Prosecution Insights
Last updated: November 20, 2026
Application No. 18/787,664

RECURSIVE-TEMPORAL MODELS FOR AUTONOMOUS OR SEMI-AUTONOMOUS PERCEPTION SYSTEMS AND APPLICATIONS

Final Rejection §102§103§DOUBLEPATENT
Filed
Jul 29, 2024
Examiner
VARNDELL, ROSS E
Art Unit
2674
Tech Center
2600 — Communications
Assignee
NVIDIA Corporation
OA Round
2 (Final)
84%
Grant Probability
Favorable
3-4
OA Rounds
0m
Est. Remaining
98%
With Interview

Examiner Intelligence

Grants 84% — above average
84%
Career Allowance Rate
541 granted / 640 resolved
+22.5% vs TC avg
Moderate +14% lift
Without
With
+13.6%
Interview Lift
resolved cases with interview
Typical timeline
2y 3m
Avg Prosecution
35 currently pending
Career history
687
Total Applications
across all art units

Statute-Specific Performance

§101
7.0%
-33.0% vs TC avg
§103
66.8%
+26.8% vs TC avg
§102
6.4%
-33.6% vs TC avg
§112
11.9%
-28.1% vs TC avg
Black line = Tech Center average estimate • Based on career data from 640 resolved cases

