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 .
Response to Amendment
Applicant’s amendment filed June 10th 2026, has been entered and made of record. Claims 1, 11 and 20 are amended. Claims 1-20 are pending.
Applicant’s remarks in view of the newly presented amendments have been considered to be persuasive with regard to the newly added claim limitation:
performing, by the device and using the video data, network calibration on the machine learning model, without retraining the machine learning model, by adjusting quantization parameters of the machine learning model based on a distribution of the video data from the target environment, to form a domain-adapted model;
A new rejection is provided relying on the previously cited reference to Hussein, titled “Brining Quantization to the Transfer Learning World” to teach the added limitations.
Regarding Applicant’s arguments that the combination of Lim and Adeel, Applicant argues improper hindsight was used and that one of ordinary skill in the art would not be motivated to combine the UI model selection of Adeel with Lim. Examiner disagrees. Adeel was merely cited to teach the limitation of: “receiving, at a device and via a user interface, a selection of a machine learning model.” Examiner maintains one of ordinary skill in the art would reasonably allow for a selection via a user interface as is exceedingly well known in the art and as is clearly taught by Adeel.
The new rejection is necessitated by the amendment and is accordingly made FINAL.
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 1-7, 9, 11-17, 19 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of USPNs 2024/0119360 to Lim et al. and 2024/0143645 to Adeel et al. and further in view of the previously cited publication titled "Bringing Quantization to the Transfer Learning World" to Hussein.
With regard to claim 1, Lim discloses a method comprising:
receiving, at a device [and via a user interface, a selection of] a machine learning model trained to perform a video analytics task using a training dataset (paragraphs [0016]-[0017], Lim discloses that a machine learning model for identifying objects in streaming video is used and that the training data is from a source domain);
obtaining, by the device, video data from a target environment that is not represented in the training dataset (paragraphs [0017]-[0020], Lim discloses that additional input data is gathered from a target domain or environment different than the source domain training data);
performing, by the device and using the video data, network calibration on the machine learning model to form a domain-adapted model (paragraphs [0021]-[0027] and [0041]-[0046], Lim describes network calibration in the form of adjusting weights of the layers of the network according to the target or shifted domain in order to adapt the network model to the shifted/target domain); and
causing, by the device, the domain-adapted model to be deployed to perform the video analytics task with respect to the target environment (paragraph [0041] “…The operations 400 can be performed, for example, by a computing system, such as a user equipment (UE) or other computing device, such as that illustrated in FIG. 5, on which a pre-trained machine learning model can be adjusted prior to deployment (e.g., during or after training) or on which a machine learning model is deployed.” The domain adapted model for the domain-shifted data is deployed for performing the task in the target/shifted domain).
Lim does not explicitly teach the newly added calibration step as recited:
performing, by the device and using the video data, network calibration on the machine learning model, without retraining the machine learning model, by adjusting quantization parameters of the machine learning model based on a distribution of the video data from the target environment, to form a domain-adapted model
Hussein discloses, as the title “Bringing Quantization to the Transfer Learning World” suggests, Quantization calibration for transfer learning by adjusting specific quantization parameters, scale “S” and Integer Zero Point “Z” (pages 19-20, section 3.4 and Fig. 3.3).
“3.4 Quantization Fine Tuning Methods
In quantization, it is essential to tune the parameters in the neural networks to quantize
the model. There are mainly two categories of neural network quantization schemes.
First, we can tune the model by retraining it and applying the quantization process on
the fly while training in a process called Quantization Aware Training (QAT). Second,
we can tune the network without retraining the model by just collecting some statistics
about the model from a sample of the dataset called calibration dataset, a process that
is named Post-Training Quantization (PTQ).
3.4.1 Post Training Quantization
PTQ aims at fine-tuning the weights and the activation units’ inputs and outputs by calculating
the clipping range [α, β] through calculation of scale "S" and Integer Zero Point
"Z" without retraining the model. The underlying idea is that a calibration dataset
is passed through the pretrained unquantized model and used for the calibration and
statistics of the model parameters. Later, these statistics are used based on the scheme
of choice to quantize the model parameters (Shown in Figure 3.3).
