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 .
Preliminary Amendment
The preliminary amendment filed on April 2, 2024, is acknowledged. Claims 1-20 of the preliminary amendment are currently pending and under consideration.
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.
1. Claims 1-4, 6-11, 13-18, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Gidaris et al., “OBoW: Online Bag-of-Visual-Words Generation for Self-Supervised Learning,” CVPR 2021, June 19-25, 2021 (“Gidaris”) in view of Chen et al. (US 2021/0327029 A1) (“Chen”).
As to claim 1, Gidaris teaches a model training method, wherein the method comprises:
obtaining a first data sequence and a perturbed first data sequence; [§ 3, paragraph 3: “To learn image representations, the student gets as input a perturbed version of the image x, denoted as
x
~
, and is trained to reconstruct the BoW representation yT(x), produced by the teacher, of the original unperturbed image x.” As shown in FIG. 1, there are perturbed and unperturbed versions of an image used as input.]
processing the perturbed first data sequence by using a first to-be-trained model to obtain a first feature sequence, and processing the first data sequence by using a second to-be-trained model to obtain a second feature sequence; [§ 3, paragraph 3: “To learn image representations, the student gets as input a perturbed version of the image x, denoted as
x
~
, and is trained to reconstruct the BoW representation yT(x), produced by the teacher, of the original unperturbed image x.” See FIG. 1 and § 3, paragraph 3: “The student S(·) is parameterized by θS and the teacher T(·) by θT.” As shown in FIG. 1, the perturbed input
x
~
is processed by Student S (corresponding to a first to-be-trained model) and unperturbed input is processed by Teacher T (corresponding to a second to-be-trained model). The models generate various outputs, including the feature sequences Tl(x) from the teacher and S(
x
~
) from the student. Additionally, yT(x) and yS(
x
~
) as shown in FIG. 1 can also be regarding as the first and second sequence.]
training the first to-be-trained model and the second to-be-trained model based on the first feature sequence and the second feature sequence to obtain a first target model and a second target model, [The models are trained as disclosed in § 3.1, which refers to “we propose a fully online training methodology that allows the teacher to be continuously updated as the student training progresses, with no need for off-line k-means stages.” That is, both the student and teacher models are trained, such that the trained models (or later stages of trained models) corresponds to a first target model and a second target model.] wherein the first target model is obtained by training the first to-be-trained model based on the first feature sequence and the second feature sequence, and the second target model is obtained based on a parameter in a training process of the first to-be-trained model; [The student model (first to-be-trained/target model) is trained using a loss function as shown in equation (2), which is based on the features S(
x
~
), since S(
x
~
) is used to compute yS(
x
~
) as shown in equation (1). Furthermore, the teacher model (second to-be-trained/target model) is updated based on the trained weights of the student model as described in § 3.1, paragraph 3: Updating the teacher network. Inspired by MoCo [33], the parameters θT of the teacher convnet are an exponential moving average of the student parameters. Specifically, at each training iteration the parameters θT are updated as:
θ
T
←
α
⋅
θ
T
+
1
-
α
⋅
θ
S
[labeled as equation (3)].”] […].
Gidaris does not explicitly teach “fine-tuning the first target model or the second target model to obtain a third target model, wherein the third target model is used to obtain a label of a data sequence.”
