Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
DETAILED ACTION
2. This action is in response to the amendment filed August 21, 2026.
3. Claims 1-22 have been examined and are pending with this action.
Response to Arguments
4. Applicant's arguments filed August 21, 2026, with respect to the rejection of claims 1-10, 12-14, 16, and 18-22, previously rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) in view of Jha et al (US 2020/0027022 A1), have been fully considered and are persuasive. Therefore, the rejection has been withdrawn.
However, upon further consideration, a new ground(s) of rejection is made in view of Chaudhuri et al. (US 2019/0378028 A1). Chaudhuri has been cited to better teach a response generated by the first machine learning model is forwarded to a second, as argued. Please see rejections set forth below. This action is Non-Final.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
5. Claims 1, 3-10, 12-14, 16, and 18-22 are rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) in view of Chaudhuri et al. (US 2019/0378028 A1).
As per claim 1, Gazzetti teaches a method, comprising:
configuring one or more first machine learning models in a first network located between a client, and a second network, the second network hosting a second machine learning model, wherein, as between at least one first machine learning model and the second machine learning model (see Gazzetti, FIG. 2; Abstract: “Embodiments for processing data with multiple machine learning models are provided… The input data is caused to be evaluated by a first machine learning model to generate a first inference result.”; col.11, lines 42-46: “Each of the machine learning models may be configured to perform, for example, object detection, object classification, motion detection, etc. The at least one QoS parameter may be associated with, for example, latency, precision, recall, etc.”; and col.16, lines 35-39: “The edge service 1604 may be implemented in any suitable computing device ( e.g., perhaps integrated into the same device as the rest API 1602) and is configured process the input data (e.g., the frame buffer 1608) with a (first) machine learning model ( e.g., an object classification model)”); and
responsive to receipt of a request by a given one of the first machine learning models, generating a response to the request (see Gazzetti, col.1, lines 42-44: “… evaluated by a first machine learning model to generate a first inference result.”; and col.11, lines 33-35: “The input data is caused to be evaluated by a first machine learning model to generate a first inference result.”).
Although Gazzetti teaches forwarding onward for additional handling (see Gazzetti, col.1, lines 47-49: “the input data is caused to be evaluated by a second machine learning model to generate a second inference result.”; col.5, lines 41-43: “The causing of the input data to be evaluated by the second machine learning model to generate a second inference result”; col.5, lines 58-61: “The first machine learning model may be implemented utilizing a first computing device, and the second machine learning model is implemented utilizing a second computing device.”; and col.15, lines 29-31: “Based on the policy and/or if appropriate, the advisor module 1406 may forward a request to the next inference service in the system 1400, such as inference service 1404”), Gazzetti does not explicitly teach the response is forwarded onward to the second machine learning model.
Chaudhuri teach the response is forwarded onward to the second machine learning model (see Chaudhuri, [0156]: “A DNN uses a cascade of many layers of non-linear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. Higher-level features are derived from lower-level features to form a hierarchical representation. The layers following the input layer may be convolution layers that produce feature maps that are filtering results of the inputs and are used by the next convolution layer.”; and [0207]: “executing the first PP to generate a set of first passing blobs that meet the first clause and a set of first failing blobs that do not meet the first clause; executing the second PP on the set of first failing blobs to generate a set of second passing blobs that meet the second clause and a set of second failing blobs that do not meet the second clause; and continuing the PP sequence with a union of the set of first passing blobs and the set of second passing blobs”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti in view of Chaudhuri so that the response is forwarded onward to the second machine learning model. One would be motivated to do so because Gazzetti teaches column 3, lines 32-35, “an inference system capable of “handing off” a user request to different (e.g., second, third, fourth, etc.) models may be provided”.
Although Gazzetti teaches the first machine learning model provides inferencing with less accuracy (see Gazzetti, col.3, lines 19-23: “To address these needs and/or the shortcomings in the prior art, in some embodiments described herein, methods and/or systems are disclosed that, for example, leverage multiple models with different levels of accuracy and achievable QoS”), Gazzetti does not explicitly teach at a lower cost.