Office Action

§102 §103 §DOUBLEPATENT
DETAILED ACTION Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Status of Claims This final rejection is in response to the amendment filed September 15, 2026. Claims 1, 3-10, and 12-22 are pending and have been examined. Claims 2 and 11 are cancelled and claims 21-22 are new. Claim Objections Claim(s) 22 is objected to because the claim recites “the state data generated during the one or more first inference iterations.” However, claim 10 only recites “a first inference iteration.” There is insufficient antecedent basis for this limitation in the claim. Appropriate correction is required. Specification The title of the invention is not descriptive. A new title is required that is clearly indicative of the invention to which the claims are directed. Reverting the title to the previous title is suggested. Response to Arguments Applicant’s arguments filed September 15, 2026 have been fully considered. The objection to claim 18 and the rejection of claim 1-9 and 12 under 35 U.S.C. 112(b) are withdrawn in view of the amendments. The rejection of claim 1-16 under 35 U.S.C. 101 is withdrawn. Amended claim 1 recited the fixed parameter subset of backbone layers and stored intermediate features, and amended claim 10 recites state carried from one inference iteration to the next, which describe the improvement from the specification’s ¶4 and integrate the judicial exception into a practical application, same as previously examined claim 17 (MPEP 2106.04(d)(1), 2106.05(a)). Applicant’s arguments with respect to claim(s) 1, 9, 10, 15, 21, and 22 are considered but are moot in view of the new ground(s) of rejection in view of Li and Nguyen, as shown below. The rejection of claim 2 and 11 are moot because the claims have been cancelled. Applicant argues the rejection does not establish that Wu anticipates amended claim 17, which recites state data, “generated during one or more prior inference iterations and stored internal to the neural network.” The argument is not persuasive because Wu’s stateful mode reset hidden states only at start-up and carries a hidden state from one time instance to the next: “a sequential model may be run in a fully stateful mode where the hidden states may only be reset at start-up” (Wu ¶ 119), and “Although the DNN may contain multiple hidden states” (Wu ¶119), so that hidden state is state data generated during prior inference iterations and stored internal to the network. Applicant traverses the double patenting rejection on the grounds that the independent claims were amended. In view of the amendments, the rejection is withdrawn with respect to claim 1, 3-9, and 15 and is not applied to new claim 21. For claims 10, 12-14, 16-20, and 22, the argument is not persuasive because the applicant does not identify limitations that distinguish the instant claim from the reference claims, and the added inference-iteration and stored-state are taught by Wu, as shown below. Double Patenting The nonstatutory double patenting rejection is based on a judicially created doctrine grounded in public policy (a policy reflected in the statute) so as to prevent the unjustified or improper timewise extension of the “right to exclude” granted by a patent and to prevent possible harassment by multiple assignees. A nonstatutory double patenting rejection is appropriate where the conflicting claims are not identical, but at least one examined application claim is not patentably distinct from the reference claim(s) because the examined application claim is either anticipated by, or would have been obvious over, the reference claim(s). See, e.g., In re Berg, 140 F.3d 1428, 46 USPQ2d 1226 (Fed. Cir. 1998); In re Goodman, 11 F.3d 1046, 29 USPQ2d 2010 (Fed. Cir. 1993); In re Longi, 759 F.2d 887, 225 USPQ 645 (Fed. Cir. 1985); In re Van Ornum, 686 F.2d 937, 214 USPQ 761 (CCPA 1982); In re Vogel, 422 F.2d 438, 164 USPQ 619 (CCPA 1970); In re Thorington, 418 F.2d 528, 163 USPQ 644 (CCPA 1969). A timely filed terminal disclaimer in compliance with 37 CFR 1.321(c) or 1.321(d) may be used to overcome an actual or provisional rejection based on nonstatutory double patenting provided the reference application or patent either is shown to be commonly owned with the examined application, or claims an invention made as a result of activities undertaken within the scope of a joint research agreement. See MPEP § 717.02 for applications subject to examination under the first inventor to file provisions of the AIA as explained in MPEP § 2159. See MPEP § 2146 et seq. for applications not subject to examination under the first inventor to file provisions of the AIA . A terminal disclaimer must be signed in compliance with 37 CFR 1.321(b). The filing of a terminal disclaimer by itself is not a complete reply to a nonstatutory double patenting (NSDP) rejection. A complete reply requires that the terminal disclaimer be accompanied by a reply requesting reconsideration of the prior Office action. Even where the NSDP rejection is provisional the reply must be complete. See MPEP § 804, subsection I.B.1. For a reply to a non-final Office action, see 37 CFR 1.111(a). For a reply to final Office action, see 37 CFR 1.113(c). A request for reconsideration while not provided for in 37 CFR 1.113(c) may be filed after final for consideration. See MPEP §§ 706.07(e) and 714.13. The USPTO Internet website contains terminal disclaimer forms which may be used. Please visit www.uspto.gov/patent/patents-forms. The actual filing date of the application in which the form is filed determines what form (e.g., PTO/SB/25, PTO/SB/26, PTO/AIA /25, or PTO/AIA /26) should be used. A web-based eTerminal Disclaimer may be filled out completely online using web-screens. An eTerminal Disclaimer that meets all requirements is auto-processed and approved immediately upon submission. For more information about eTerminal Disclaimers, refer to www.uspto.gov/patents/apply/applying-online/eterminal-disclaimer. Claims 10, 12-14, 16-20, and 22 are provisionally rejected on the ground of nonstatutory double patenting as being unpatentable over the claims of copending Application No. 18/677,445 (reference application) in view of Wu et al., US 2020/0293064 A1 (“Wu”). Although the claims at issue are not identical, they are not patentably distinct from each