rate of the trained artificial intelligence model is greater than the target error rate, re-training the trained artificial intelligence model.
However, Natarajan teaches:
further comprising: determining whether an error rate of the trained artificial intelligence model is greater than the target error rate; “An accuracy value associated with machine-learning models 132 may be used to trigger re-training or provisioning new machine-learning models. If the accuracy value falls below a first threshold value then the re-training or provisioning may be triggered” [Natarajan ¶ 30].
and in response to determining that the error rate of the trained artificial intelligence model is greater than the target error rate, re-training the trained artificial intelligence model. “An accuracy value associated with machine-learning models 132 may be used to trigger re-training or provisioning new machine-learning models. If the accuracy value falls below a first threshold value then the re-training or provisioning may be triggered” [Natarajan ¶ 30].
Natarajan is considered to be analogous to the claimed invention because it is in the same field of machine learning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora in view of Rakshit to incorporate the teachings of Natarajan and include: determining whether an error rate of the trained artificial intelligence model is greater than the target error rate; and in response to determining that the error rate of the trained artificial intelligence model is greater than the target error rate, re-training the trained artificial intelligence model. Doing so would allow the system to address the issue of machine learning model accuracy falling below a target level. “If the accuracy value falls below a first threshold value then the re-training or provisioning may be triggered” [Natarajan ¶ 30].
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11,698,863 B1) in view of Dubeyko (US 2020/0257562 A1) in view of Arora (US 2022/0206774 A1) in view of Rakshit (US 11,553,038 B1) in view of Natarajan (US 2021/0064418 A1) in view of Merrill (US 2019/0279111 A1).
With regard to claim 10, Wang in view of Dubeyko in view of Arora in view of Rakshit in view of Natarajan teaches the method of claim 9, as referenced above. Wang in view of Dubeyko fails to teach selecting, from the plurality of artificial intelligence models, a different artificial intelligence model that is configured to perform the task and is compatible with the attributes of the input training dataset; and training, using the input training dataset, the different artificial intelligence model to perform the task.
However, Arora teaches:
selecting, from the plurality of artificial intelligence models, a different artificial intelligence model that is configured to perform the task “For example, a facial recognition model may be associated with images. In this case, the system may identify one or more machine learning models associated with images. For each of the one or more machine learning models associated with the data type, the system may generate a respective metric value indicating a performance of the machine learning model over the input data; and select a machine learning model from among the one or more machine learning models by comparing metric values generated for each of the one or more machine learning models” [Arora ¶ 32].
and is compatible with the attributes of the input training dataset; “As discussed above, in some embodiments the system may identify one or more machine learning models (e.g., based on matching the data type associated with the machine learning models to user's data), and automatically select a suitable machine learning model” [Arora ¶ 31].
and training, using the input training dataset, the different artificial intelligence model to perform the task “For example, to deploy a facial recognition system, a user may select several components in the workflow, including preparing training data containing facial recognition training images, selecting a machine learning model for facial recognition, training the machine learning model using the training data, evaluating the trained machine learning model, configuring business logics such as notification upon detection of certain subjects, and deploying the machine learning model” [Arora ¶ 53].
Wang in view of Dubeyko in view of Arora in view of Rakshit fails to explicitly teach further comprising: in response to determining that an error rate of the re-trained artificial intelligence model is greater than the target error rate, selecting, from the plurality of artificial intelligence models, a different artificial intelligence model and training, using the input training dataset, the different artificial intelligence model to perform the task at an error rate not greater than the target error rate.
However, Natarajan teaches:
further comprising: in response to determining that an error rate of the re-trained artificial intelligence model is greater than the target error rate, “If the accuracy falls below both the first threshold and the second threshold, then a new machine-learning model may be provisioned. The new machine-learning model may be trained in the same manner as described above” [Natarajan ¶ 30].
selecting, from the plurality of artificial intelligence models, a different artificial intelligence model “In the instance of provisioning, the machine-learning model may be replaced with a new machine-learning model. The new machine-learning model may be trained in the same manner as described above” [Natarajan ¶ 30].
and training, using the input training dataset, the different artificial intelligence model to perform the task at an error rate not greater than the target error rate. “In another example, training may continue until the machine-learning model 132 reaches a predetermined accuracy threshold” [Natarajan ¶ 29].