Chaudhuri teaches the first machine learning model provides inferencing at a lower cost (see Chaudhuri, [0076]: “The basic intuition behind probabilistic predicates is akin to that of cascading classifiers in machine learning: a more efficient but inaccurate classifier can be used in front of an expensive classifier to lower the overall cost.”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti in view of Chaudhuri so that the first machine learning model provides inferencing at a lower cost. One would be motivated to do so because this enables the requestor to be able to select based on preference and Gazzetti explicitly teaches the ability to select inferencing threshold (see Gazzetti, col.17, lines 28-36: “Mechanisms for threshold computation for one or more model performing inferences in sequence may also be provided, routing inference requests to different models based on policies, and selecting an inference threshold based on specified QoS parameters may also be provided. Additionally, the system may evaluate inference results at runtime (or in "real-time") and redirect requests to different models based on the threshold(s) selected.”).
As per claim 3, which depends on claims 1, Gazzetti further teaches wherein, as between the first machine learning model and the second machine learning model, the first machine learning model is smaller in a scale of a language corpus on which the model is trained (see Gazzetti, col.11, lines 22-25: “In some embodiments, the level of QoS that may be expected when two or more networks ( or models) are utilized together to classify, detect, etc. elements in input data is evaluated.”; and col.16, lines 60-64: “reroute the response to the user or forward the input data to the cloud connector 1616 for a more specific classification ( or to be processed by a second model), and generate a set of labels that respect the user-defined constraints (e.g., the QoS parameters).”).
As per claim 4, which depends on claims 1, Gazzetti further teaches wherein the first network is one of: an edge region of a Content Delivery Network (CDN), and a datacenter hosting cloud compute infrastructure (see Gazzetti, col.16, lines 40-52: “If determined to be appropriate, the edge service 1604 may also interact with the cloud platform (or remote cloud platform) 1606 to request that the input data be processed with another (e.g. a second) machine learning model. As such, the cloud platform 1606 may include, for example, remote services/systems deployed in a cloud infrastructure, which are configured to process the input data with the second ( or third, fourth, etc.) machine learning model. It should be noted that because of the potentially (nearly) unlimited resource constraints of such a service ( e.g., on the cloud), the accuracy of inferences generated by such a system may exceed that of the edge service 1604 (e.g., implemented locally, on a single device, etc.).”).
As per claim 5, which depends on claims 1, Gazzetti further teaches wherein the response to the request represents an initial processing of the request (see Gazzetti, col.13, lines 34-41: “The upper bound may correspond to a fraction or portion of the data set that are to be evaluated or detected by a first model (e.g., Model A), which may be implemented locally (e.g. in an "edge" device). For example, in operation, if all of the predictions generated by the first model have a score that is above the upper bound, the predictions generated by the first model are utilized (and returned to the user).”).
As per claim 6, which depends on claims 5, Gazzetti further teaches wherein the initial processing performs one of: augmenting the request with additional context, changing the request, and screening the request (see Gazzetti, col.5, lines 2-7: “Utilizing the models, the performance (or operation) of the system (e.g., utilizing/based on new inputs) may be predicted and/or the performance of the system may be optimized by investigating how changes in the input(s) effect the output(s).”).
As per claim 7, which depends on claims 1, Gazzetti further teaches wherein the request originates at the client and is originally directed to the second machine learning model (see Gazzetti, col.11, lines 18-21: “Additionally, an inference system capable of "handing off' a user request to different (or second, third, fourth, etc.) models may be provided”).
As per claim 8, which depends on claims 1, Gazzetti and Chaudhuri further teach wherein the lower cost results from execution of the first machine learning model on infrastructure in the first network having an operating cost that is less than infrastructure in the second network on which the second machine learning model executes (see Gazzetti, col.16, lines 60-64: “reroute the response to the user or forward the input data to the cloud connector 1616 for a more specific classification ( or to be processed by a second model), and generate a set of labels that respect the user-defined constraints (e.g., the QoS parameters).”).