Chen teaches “fine-tuning the first target model or the second target model to obtain a third target model, wherein the third target model is used to obtain a label of a data sequence.” [[0046]: “The task specific model 250 and/or the base encoder neural network 204 can be additionally trained (e.g., “fine-tuned”) on additional training data (e.g., which may be task specific data). The additional training can be, for example, supervised learning training.” That is, the base encoder neural network 204 is analogous to the first/second target models of the instant claim (as shown in FIG. 2A, the base encoder neural network 2024 was trained using different perturbations of the data in contrastive learning). This fine tuning is performed to obtain a fine-tuned third model, which is the combination of the base encoder neural network and the task specific model, as shown in FIG. 2B. FIG. 2B also shows that this third model is used to generate task-specific predictions (a label) of additional input (data sequence). Note that the additional input in this case includes an image (because the base encoder is configured to receive an image as an input) and a classification of it (see [0047]: “As examples, the task-specific prediction 256 can be a classification prediction; a detection prediction; a recognition prediction; a segmentation prediction; and/or other prediction tasks.”).]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Gidaris with the teachings of Chen by implementing a further fine-tuning step as taught in Chen, so as to arrive at the claimed invention. The motivation for doing so would have been to utilize the trained model for a task specific model, as suggested by Chen (see parts cited above and [0045]: “In particular, the base encoder neural network 204 has been extracted and an additional task specific model 250 has been appended to the base encoder neural network 204. For example, the task specific model 250 can be any kind of model including linear models or non-linear models such as neural networks.”).
As to claim 2, the combination of Gidaris and Chen teaches the method according to claim 1, wherein the training the first to-be-trained model and the second to-be-trained model based on the first feature sequence and the second feature sequence to obtain a first target model and a second target model comprises:
obtaining a first loss based on the first feature sequence and the second feature sequence, wherein the first loss indicates a difference between the first feature sequence and the second feature sequence; [Gidaris, § 3, paragraph 4 teaches the cross entropy loss between the yT(x) and yS(
x
~
) which are a function of the feature Tl(x) and S(
x
~
) from the student and teacher, respectively, and is thus a loss that indicates a difference of those feature outputs. Note that although yT(x) is a feature map and yS(
x
~
) is just a feature vector, they are nonetheless compared with each other indirectly through the loss function. This is because the student and teacher have the same architecture (see text below equation (3): “teacher has to share exactly the same architecture as the student”) and the feature map includes the crop of the input into the student model, the yT(x) and yS(
x
~
) outputs are both indications of the predicted softmax probabilities for the loss function (see text above equation (2)). Alternatively, if yT(x) and yS(
x
~
) are regarded as the feature outputs of the models, then the loss described above indicates the difference between yT(x) and yS(
x
~
).]
updating a parameter of the first to-be-trained model based on the first loss, and updating a parameter of the second to-be-trained model based on the updated parameter of the first to-be-trained model until a model training condition is met, to obtain the first target model and the second target model. [In GIdaris, the student model (first to-be-trained/target model) is trained using the loss function as shown in equation (2). Furthermore, the teacher model (second to-be-trained/target model) is updated based on the trained weights of the student model as described in § 3.1, paragraph 3: Updating the teacher network. Inspired by MoCo [33], the parameters θT of the teacher convnet are an exponential moving average of the student parameters. Specifically, at each training iteration the parameters θT are updated as: [see equation (3)].”]
As to claim 3, the combination of Gidaris and Chen teaches the method according to claim 2, wherein the updated parameter of the second to-be-trained model is determined based on the updated parameter of the first to-be-trained model, the parameter of the second to-be-trained model, and a preset weight. [As noted in the rejection of the parent independent claim, equation (3) of Gidaris teaches the weight update formula of
θ
T
←
α
⋅
θ
T
+
1
-
α
⋅
θ
S
for the teacher model, where
θ
T
is the updated parameter of the teacher model (second to-be-trained model) and
θ
S
is the parameter of the student model (first to-be-trained model). Note that the student model is trained in an online manner, as described in § 3.1, paragraph 2: “a fully online training methodology that allows the teacher to be continuously updated as the student training progresses.” Furthermore,
α
corresponds to a preset weight, and is exemplified as 0.99 in the text in Gidaris below equation (3): “With a proper tuning of α, e.g., α = 0.99, this update rule allows slow and continuous updates of the teacher.”]
As to claim 4, the combination of Gidaris and Chen teaches the method according to claim 1, as set forth above.