other claim 10 of the ‘455 application recites “storing the first feature value in a cache” and “generating a perception output associated with the second sensor data frame based at least on the first feature data retrieved from the cache and the second feature data,” and claim 16 adds that “the temporal model includes the neural network as a backbone and one or more additional layers for temporal context”; claims 18 and 14 supply “a control system for an autonomous or semi-autonomous machine” and an output including “path geometry, a path class, a path uncertainty, or one or more path attributes.” The difference is that the claims at issue, as amended, carry temporal context forward from a prior inference iteration to a subsequent inference iteration as state data stored internal to the network instead of as cached feature data, but Wu discloses both approaches. Wu teaches a buffer “used to store each of the three previous feature maps” (Wu ¶ 59) and a stateful mode where “a sequential model may be run in a fully stateful mode where the hidden states may only be reset at start-up” (Wu ¶ 119) and where “the sequential DNN may be trained to maintain, or store, a hidden state across sequences of frames without resetting the hidden state” (Wu ¶ 120). It would have been obvious to carry the ‘455 temporal context forward as stored recurrent state, avoiding per-frame computation, because Wu teaches a stateful single pass mode, head layers that generate the prediction (claims 10,13,18, and 22) and GRU sequential layers (claims 12, 19, and 22) , while claim 14’s earlier-image features correspond to the first execution of the ‘445 backbone (‘445 claims 10 and 16) and claims 16 and 20 recite the system list of ‘445 claim 18 (Wu ¶¶ 40, 69, 119). This is a provisional nonstatutory double patenting rejection because the patentably indistinct claims have not in fact been patented. Claim Rejections - 35 USC § 102 The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action: A person shall be entitled to a patent unless – (a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention. Claim(s) 17, 19, and 20 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Wu et al., US 2020/0293064 A1 (hereinafter “Wu”). Claim 17. Wu discloses one or more processors (FIG. 14C) comprising: processing circuitry to cause performance of one or more operations of a machine based at least on one or more outputs of a neural network, the one or more outputs computed based at least on the neural network processing, during a subsequent inference iteration, an instance of sensor data obtained using one or more sensors of the machine along with state data generated during one or more prior inference iteration and stored internal to the neural network, the state data computed using one or more temporal layers of the neural network and based at least on the neural network processing a plurality of instances of sensor data prior to the instance of sensor data (Wu: “The image data 102 may include image data generated by one or more cameras of an autonomous vehicle” (¶ 33), which is the instance of sensor data; “a sequential model may be run in a fully stateful mode where the hidden states may only be reset at start-up” (Wu ¶ 119); “where y is a prediction, x is the input (e.g., an image), h denotes a hidden state, t is a time index” (Wu ¶ 119), so the hidden state from the prior pass is processed with x(t) and is generated during prior inference iterations and stored internal to the network; “Although the DNN may contain multiple hidden states” (Wu ¶ 119); because the hidden state is reset only at start-up, h(t-1) is computed from every image processed since start-up, which is the plurality of prior instances of sensor data; “stateful training may be used for the sequential DNN 104 to account for, and remedy, this contrast between training and inference” (Wu ¶120) ties the stateful mode of ¶ 119 to sequential DNN 104; “The sequential DNN(s) 104 may include a recurrent neural network (RNN), a gated recurrent unit (GRU) DNN, a long short term memory (LSTM) DNN” (Wu ¶ 40), which are temporal layers; “world model management, obstacle avoidance, path planning, etc.” (Wu ¶ 34), which are operations.). Claim 19. Wu discloses the one or more processors of claim 17, wherein the one or more temporal layers include one or more gated recurrent unit (GRU) layers, one or more long short-term memory (LSTM) layers, one or more recursive neural network layers, or one or more recurrent neural network (RNN) layers (Wu: "a recurrent neural network (RNN), which may have one or more long short term memory (LSTM) layers and/or gated recurrent unit (GRU) layers" (¶0029).). Claim 20. Wu discloses the one or more processor of claim 17, wherein the processor is comprised in at least one of: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine (Wu is directed to "autonomous machine applications" and "perform[s] one or more operations by the vehicle" ( claim 1; ¶0027); it is the perception and control system of an autonomous vehicle.); a system for performing one or more simulation operations; a system for performing one or more digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing one or more deep learning operations; a system implemented using an edge device; a system implemented using a robot; a system for performing one or more generative AI operations; a system for performing operations using one or more large language models (LLMs); a system for performing operations using one or more vision language models (VLMs); a system for performing operations using one or more multi-modal language models (VLMs); a system for performing one or more conversational AI operations; a system for generating synthetic data; a system for presenting at least one of virtual reality content, augmented reality content, or mixed reality content; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources. Claim Rejections - 35 USC § 103 The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows: 1. Determining the scope and contents of the prior art. 2. Ascertaining the differences between the prior art and the claims at issue. 