As per claim 9, which depends on claims 1, Gazzetti further teaches wherein the second machine learning model is a Large Language Model (LLM) (see Gazzetti, col.4, lines 16-22: “The cognitive analysis may include natural language processing (NLP) or a NLP technique, such classifying natural language, analyzing tone, and analyzing sentiment ( e.g., scanning for keywords, key phrases, etc.) with respect to, for example, content (or data), communications sent to and/or received by users, and/or other available data sources”).
As per claim 10, which depends on claims 1, Gazzetti further teaches wherein the first network is an edge network, and the second network is a network-accessible cloud compute infrastructure (see Gazzetti, col.5, lines 58-64: “The first machine learning model may be implemented utilizing a first computing device, and the second machine learning model is implemented utilizing a second computing device. The second computing device may be remote from the first computing device and in operable communication with the first computing device via a communications network ( e.g., on the "cloud").”; and col.16, lines 35-39: “The edge service 1604 may be implemented in any suitable computing device ( e.g., perhaps integrated into the same device as the rest API 1602) and is configured process the input data (e.g., the frame buffer 1608) with a (first) machine learning model ( e.g., an object classification model)”).
As per claim 12, which depends on claims 11, Gazzetti further teaches wherein the set of first machine learning models are configured to provide a collaborative learning inferencing task (see Gazzetti, col.16, lines 40-52: “If determined to be appropriate, the edge service 1604 may also interact with the cloud platform (or remote cloud platform) 1606 to request that the input data be processed with another (e.g. a second) machine learning model. As such, the cloud platform 1606 may include, for example, remote services/systems deployed in a cloud infrastructure, which are configured to process the input data with the second ( or third, fourth, etc.) machine learning model. It should be noted that because of the potentially (nearly) unlimited resource constraints of such a service ( e.g., on the cloud), the accuracy of inferences generated by such a system may exceed that of the edge service 1604 (e.g., implemented locally, on a single device, etc.).”).
As per claim 13, which depends on claims 11, Gazzetti further teaches wherein the set of first machine learning models collectively comprise a neural network (see Gazzetti, col.4, lines 22-28: “In some embodiments, natural language processing (NLP), Mel-frequency cepstral coefficients (MFCCs) (e.g., for audio content detected by a microphone), and/or region-based convolutional neural network (R-CNN) pixel mapping (e.g., for object detection/classification in images/videos), as are commonly understood, are used.”).
As per claim 14, which depends on claims 13, Gazzetti further teaches wherein at least one of the set of first machine learning models implements back-propagation to provide feedback or hints to others of the set of first machine learning models (see Gazzetti, col.5, lines 3-11: “Utilizing the models, the performance (or operation) of the system (e.g., utilizing/based on new inputs) may be predicted and/or the performance of the system may be optimized by investigating how changes in the input(s) effect the output(s). Feedback received from (or provided by) users and/or administrators may also be utilized, which may allow for the performance of the system to further improve with continued use.”; and col.18, lines 14-16: “In some embodiments, feedback from users may also be utilized to improve the performance of the system over time.”).
As per claim 16, which depends on claims 11, Gazzetti further teaches wherein the set of first machine learning models are configured to perform preliminary work on the request before processing at the second machine learning model (see Gazzetti, col.16, lines 4-12: “If the confidence results do not meet the threshold ( or defined QoS parameters), the request handler may forward requests to a cloud connector that causes a cloud inference service (and/or second object detection model) to be utilized to evaluate or process the images. The results of the processing with the cloud inference service may then be returned to the user. However, if the confidence results meet/exceed the threshold, the response/result from the edge object detection model is utilized and provided to the user.”).