Wang in view of Dubeyko in view of Arora in view of Rakshit in view of Natarajan fails to explicitly teach further comprising: in response to determining that an error rate of the re-trained artificial intelligence model.
However, Merrill teaches further comprising: in response to determining that an error rate of the re-trained artificial intelligence model “In some embodiments, accuracy of a retrained model is determined to determine the impact on model accuracy from removal of the potentially suspect features. In some embodiments, information indicating accuracy of the retrained model is transmitted to the operator device (e.g., as a message, a data object, a user interface element, and the like)” [Merrill ¶ 106].
Merrill is considered to be analogous to the claimed invention because it is in the same field of machine learning. Therefore, it would be obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Wang in view of Dubeyko in view of Arora in view of Rakshit in view of Natarajan to incorporate the teachings of Merrill and include: in response to determining that an error rate of the re-trained artificial intelligence model. Doing so would allow for comparison between the accuracy of the trained machine learning model and the accuracy of the re-trained machine learning model. “In some embodiments, the model evaluation system provides a graphical user interface to the operator device that displays the accuracy of the retrained model, a and the potentially suspect features removed from the original model. In some embodiments, the graphical user interface displays a selection element for receiving user selection of one of the original model and the retrained model for use in production” [Merrill ¶ 106].
Response to Arguments
Applicant's arguments filed 01/05/2026 have been fully considered but they are not persuasive. Applicant argues in substance:
I. A. Step 2A, Prong 1: The Claims are Not Directed to an Abstract Idea
The amended claims are not directed to a mental process. They recite a specific technical method for provisioning computing resources within a distributed system. The claims require the system to query "processing and networking limits" of physical hardware and determine if those physical limits allow for the "successful completion of the task within the time limit."
A human cannot perform these steps mentally. A human cannot look at a dataset and a server rack and mentally calculate whether the server's specific packet-switching network limits and CPU cycle limits will allow a complex Al training task to finish within a specific duration (e.g., 2 hours). This requires complex computation of data throughput vs. hardware capability.
a) Examiner respectfully disagrees. In response to applicant's argument that the amended claims are not directed to a mental process, it is noted that the features upon which applicant relies (i.e., “require the system to query "processing and networking limits" of physical hardware” and “calculate whether the server's specific packet-switching network limits and CPU cycle limits will allow a complex Al training task to finish within a specific duration (e.g., 2 hours)”) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). Further, as detailed in the rejection above, the independent claims recite mental processes of determining, identifying, and selecting. These limitations as claimed are able to be performed in the human mind through observation and evaluation; a person can mentally compare a time limit to processing and network abilities of a computing resource. Further, the additional elements of the claims and the claims as a whole fail to integrate the judicial exception of the abstract ideas into a practical application, thus, the claims are directed to the judicial exception. The arguments have been considered but were not found to be persuasive.
II. B. Step 2A, Prong 2: Integration into a Practical Application
The claims are integrated into a practical application because they provide a specific improvement to the functioning of the computing system itself, specifically in how distributed hardware resources are provisioned to prevent technical failure.
The Specification explicitly describes the technical problem addressed: avoiding "software or hardware failures (e.g., crashes)" when processing Al tasks. (See Specification at [0044]) The Specification notes that "users are unable to identify the best A.I. models... as there are many factors to consider such as the availability of required computing resources". (See Specification at [0003])
The amended claims solve this by enforcing a specific hardware-selection logic: determining if "processing and networking limits enable successful completion of the task within the time limit." This is not a mere administrative task; it is a technical calculation of hardware throughput. The Specification provides specific support for how this improves the computer's operation: Preventing Time-Outs via Hardware Selection: In Paragraph [0038], the Specification describes how the system resolves conflicts between budget and time constraints by selecting specific hardware architectures (e.g., GPU vs. CPU). It states: "Using a CPU may cause for the amount of time it takes to train the algorithm to exceed the time constraint. In order to meet the time constraint, a GPU may be needed". The claims recite this specific technical step of selecting the resource that meets the time limit based on processing limits. Hardware-Specific Analysis: Paragraph [0032] details the analysis of specific hardware attributes, noting that "Server 1 is the more capable computing resource... beyond a certain limit, server 1 is a better option". Ensuring Successful Execution: Paragraph [0044] defines "accommodating" as specifically enabling "successful completion of the task... without any software or hardware failures". By tying the resource selection to specific "processing and networking limits" to meet a "time limit," the claims prevent the technical problem of a task failing due to insufficient hardware throughput or timing out. This improves the efficiency and reliability of the computing system, satisfying Step 2A, Prong 2.