One of the advantages of PTQ is the very low overhead of quantizing the model, which
could even be negligible in some cases. Unlike QAT, which requires the training dataset
for retraining the model (more on that in the following sub-section), PTQ is privileged
because it only needs a subset dataset from either the training or test datasets.”
Hussein clearly teaches that Post Training Quantization (PTQ) is used without the need for retraining as claimed. Hussein also teaches that Quantization Aware Training (QAT) can be done with retraining (See pages 21-22, section 3.4.2 and Fig. 3.4). Hussein explains that both methods are beneficial.
Therefore, it would have been obvious to one of ordinary skill in the art before time of filing to use Post Training Quantization without retraining as taught by Hussein in combination with domain shifted models of Lim, i.e. as an additional step after training in order to calibrate and refine the model. One of ordinary skill in the art will readily recognize that a model may be updated without retraining the model with quantization as taught by Hussein, and that a model may be retrained with quantization aware training also taught by Hussein. Both are useful in forming a domain adapted model.
Lim discloses that the method is implemented on a user device (paragraphs [0016] and [0041]), but does not explicitly teach the step of receiving a user selection of a machine learning model.
Adeel teaches an image analysis system that uses machine learning to process and further teaches “using a user interface page provided by the user application 108, the user 112 may select the single machine-learning model that is trained to detect various items to process the multimedia file” (paragraph [0017]).
Therefore, it would have been obvious to one of ordinary skill in the art before time of filing to allow a user to select a machine learning model as taught by Adeel in combination with the machine learning model use of Lim in order to allow a user selection of a machine learning model.
With regard to claim 2, Lim discloses the method as in claim 1, wherein the video analytics task comprises at least one of:
image classification (paragraph [0058]), object re-identification, or object detection (paragraphs [0003], [0016], [0053] and [0058], Lim teaches object detection and image classification).
With regard to claim 3, Lim discloses the method as in claim 1, wherein causing the domain-adapted model to be deployed to perform the video analytics task with respect to the target environment comprises:
providing the domain-adapted model to an edge device in the target environment for execution (paragraphs [0003]-[0004], [0017] and [0041], Lim discloses that the domain adapted machine learning model is deployed to a target environment and gives the example of autonomous driving, in which case the edge device would be the processor in car or an end user equipment device).
With regard to claims 4, Lim and Adeel disclose the method as in claim 1, but do not disclose wherein the network calibration de-quantizes the machine learning model to form the domain-adapted model. Hussein teaches a method for transfer learning for training a model from source domain data and then updating the model for target domain (pages 6-8 and Fig. 2.1), and further teaches that the domain adapted model is generated or fine-tuned by dequantization of the source trained model (pages 16-19, dequantization of the layers is discussed at section 3.3.2 on page 17). Therefore it would have been obvious to one of ordinary skill in the art before time of filing to use the dequantization taught by Hussein in order to retrain the layers of the target domain model as taught by Lim in order to adapt the model to the target domain.
With regard to claim 5, Lim discloses the method as in claim 1, wherein the video data from the target environment that is not represented in the training dataset depicts a feature not depicted in the training dataset (paragraphs [0004], [0017], Lim gives examples of environment conditions for the shifted domain such as urban/rural environment, weather conditions, lighting, i.e. bright versus dim, as well as blurring and noise).
With regard to claim 6, Lim discloses the method as in claim 5, wherein the feature comprises at least one of: a camera angle, a lighting condition, a cosmetic style of a type of object, or an image background (paragraphs [0004], [0017], Lim gives examples of environment conditions for the shifted domain such as urban/rural environment/background, weather conditions, lighting, i.e. bright versus dim, as well as blurring and noise.