As per claim 18, which depends on claims 11, Gazzetti further teaches wherein at least some of the first machine learning models inform each other of respective inferencing outputs (see Gazzetti, FIG. 5 & FIG. 6; and col.11, line 66-col.12, line 3: “For each model (e.g., Model A and Model B), all of the available annotations (or labels) produced by running an inference process over the entirety of a test data set are collected ( or calculated) utilizing the models.”).
As per claim 19, which depends on claims 11, Gazzetti further teaches wherein the set of first machine learning models are configured in a tiered arrangement having one or more outputs generated at a first tier are supplied as one or more inputs to a second tier (see Gazzetti, col.11, lines 25-29: “As a particular example, it should be assumed that the input data includes a single category to be identified. However, the procedure may be repeated in scenarios in which multiple classes are present (e.g., by repeating the procedure for each class).”; and col.16, lines 40-52: “If determined to be appropriate, the edge service 1604 may also interact with the cloud platform (or remote cloud platform) 1606 to request that the input data be processed with another (e.g. a second) machine learning model. As such, the cloud platform 1606 may include, for example, remote services/systems deployed in a cloud infrastructure, which are configured to process the input data with the second ( or third, fourth, etc.) machine learning model. It should be noted that because of the potentially (nearly) unlimited resource constraints of such a service ( e.g., on the cloud), the accuracy of inferences generated by such a system may exceed that of the edge service 1604 (e.g., implemented locally, on a single device, etc.).”).
As per claim 20, which depends on claims 11, Gazzetti further teaches wherein at least a first one of the set of first machine learning models have a different capability relative to a second one of the set of first machine learning models (see Gazzetti, col.5, lines 36-40: “Each of the first machine learning model and the second machine learning model may be configured to perform the same prediction task (e.g., object detection, object classification, or any other suitable machine learning process/technique).”).
As per claim 21, which depends on claims 11, Gazzetti further teaches wherein at least a first one of the set of first machine learning models is associated with additional operating logic (see Gazzetti, col., lines: “The first machine learning model may be implemented utilizing a first computing device, and the second machine learning model is implemented utilizing a second computing device.”).
As per claim 22, which depends on claims 11, Gazzetti further teaches wherein the set of first machine learning models comprise a model chain (see Gazzetti, col.14, lines 4-8: “performance that may be achieved utilizing a single model (e.g., Model A or "edge model") and a chain inference (e.g., utilizing both Model A and Model B or "chain model").”.
6. Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) and Chaudhuri et al. (US 2019/0378028 A1), and still further in view of Jha et al (US 2020/0027022 A1).
As per claim 2, which depends on claims 1, Gazzetti and Chaudhuri do not explicitly teach wherein the at least one first machine learning model executes primarily on one or more Central Processing Units (CPUs), and the second machine learning model executes primarily on one or more Graphics Processing Units (GPUs).
Jha teaches a first machine learning model executes primarily on one or more Central Processing Units (CPUs), and a second machine learning model executes primarily on one or more Graphics Processing Units (GPUs) (see Jha, [0139]: “The machine (e.g., computer system) 1400 may include a hardware processor 1402 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a hardware processor core, or any combination thereof)”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti and Chaudhuri in view of Jha so that first machine learning model executes primarily on one or more Central Processing Units (CPUs), and the second machine learning model executes primarily on one or more Graphics Processing Units (GPUs). One would be motivated to do so because Gazzetti teaches in paragraph 16, lines 35-36, “The edge service 1604 may be implemented in any suitable computing device”.
7. Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) and Chaudhuri et al. (US 2019/0378028 A1), and still further in view of Kanitkar et al. (US 11,490,307 B2).
As per claim 11, which depends on claims 1, Gazzetti and Chaudhuri do not explicitly teach wherein the one or more first machine learning models comprise a set of first machine learning models that execute in or across an overlay network operating region.