3. Resolving the level of ordinary skill in the pertinent art. 4. Considering objective evidence present in the application indicating obviousness or nonobviousness. This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention. Claims 10, 12-14, 16, 18, and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Wu. Claim 10. Wu discloses a system comprising: one or more processors (FIG. 14C) to: during a first inference iteration, generate, using a machine learning model and based at least on one or more first images, state data representative of a recursive combination of one or more first features (Wu: a “computed hidden state of the first set of sequential layers 608A ... may be passed to the second set of sequential layers 608B, a computed hidden state of the second set of sequential layers 608B (e.g., using ... the hidden state from the first set of sequential layers 608A) may be passed to the third set” (¶0069); the hidden state recursively combines features from the first images; in the stateful mode, Wu teaches that “a sequential model may be run in a fully stateful mode where the hidden states may only be reset at start-up” (¶ 119), so the hidden state at a first inference iteration combines the features of earlier images.); during a subsequent inference iteration, generate, using one or more backbone layers of the machine learning model and based at least on a second image, one or more second features corresponding to the second image (Wu: the image 602 captured at time T is the second image and feature extractor 604 is the backbone, since the image “may be input to a feature extractor 604 to generate a feature map” (¶ 68).); generate, during the subsequent inference iteration and using one or more recursive layers and one or more head layers of the machine learning model, one or more outputs based at least on updating the state data generated during the inference iteration using at least a portion of the one or more second features (Wu: the sequential layers 608 are the recursive layers and the additional layer(s) 610 are the head layers; a feature map “may be fed into each set of sequential layers 608 ( e.g., GRU, LSTM, and/or other sequential layer types)” (¶ 68), the results “may be provided as input to additional layer(s) 610” (¶ 69), and in the stateful mode each pass updates the prior hidden state h(t-1) with the current input, as in Wu’s description “where y is a prediction, x is the input ( e.g., an image), h denotes a hidden state, t is a time index” (¶ 119).); and cause a machine to perform one or more planning or control operations based at least on the one or more outputs (Wu: "performing one or more operations by the vehicle based at least in part on at least one of the first output or the second output" (claim 1); the operations include “world model management, obstacle avoidance, path planning, etc.” (¶ 34).). Wu teaches the backbone, recursive, and head layers of the FIG. 6A architecture: the image data “may be input to a feature extractor 604 to generate a feature map” (Wu ¶ 68), the feature map “may be fed into each set of sequential layers 608 (e.g., GRU, LSTM, and/or other sequential layer types” (Wu ¶ 68), and the sequential output “may be provided as input to additional layer(s) 610” and “to generate a prediction 612” (Wu ¶ 69). Wu separately teaches stateful inference in which “a sequential model may be run in a fully stateful mode where the hidden states may only be reset at start-up” (Wu ¶¶ 119). It would have been obvious to one of ordinary skill in the art before the effective filling data to run Wu’s FIG. 6A network in the stateful mode of ¶ 119 to cut the per-frame computation from “N hidden states may be computed for N different length sequences of images in parallel at every time instant” (Wu ¶ 71) to a “a single forward pass computation may be required per time instance” (Wu ¶ 119). Wu states that, “depending on the implementation of the sequential DNN, the present system implements stateless or stateful training and inference methods” (Wu ¶ 6) and teaches that “stateful training may be used for the sequential DNN 104 to account for, and remedy, this contrast between training and inference” (Wu ¶ 120), so there was a reasonable expectation of success. KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398 (2007). Claim 12. Wu discloses the system of claim 10, the one or more processors further to update a state associated with a Gated Recurrent Unit (GRU) of the machine learning model based at least on a previous state associated with the GRU and the portion of the one or more second features, wherein the updating of the state data comprises updating the state from the previous state to a current state (The hidden state of one set of sequential layers is recomputed using the prior set's hidden state and the current feature map – Wu: the hidden state of the second set of sequential layers 608B is computed "using the first feature map ... and the hidden state from the first set of sequential layers 608A", and "may be passed to the third set of sequential layers 608C, and so on" ¶0069); the sequential layers are "GRU, LSTM" layers (¶0058). The prior set's hidden state is the previous state and the recomputed hidden state is the current state.). Claim 13. The system of claim 10, wherein the one or more outputs are generated using one or more head layers of the machine learning model to process an updated version of the state data, the one or more outputs indicating one or more predictions associated with an environment in which the machine is operating (Wu: "[t]he output of the last set of sequential layers 608 ... may be provided as input to additional layer(s) 610 of the DNN ... to generate a prediction 612" (¶0069); the prediction concerns objects in the vehicle's environment, including a velocity and a current and future location of an object (claim 1). The additional layer(s) 610 process the sequential layers' hidden state to produce the prediction.). Claim 14. Wu discloses the system of claim 10, wherein the one or more first images correspond to a temporal series of images associated with one or more previous instances of time that precede a first instance of time associate with the second image (Wu: "an image captured at times previous to (or after, in some embodiments) the image 502" (¶0058); the "previous to" embodiment is the recited preceding instances of time.), the temporal series of images processed using one or more backbone layers of the machine