Chen further teaches wherein the fine-tuning the first target model or the second target model to obtain a third target model comprises:
obtaining a second data sequence; [[0046]: “The task specific model 250 and/or the base encoder neural network 204 can be additionally trained (e.g., “fine-tuned”) on additional training data (e.g., which may be task specific data). The additional training can be, for example, supervised learning training.”]
fusing one of the first target model and the second target model with a preset model to obtain a third to-be-trained model; [[0045]: “FIG. 2B depicts a graphical diagram of an example use of a base encoder neural network trained after it has been trained in the example framework shown in FIG. 2A. In particular, the base encoder neural network 204 has been extracted and an additional task specific model 250 has been appended to the base encoder neural network 204. For example, the task specific model 250 can be any kind of model including linear models or non-linear models such as neural networks”]
processing the second data sequence by using the third to-be-trained model, to obtain a predicted label of the second data sequence; [FIG. 2B teaches that the combined model (the third to-be-trained model) outputs a task-specific prediction. Furthermore, [0047] teaches a “classification prediction,” which corresponds to a predicted label. Furthermore, the second data sequence is taught in the form of images. See [0083]: “Transfer learning performance was also evaluated across 12 natural image datasets in both linear evaluation (fixed feature extractor) and fine-tuning settings” (also referring to “image classification”).] and
training the third to-be-trained model based on a real label and the predicted label of the second data sequence, to obtain the third target model. [[0046]: “The additional training can be, for example, supervised learning training.” Given the context discussed above, supervised learning when applied to image classification implies the use of a real label. See also [0031]: “Further, when fine-tuned on only 1% of the labels, example implementations of the proposed techniques achieve 85.8% top-5 accuracy, outperforming AlexNet with 100 fewer labels”; [0082]: “The whole base network can be finetined on the labeled data without regularization” (where “finetined” is understood as “fine-tuned”).]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Gidaris and Chen, including the teachings of Chen cited above, so as to also arrive at the claimed invention of the instant dependent claim. The motivation for doing so is covered by the motivation given for Chen in the rejection of the parent independent claim.
As to claim 6, the combination of Gidaris and Chen teaches the method according to claim 1, as set forth above.
Chen further teaches “wherein the perturbation comprises at least one of adding noise, adding reverberation, or adding a mask in a time-frequency domain.” [[0056]: “FIG. 4 visualizes the augmentations were considered and can optionally be included in implementations of the present disclosure, which include the following examples visualized relative to the original image: crop and resize; crop, resize (and flip); color distortion (drop); color distortion (jitter); rotate; cutout; Gaussian noise; Gaussian blur; and Sobel filtering.” That is, the alternative of “adding noise” is taught.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Gidaris and Chen, including the teachings of Chen cited above by modifying the perturbation to further include added noise, so as to also arrive at the claimed invention of the instant dependent claim. The motivation for doing so would have been to use multiple forms of augmentation, including noise, to define a contrastive prediction task that yields effective representations, as suggested by Chen (see [0028]: “One example aspect of the present disclosure is directed to particular compositions of data augmentations which enable the system to define effective predictive tasks. Composition of multiple data augmentation operations is crucial in defining the contrastive prediction tasks that yield effective representations.”).
As to claim 7, the combination of Gidaris and Chen teaches the method according to claim 4, as set forth above.
Chen further teaches “wherein the first data sequence and the second data sequence are speech sequences, and the third target model is used to obtain a recognition result of the speech sequence; or the first data sequence and the second data sequence are text sequences, and the third target model is used to obtain a recognition result of the text sequence; or the first data sequence and the second data sequence are image sequences, and the third target model is used to obtain a classification result of the image sequence.” [Noting that this is an alternative expression, the third alternative of “…are image sequences, and the third target model is used to obtain a classification result of the image sequence” is taught. FIG. 2B teaches that the combined model (the third to-be-trained model) outputs a task-specific prediction and [0047] teaches a “classification prediction” (i.e., a classification result). Furthermore, the first data sequence and the second data sequences are images. See [0005]: “a base encoder neural network configured to process an input image to generate an intermediate representation of the input image… The operations include: obtaining a training image; performing a plurality of first augmentation operations on the training image to obtain a first augmented image” (i.e., the first sequence is an input image); [0083]: “Transfer learning performance was also evaluated across 12 natural image datasets in both linear evaluation (fixed feature extractor) and fine-tuning settings” (i.e., the fine-tuning uses image data.).]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Gidaris and Chen, including the teachings of Chen cited above, so as to also arrive at the claimed invention of the instant dependent claim. The motivation for doing so is covered by the motivation given for Chen in the rejection of the parent independent claim.