a) Examiner respectfully disagrees. As detailed in the rejection above, the independent claims recite mental processes of determining, identifying, and selecting. The improvements in Applicants argument are directed to these mental processes. A person can still use mental processes to select computing resources by taking into account processing and networking limits and a time limit. Improvements to a determination and selection of resources does not represent improvements to computer functionality or other technology, as per MPEP: “… it is important to keep in mind that an improvement in the abstract idea itself (e.g. a recited fundamental economic concept) is not an improvement in technology. For example, in Trading Technologies Int’l v. IBG, 921 F.3d 1084, 1093-94, 2019 USPQ2d 138290 (Fed. Cir. 2019), the court determined that the claimed user interface simply provided a trader with more information to facilitate market trades, which improved the business process of market trading but did not improve computers or technology” [MPEP § 2106.05(a) II]. An improvement to the abstract ideas of determining, identifying, and selecting does not amount to significantly more than the abstract ideas. The arguments have been considered but were not found to be persuasive.
III. C. Step 2B: The Claims Amount to "Significantly More"
Even if the claims were viewed as directed to an abstract idea, the ordered combination of claim elements amounts to "significantly more" than the abstract idea itself.
The claims require a specific, non-conventional sequence of steps: 1. Receiving a user-defined time limit for an AI task; 2. Querying and identifying the respective processing and networking limits of a plurality of physical computing resources; and 3. Selecting a specific resource only if its physical limits allow the task to finish within that time limit.
This is not a "well-understood, routine, conventional activity." Generic computer usage involves assigning a task to any available processor. The claimed invention involves a specialized "Computing Evaluation Module" (see Specification at [0030], [0037]) that performs a complex pre-execution analysis of hardware capabilities against temporal constraints. As noted in Paragraph [0045], this involves determining that "processing and network limits enable successful completion... within the time limit".
This adds a specific inventive concept: the predictive matching of hardware physics (networking/processing limits) to task duration requirements to ensure successful execution. This transforms the abstract idea of "provisioning" into a specific, patent-eligible technical process for reliable distributed computing.
Therefore, the claims are eligible under § 101.
a) Examiner respectfully disagrees. In response to applicant's argument that the claims amount to significantly more than the abstract ideas, it is noted that the features upon which applicant relies (i.e., “querying … the respective processing and networking limits of a plurality of physical computing resources”, “The claimed invention involves a specialized "Computing Evaluation Module"… that performs a complex pre-execution analysis of hardware capabilities against temporal constraints”) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
Further, the additional elements of the claims amount to no more than generic computing components, insignificant extra solution activity, and technological environment/field of use which, when considered alone and in combination, do not amount to significantly more than the abstract ideas. Receiving a time limit is merely a recitation of data reception which is insignificant extra solution activity that is well-understood, routine, and conventional activity [MPEP§ 2106.05(d)(II)]. As detailed above, identifying processing and networking limits is a process which can be performed in the human mind and is thus an abstract idea. Selecting a resource to enable successful completion of a task within a time limit is also a process which can be performed in the human mind or with pencil and paper. “As explained by the Supreme Court, in order to make a claim directed to a judicial exception patent-eligible, the additional element or combination of elements must do "‘more than simply stat[e] the [judicial exception] while adding the words ‘apply it’". Alice Corp. v. CLS Bank, 573 U.S. 208, 221, 110 USPQ2d 1976, 1982-83 (2014) (quoting Mayo Collaborative Servs. V. Prometheus Labs., Inc., 566 U.S. 66, 72, 101 USPQ2d 1961, 1965). Thus, for example, claims that amount to nothing more than an instruction to apply the abstract idea using a generic computer do not render an abstract idea eligible” [MPEP§ 2106.05(f)]. Currently the claims do not include additional elements amounting to significantly more than the recited judicial exception. The arguments have been considered but were not found to be persuasive.
IV. Claim 1 as amended requires: 1. Receiving a "time limit" for the task. 2. Identifying "processing and networking limits" of the computing resources. 3. Selecting the resource "in response to determining that the processing and network limits enable successful completion of the task within the time limit."