learning model to generate the one or more first features (Wu: "an image 502 ... may be input to a feature extractor 504 to generate a feature map F 508A" (¶0058), the feature extractor processing each image of the sequence.). Claim 16. Wu discloses the system of claim 10, wherein the system is comprised in at least one of: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine (Wu is directed to "autonomous machine applications" and "perform[s] one or more operations by the vehicle" ( claim 1; ¶0027); it is the perception and control system of an autonomous vehicle.); a system for performing one or more simulation operations; a system for performing one or more digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing one or more deep learning operations; a system implemented using an edge device; a system implemented using a robot; a system for performing one or more generative AI operations; a system for performing operations using one or more large language models (LLMs); a system for performing operations using one or more vision language models (VLMs); a system for performing operations using one or more multi-modal language models (VLMs); a system for performing one or more conversational AI operations; a system for generating synthetic data; a system for presenting at least one of virtual reality content, augmented reality content, or mixed reality content; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources. Claim 18. Wu discloses the one or more processors of claim 17, wherein an intermediate representation of the instance of the sensor data along with the state data is processed using one or more head layers of the neural network to compute the one or more outputs (Wu: the feature extractor generates a feature map for image 502 (¶0058), and "[t]he output of the last set of sequential layers 608 ... may be provided as input to additional layer(s) 610 of the DNN to generate a prediction 612" (¶0069).). Claim 18 is rejection of Wu for the reasons given in claim 10, since the intermediate representation and head layers are shown in FIG. 6A (feature map from the feature extractor 604 an additional layer(s) 610) run in the stateful mode of ¶ 119. Claim 22. (New) Wu discloses the system of claim 10, wherein the one or more recursive layers comprise a Gated Recurrent Unit (GRU) (Wu: “The feature map FT 606A corresponding to the image 602 may be fed into each set of sequential layers 608 ( e.g., GRU …” (¶ 68); ), and, during the subsequent inference iteration, inputs to the GRU include at least the state data generated during the one or more first inference iterations and the one or more second features corresponding to the second image, the GRU updates the state data based on the inputs (Wu: “where y is a prediction, x is the input ( e.g., an image), h denotes a hidden state, t is a time index” (¶ 119). Wu’s inputs x(t), the features of the current image, and h(t-1), the state data from the prior inference iterations, and the outputs include the updated hidden state h(t), so the GRU updates the state data based on the inputs.), and the one or more head layers generate the one or more outputs based at least on the updated state data (Wu: the updated hidden state “may be provided as input to additional layer(s) 610 of the DNN … to generate a prediction 612” (¶ 69); the head layers are the additional layer(s) 610). The combination is obvious for the reasons given in claim 10. Claims 1 and 3-8 are rejected under 35 U.S.C. 103 as being unpatentable over Wu in view of Li et al., “Temporal Modeling Approaches for Large-scale Youtube-8M Video Understanding” (hereinafter “Li”) and Nguyen, “Scene Understanding for Autonomous Manipulation with Deep Learning” (hereinafter “Nguyen”). Claim 1. Wu discloses a method comprising: causing a machine to perform one or more planning or control operations to navigate within one or more outputs of a machine learning model, the one or more outputs including one or more first predictions associated with the environment (Wu: "performing one or more operations by the vehicle based at least in part on at least one of the first output or the second output" (claim 1); the operations include “world model management, obstacle avoidance, path planning, etc.” (¶ 34); the outputs are computed by a sequential deep neural network and correspond to predictions about objects in the vehicle's environment.), wherein the machine learning model is trained, at least, by: applying, to one or more first layers of the machine learning model, training data including at least a temporal series of stored intermediate features generated over respective iterations by applying respective images of a temporal series of images to a subset of backbone layers having fixed parameters (Wu: "The feature map FT ... may be fed into sequential layer(s) 510 (e.g., GRU, LSTM ... ) of a DNN … In addition to the feature map FT 508A from the image 502, feature maps FT-1 to FT-(N-1) (508B-508D) may be fed into the sequential layer(s) 510" (¶0058); the feature maps correspond to a sequence of images captured over time. The sequential DNN "may include a recurrent neural network (RNN), which may have one or more long short term memory (LSTM) layers and/or gated recurrent unit (GRU) layers" (¶0029), and is trained (¶0029; FIG. 11).); applying, to one or more second layers of the machine learning model, state data generated using the one or more first layers, the state data representative of a combination of at least the temporal series of stored intermediate features and one or more additional features corresponding to a current image (Wu: the sequential layers 608 compute a "hidden state" combining the current image's feature map FT with the prior feature maps FT-1 to FT-(N-1) (¶¶ 0058, 0069), and "[t]he output of the last set of sequential layers 608 ... may be provided as input to additional Iayer(s) 610 of the DNN ... to generate a prediction 612" (¶0069). The additional Iayer(s) 610 are the second layers, to which the hidden-state state data generated by the sequential layers is applied.); obtaining one or more second predictions using the one or more second layers to process at least a portion of the state data (Wu: the additional layer(s) 610 process the hidden state "to generate a prediction 612" (¶0069); the predictions include