Kanitkar teaches wherein the one or more first machine learning models comprise a set of first machine learning models that execute in or across an overlay network operating region (see Kanitkar, Abstract: “Edge server compute capacity demand in an overlay network is predicted and used to pre-position compute capacity in advance of application-specific demands. Preferably, machine learning is used to proactively predict anticipated compute capacity needs for an edge server region (e.g., a set of co-located edge servers)… Overlay network mapping (globally) may be used for more long-term positioning, with short-duration scheduling then being done in-region as needed. Compute instances and associated state are migrated intelligently based on predicted (e.g., machine-learned) demand, and with full data consistency enforced).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti and Chaudhuri in view of Kanitkar so that the one or more first machine learning models comprise a set of first machine learning models that execute in or across an overlay network operating region. One would be motivated to do so because overlay networks are well-known, routine, and conventional for intelligently routing requests intelligently in edge inference servers and cloud inference endpoints such as those taught by Gazzetti.
8. Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) and Chaudhuri et al. (US 2019/0378028 A1), and still further in view of Van Vredendaal et al. (US 2020/0327443 A1).
As per claim 15, which depends on claims 11, Gazzetti and Chaudhuri do not explicitly teach wherein the set of first machine learning models are configured to protect the second machine learning model against a given execution vulnerability.
Van Vredendaal teaches wherein the set of first machine learning models are configured to protect the second machine learning model against a given execution vulnerability (see, Van Vredendaal, [0010]: “By using a plurality of additional ML models as described herein, a trained ML model is protected from extraction, or cloning, by an attacker without significantly impacting the functioning of the protected ML model for problem domain input samples. Also, the ML model does not require ML model architecture changes to provide the protection. In addition, no retraining of the ML model is required. Constructing the plurality of additional ML models to have a reduced accuracy provides a reduction in effort, time, and cost to construct the plurality of ML models.”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti and Chaudhuri in view of Van Vredendaal so that the set of first machine learning models are configured to protect the second machine learning model against a given execution vulnerability. One would be motivated to do so because modern machine learning systems are vulnerable to multiple kinds of attacks and failures during inference or execution.
9. Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Gazzetti et al. (US 11,605,028 B2) and Chaudhuri et al. (US 2019/0378028 A1), and still further in view of Huang et al. (US 2025/0111159 A1).
As per claim 17, which depends on claims 11, Gazzetti and Chaudhuri do not explicitly teach wherein at least one of the first machine learning models is associated with a Retrieval Augmented Generation (RAG) process that adds context to the request to generate an augmented request that is forwarded to the second machine learning model.
Huang teaches wherein at least one of the first machine learning models is associated with a Retrieval Augmented Generation (RAG) process that adds context to the request to generate an augmented request that is forwarded to the second machine learning model (see, Abstract: “Certain aspects of the present disclosure provide techniques retrieval augmented generation of language model responses using an embedding database. Embeddings for data is stored in an embedding database. When a prompt related to the data is received, relevant embeddings may be retrieved from the database and used generate an augmented prompt based on the initial prompt and the retrieved embeddings from the database. The augmented prompt can be input into a machine learning model.”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the system of Gazzetti and Chaudhuri in view of Huang so that at least one of the first machine learning models is associated with a Retrieval Augmented Generation (RAG) process that adds context to the request to generate an augmented request that is forwarded to the second machine learning model. One would be motivated to do so because the first model is optimized for retrieving, filtering, grounding, or enriching information, whereas the second model is optimized for reasoning, generation, decision-making, or task execution.
Conclusion
10. For the reasons above, claims 1-22 have been rejected and remain pending.
11. Any inquiry concerning this communication or earlier communications from the examiner should be directed to MICHAEL Y WON whose telephone number is (571)272-3993. The examiner can normally be reached on Wk.1: M-F: 8-5 PST & Wk.2: M-Th: 8-7 PST.
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, Nicholas R Taylor can be reached on 571-272-3889. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Michael Won/Primary Examiner, Art Unit 2443