Wang teaches a storage-based scheduling method. Wang checks if a host node has "sufficient node storage space" (Wang, Abstract; Col. 3, lines 49-51). Wang is concerned with caching datasets and deleting obsolete caches. Wang does not teach receiving a time limit from the user for the task, nor does it teach calculating whether a node's processing (CPU speed) and networking (bandwidth) limits will allow the task to finish within that time limit. Wang merely checks if the data fits (storage).
Dubeyko teaches memory management within a single machine. It distinguishes between "hot" and "cold" data to decide whether to place data in volatile vs. non-volatile memory (Dubeyko, Para [0050], [0077]). Dubeyko is about memory tiering to optimize local execution speed or durability. It does not teach selecting a computing resource (e.g., a server from a plurality of servers) based on whether that server's network and processing limits allow a task to meet a user-defined time limit.
Arora teaches a workflow for building AI applications (Arora, Abstract). It discusses selecting a model based on data type (e.g., image vs. text). It does not teach the hardware provisioning logic of analyzing hardware processing network limits against a specific time constraint to select a specific server.
a) The prior art sources described in Applicants argument, Wang, Dubeyko, and Arora, are not cited in the rejection above to teach the argued claim limitations. As detailed in the rejection above, Rakshit teaches the claimed: wherein the indication of the task includes a time limit for performing the task; [Rakshit Col. 3 Lines 15-18, Col. 4 Lines 9-11] identifying respective processing and networking limits of each of a plurality of computing resources; [Rakshit Col. 16 Lines 1-11] by selecting the at least one computing resource in response to determining that the processing and network limits enable successful completion of the task within the time limit; [Rakshit Col. 4 Lines 3-11, Col. 3 Lines 56-64]. The arguments have been considered but were not found to be persuasive.
V. The Examiner argues that Dubeyko allows for "efficient execution." However, Dubeyko optimizes memory allocation (RAM vs SSD), not server provisioning based on time constraints. There is no motivation to combine Wang's storage-based cluster scheduler with Dubeyko's single- node memory tiering to arrive at the Applicant's invention of time-based, processing-limit-based resource provisioning. Wang cares about disk space; Dubeyko cares about data frequency. Neither cares about meeting a specific "2-hour" (e.g.) deadline by analyzing network throughput.
The combination fails to teach the specific claim limitations regarding the time limit and the comparison of processing network limits to that time limit.
The amended claims recite patent-eligible subject matter and contain limitations not found in the prior art. Withdrawal of the rejections and allowance of the claims are earnestly solicited.
a) Examiner respectfully disagrees. In response to applicant’s argument that there is no teaching, suggestion, or motivation to combine the references, the examiner recognizes that obviousness may be established by combining or modifying the teachings of the prior art to produce the claimed invention where there is some teaching, suggestion, or motivation to do so found either in the references themselves or in the knowledge generally available to one of ordinary skill in the art. See In re Fine, 837 F.2d 1071, 5 USPQ2d 1596 (Fed. Cir. 1988), In re Jones, 958 F.2d 347, 21 USPQ2d 1941 (Fed. Cir. 1992), and KSR International Co. v. Teleflex, Inc., 550 U.S. 398, 82 USPQ2d 1385 (2007). In this case, as detailed in the rejection above, combining Wang with the teachings of Dubeyko would provide further efficiency to the system by improving the use of memory based on features of application datasets.
Further, the combination of Wang in view of Dubeyko is not cited in the rejection above to teach the argued claim limitations; these limitations are taught by the prior art Rakshit. The arguments have been considered but were not found to be persuasive.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Examiner respectfully requests, in response to this Office action, support be shown for language added to any original claims on amendment and any new claims. That is, indicate support for newly added claim language by specifically pointing to page(s) and line number(s) in the specification and/or drawing figure(s). This will assist Examiner in prosecuting the application.
When responding to this Office Action, Applicant is advised to clearly point out the patentable novelty which he or she thinks the claims present, in view of the state of the art disclosed by the references cited or the objections made. He or she must also show how the amendments avoid such references or objections. See 37 CFR 1.111(c).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ARI F RIGGINS whose telephone number is (571)272-2772. The examiner can normally be reached Monday-Friday 7:00AM-4:30PM.
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, Bradley Teets can be reached at (571) 272-3338. 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.
/A.F.R./Examiner, Art Unit 2197
/BRADLEY A TEETS/Supervisory Patent Examiner, Art Unit 2197
Prosecution Insights