a velocity output and a current/future location output (claim 1).); and updating one or more parameters of the machine learning model based at least on an evaluation of the one or more second predictions with respect to ground truth data associated with at least the current image (Wu: "USE THE FIRST GROUND TRUTH DATA AND THE SECOND GROUND TRUTH DAT A TO TRAIN A NEURAL NETWORK" (FIG. 11); the sequential DNN "is trained using ground truth data generated by correlating training sensor data with training image data representative of a training sequences of images" (claim 5). Training adjusts the network's parameters based on the ground truth for the images.). Wu discloses training its sequential DNN and storing previously computed feature maps so they do not need to be generated again, as Wu teaches, “such that the feature maps do not need to be generated again at each iteration of the DNN” (Wu ¶ 59), but Wu does not state that the temporal series of features applied to the recursive layers in training is a series of stored intermediate features generated by applying the images to a subset of backbone layers having fixed parameters. Li, in the field of video recognition using RNNs, teaches that recurrent sequence models are trained on stored per-frame features generated by an image CNN whose parameters are fixed (Li p. 1: “only provides the pre-extracted visual and audio features from every second of video instead of raw videos. We can neither train different CNNs architectures”; p. 4: “Visual features are obtained by the Google Inception CNN pre-trained on the ImageNet”; p. 1: “Our two stream sequence models build upon the bidirectional LSTM”). The pre-extracted features are the stored intermediate features, and the pretrained CNN that cannot be trained is the backbone having fixed features. Nguyen teaches that each video frame’s visual feature is first extracted offline from an internal layer of a pretrained image-classification network and fed into the firs recurrent layer (Nguyen p. 44: “the visual feature from each video frame is first extracted offline then fed into the captioning module”; p. 44: “for the VGG16 network, the features are extracted from its last fully connected fc2 layer … we use the features from pool5 layer of the ResNet50 network”; p. 45: “These features are encoded in the first RNN layer to create the encoder hidden state”). The layers up to fc2 or pool5 are a subset of the backbone of the single-frame network. It would have been obvious to one of ordinary skill in the art before the effective filling date to train Wu’s recursive layers on a temporal series of stored intermediate features generated by a fixed subset of backbone layers, as taught by Li and Nguyen, to avoid recomputing the backbone features for every frame at every training pass, because Nguyen explains that offline extraction is necessary when many frames are processed (Nguyen p. 44: “This step is necessary since we usually have a lot of frames from the input videos”) and Wu teaches storing feature maps “such that the feature maps do not need to be generated again at each iteration of the DNN” (Wu ¶ 59). There was a reasonable expectation of success because Wu’s sequential layers already receive per-image feature maps from its feature extractor (Wu ¶ 58), so training them on stored feature maps changes only where the feature come from. KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398 (2007). Claim 3. Wu in view of Li and Nguyen discloses the method of claim 1, wherein the current image corresponds to a first instance of time and the temporal series of images correspond to one or more second instances of time that precede the first instance of time (Wu: "an image captured at times previous to (or after, in some embodiments) the image 502" (¶0058); the "previous to" embodiment is the recited preceding instances of time.). Claim 4. Wu in view of Li and Nguyen discloses the method of claim 1, wherein the one or more first layers correspond to one or more recursive layers, the one or more recursive layers to recursively combine the temporal series of stored intermediate features and the one or more additional features to update one or more state vectors associated with the one or more recursive layers, the state data including the one or more state vectors (Wu: the "sets of sequential layer(s) 608 may generate updated values of hidden states ... using a number of feature maps 606 from a number of images in addition to using a hidden state from a previous set of sequential layers 608" (¶0068); the updated hidden state is the recited state vector.). Claim 5. Wu in view of Li and Nguyen discloses the method of claim 4, wherein the one or more recursive layers correspond to one or more Gated Recurrent Units (GRUs) disposed between one or more backbone layers and one or more head layers of the machine learning model (The sequential layers are Wu’s "GRU, LSTM" layers (¶0058), they receive the feature maps from the feature extractor 504/604 (the backbone), and their output is provided to a downstream stage – Wu: "[t]he output of the last set of sequential layers 608 ... may be provided as input to additional layer(s) 610 of the DNN ... to generate a prediction 612" (¶0069). The GRU sequential layers are therefore disposed between the feature extractor backbone and the additional layer(s) 610 (the head layers).). Claim 6. Wu in view of Li and Nguyen discloses the method of claim 1, wherein the one or more second layers correspond to one or more head layers of the machine learning model, the one or more head layers to generate the one or more second predictions using the portion of the state data (Wu: "[t]he output of the last set of sequential layers 608 ... may be provided as input to additional layer(s) 610 of the DNN ... to generate a prediction 612" (¶0069); the prediction concerns objects in the vehicle's environment, including a velocity and a current and future location of an object (claim 1). The additional layer(s) 610 process the sequential layers' hidden state to produce the prediction.). Claim 7. Wu in view of Li and Nguyen discloses the method of claim 1, wherein the machine learning model is trained, at least, by further: determining a previous state associated with the one or more first layers based at least on a first combination