With regard to claim 7, Lim discloses the method as in claim 1, wherein performing network calibration on the machine learning model to form the domain-adapted model comprises:
determining an amount of distribution shift between the video data from the target environment and the training dataset (paragraphs [0018], “…The magnitude of the performance reduction may be related to the magnitude of the shift between the source data set and the data input into the machine learning model at inference time. Generally, smaller domain shifts between the source data set and the data input into the machine learning model at inference time may result in better inference performance than larger domain between the source data set and the data input into the machine learning model at inference time.”).
With regard to claim 9, Hussein discloses wherein performing network calibration on the machine learning model to form the domain-adapted model comprises: computing a scaling factor and zeropoint for the network calibration based on the video data from the target environment ” (pages 19-20, section 3.4 and Fig. 3.3).
“3.4.1 Post Training Quantization
PTQ aims at fine-tuning the weights and the activation units’ inputs and outputs by calculating
the clipping range [α, β] through calculation of scale "S" and Integer Zero Point
"Z" without retraining the model. The underlying idea is that a calibration dataset
is passed through the pretrained unquantized model and used for the calibration and
statistics of the model parameters. Later, these statistics are used based on the scheme
of choice to quantize the model parameters (Shown in Figure 3.3).”
With regard to claim 11, the discussion of claim 1 applies. Lim discloses an apparatus (Fig. 5, 500) with a network interface (512, 514 wireless connectivity), processor (502 CPU) and memory (524) is configured to store a process that is executed by the processor as discussed with regard to the method of claim 1.
With regard to claims 12-17 and 19 the discussions of claims 2-7 and 9 apply respectively.
With regard to claim 20, the discussion of claims 1 and 11 apply. Lim discloses a computer program product for performing the method discussed in claim 1 (paragraph [0006]).
Claims 8 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of USPNs 2024/0119360 to Lim et al. and 2024/0143645 to Adeel et al., and "Bringing Quantization to the Transfer Learning World" to Hussein, and further in view of USPN 2021/0042643 to Hong et al.
With regard to claims 8 and 18, the combination of Lim, Adeel and Hussein discloses the method of claim 1, but do not disclose, further comprising:
computing, by the device, an accuracy of the domain-adapted model; and
providing, by the device, an indication of the accuracy of the domain-adapted model to the user interface.
Measuring the accuracy of a machine learning model is a well known endeavor in art. Hong teaches a system for adapting machine learning models and further teaches evaluating the accuracy and presenting the evaluated accuracy (paragraphs [0009], [0096]-[0100] and [0153] and Fig. 5). Therefore it would have been obvious to one of ordinary skill in the art before time of filing to compute and provide an indication of the accuracy of the domain adapted machine learning model in order to evaluate and improve the machine learning.
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of USPNs 2024/0119360 to Lim et al. and 2024/0143645 to Adeel et al., and "Bringing Quantization to the Transfer Learning World" to Hussein, and further in view of USPN 2023/0267720 to Marvasti et al.
With regard to claim 10, Lim, Adeel and Hussein disclose the method as in claim 1, but do not disclose wherein the machine learning model is a You Only Look Once (YOLO) model. Marvasti discloses an adapted network in a vehicle image processing system that uses a YOLO architecture (paragraph [0061]). Therefore, it would have been obvious to one of ordinary skill in the art before time of filing to use a YOLO model as taught by Marvasti in the vehicle operating environment taught by Lim in order to enable an effective image recognition domain adapted model.
FINAL REJECTION
Applicant’s amendment necessitated the new grounds of rejection presented in the 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 37CFR 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 extension fee 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 date of this final action.
Contact Information
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WESLEY J TUCKER whose telephone number is (571)272-7427. The examiner can normally be reached 9AM-5PM Monday-Friday.
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, JOHN VILLECCO can be reached at 571-272-7319. 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.
/WESLEY J TUCKER/Primary Examiner, Art Unit 2661
Prosecution Insights