of one or more first features of the temporal series of stored intermediate features with one or more second features of the temporal series of stored intermediate features; and determining a current state associated with the one or more first layers based at least on a second combination of the previous state with the one or more additional features, wherein the state data corresponds to the current state associated with the one or more first layers (The hidden state of one set of sequential layers is recomputed using the prior set's hidden state and the current feature map – Wu: the hidden state of the second set of sequential layers 608B is computed "using the first feature map ... and the hidden state from the first set of sequential layers 608A", and "may be passed to the third set of sequential layers 608C, and so on" ¶0069); the sequential layers are "GRU, LSTM" layers (¶0058). The prior set's hidden state is the previous state and the recomputed hidden state is the current state.). Claim 8. Wu in view of Li and Nguyen discloses the method of claim 1, wherein the updating of the one or more parameters of the machine learning model comprises at least one of: updating one or more first parameters associated with the one or more first layers of the machine learning model; or updating one or more second parameters associated with the one or more second layers of the machine learning model (Wu: the ground truth data is used "to train a neural network" (FIG. 11).). Claims 9 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Wu further in view of Gong et al., US 10,885,345 B2 (hereinafter “Gong”), and in view of Li and Nguyen (for claim 9). Claims 9 and 15. For claim 9, Wu in view of Li and Nguyen teaches the method of claim 1 and for claim 15 Wu teaches the system of claim 10, as described above. Wu discloses training the sequential DNN, including a feature extractor (backbone) and recurrent (sequential) layers (¶¶0029, 0058; FIG. 11), but does not state that the parameters of a subset of the backbone layers are Gong, in the field of training a neural network that combines a convolutional feature extractor with a recurrent (LSTM) stage on a temporal series of video frames, teaches fixing the parameters of the convolutional layers while the recurrent layers are trained (Gong: "weights of the CNNs are frozen during the initializing of the LSTM units and the training of the LSTM units using the reinforcement learning" (claim 12); see also "weights of the CNNs are frozen during the initializing of the LSTM units and the reinforcement learning" (C10:L15-20). The CNNs are the backbone layers whose weights are fixed, and the LSTM units are the recursive layers that are trained while the backbone is fixed.). It would have been obvious to one of ordinary skill in the art before the effective filing date to fix the parameters of a subset of Wu's feature-extractor (backbone) layers and, subsequent to the fixing, apply the temporal series of images to those fixed layers to generate the feature maps that are applied to the recursive layers, as taught by Gong. The motivation is to reduce the cost and memory required to train the temporal portion of the network by reusing the learned feature representation rather than retraining the backbone, which is the use of a known training technique to improve a similar network in the same way, with a reasonable expectation of success. KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007); MPEP § 2143. Claim 21 is rejected under 35 U.S.C. 103 as being unpatentable over Wu in view of Li and Nguyen, and further in view of He et al., US 10,713,794 B1 (hereinafter "He"). Claim 21. (New) Wu in view of Li and Nguyen teaches the method of claim 1, further comprising: extracting the stored intermediate features from one or more first backbone layers of a single-frame version of the machine learning model (Nguyen p. 44: “for the VGG16 network, the features are extracted from its last fully connected fc2 layer”); and (Nguyen p. 44: “we use the features from pool5 layer of the ResNet50 network”; the layers up to fc2 or pool5 are the first backbone layers of the single-frame network.). Wu does not state that “applying the stored intermediate features to one or more second backbone layers of the single-frame version of the machine learning model subsequent to the extracting.” He, in the field of single-frame convolutional image analysis backbones, teaches applying the stored intermediate features to one or more second backbone layers of the single-frame version of the machine learning model subsequent to the extracting (He: a Feature Pyramid Network “uses a top-down architecture with lateral connections to build an in-network feature pyramid from a single-scale input. Faster R-CNN with an FPN backbone extracts RoI features from different levels of the feature pyramid according to their scale” (col. 34-35); the features taken from the earlier, intermediate stages of the backbone (the first backbone layers) are combined through the lateral connections with the later, top-down stages (the second backbone layers).). It would have been obvious to one of ordinary skill in the art before the effective filing date to generate Wu's per-image feature maps by extracting intermediate features from different levels of a single-frame backbone, as taught by He, and to apply those intermediate features to later backbone layers. The motivation is He's teaching that extracting features from multiple levels of the single-frame backbone gives (He: “excellent gains in both accuracy and speed” (col. 35).). This is the use of a known feature-extraction technique to improve a similar image-analysis backbone in the same way, with a reasonable expectation of success. KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007); MPEP § 2143. Conclusion Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to Ross Varndell whose telephone number is (571)270-1922. The examiner can normally be reached M-F, 9-5 EST. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, O’Neal Mistry can be reached at (313)446-4912. 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 https://ppair-my.uspto.gov/pair/PrivatePair. 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. /Ross Varndell/Primary Examiner, Art Unit 2674
Read full office action

Prosecution Timeline

Jul 29, 2024
Application Filed
Jun 17, 2026
Non-Final Rejection mailed — §102, §103, §DOUBLEPATENT
Sep 07, 2026
Interview Requested
Sep 15, 2026
Examiner Interview Summary
Sep 15, 2026
Response Filed
Sep 15, 2026
Applicant Interview (Telephonic)
Oct 26, 2026
Final Rejection mailed — §102, §103, §DOUBLEPATENT (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12762851
LIDAR-CAMERA SPATIO-TEMPORAL ALIGNMENT USING NEURAL RADIANCE FIELDS
3y 5m to grant Granted Oct 13, 2026
Patent 12755511
IMAGE ANALYZER
2y 4m to grant Granted Oct 06, 2026
Patent 12749288
EXPLOITING HIERARCHICAL STRUCTURE LEARNING WITH HYPERBOLIC DISTANCE TO ENHANCE OPEN WORLD OBJECT DETECTION
2y 12m to grant Granted Sep 29, 2026
Patent 12749290
STORAGE MEDIUM, DATA GENERATION METHOD, AND INFORMATION PROCESSING DEVICE
2y 11m to grant Granted Sep 29, 2026
Patent 12731395
PROCESSING METHOD, AND PROCESSING SYSTEM
2y 8m to grant Granted Sep 08, 2026
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

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

Prosecution Projections

3-4
Expected OA Rounds
84%
Grant Probability
98%
With Interview (+13.6%)
2y 3m (~0m remaining)
Median Time to Grant
Moderate
PTA Risk
Based on 640 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

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

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

Free tier: 3 strategy analyses per month