As to claims 8-11 and 13-14, these claims are directed to an apparatus for performing operations that are the same or substantially the same as those of claims 1-4 and 6-7. Therefore, the rejections made to claims 1-4 and 6-7 are applied to claims 8-11 and 13-14, respectively.
Furthermore, Gidaris teaches “a model training apparatus, wherein the apparatus comprises: at least one processor; and one or more memories coupled to the at least one processor and storing programming instructions for execution by the at least one processor to cause the apparatus to…” [Since Gidaris teaches a “fully online teacher-student learning scheme for BoW-based self-supervised training” (§ 1, paragraph 5, item 2) that addresses the problem of “inefficient usage of the computational budget for training” (§ 1, paragraph 4), it is understood that Gidaris’ method is implemented on a computer. Therefore, the instant limitations which are generic computer components are implied by the teachings of Gidaris.]
As to claims 15-18 and 20, these claims are directed to an apparatus for performing operations that are the same or substantially the same as those of claims 1-4 and 6. Therefore, the rejections made to claims 1-4 and 6 are applied to claims 8-11 and 13, respectively.
Furthermore, Gidaris teaches “a non-transitory computer storage medium, wherein the computer storage medium stores one or more instructions, and when the one or more instructions are executed by one or more computers, the one or more computers are enabled to…” [Since Gidaris teaches a “fully online teacher-student learning scheme for BoW-based self-supervised training” (§ 1, paragraph 5, item 2) that addresses the problem of “inefficient usage of the computational budget for training” (§ 1, paragraph 4), it is understood that Gidaris’ method is implemented on a computer. Therefore, the instant limitations which are generic computer components are implied by the teachings of Gidaris.]
2. Claims 5, 12, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Gidaris in view of Chen, and further in view of Hashemi et al., “Enlarging smaller images before inputting into convolutional neural network: zero‑padding vs. interpolation,” Hashemi J Big Data (2019) 6:98 (“Hashemi”).
As to claim 5, the combination of Gidaris and Chen teaches the method according to claim 1, as set forth above, but does not teach the further limitations of the instant dependent claim.
Hashemi teaches “wherein before the processing the perturbed first data sequence by using a first to-be-trained model to obtain a first feature sequence, and processing the first data sequence by using a second to-be-trained model to obtain a second feature sequence, the method further comprises: adding padding elements to two ends of the first data sequence or two ends of the perturbed first data sequence.” [Abstract: “What this study focuses on is the requirement that feature tensors must be of the same size… In classifying an image using a convolutional neural network (CNN), the input is a three-dimensional tensor, where the value of each pixel in each channel is one feature. The three-dimensional feature tensor must be the same size for all images. However, images are not usually of the same size and so are not their corresponding feature tensors… This study proposes zero-padding for resizing images to the same size and compares it with the conventional approach of scaling images up (zooming in) using interpolation. Our study showed that zero-padding had no effect on the classification accuracy but considerably reduced the training time.” As shown in FIG. 2, the zero-padding is applied at both ends (left and right side) of the image, which is analogous to the perturbed first data sequence. The limitation of “before the processing the perturbed first data sequence by using a first to-be-trained model to obtain a first feature sequence” is met because the padding as described in Hashemi is performed prior to use with a neural network. Therefore, if the padded image is the “perturbed first data sequence,” then it would be padded before the image is input into the first to-be-trained model. The limitation of “[before] processing the first data sequence by using a second to-be-trained model to obtain a second feature sequence” is not related to the functionality of the alternative of padding the perturbed sequence. Thus, in the combined teachings of the references as combined below, this limitation would have been obvious under the principles of MPEP § 2144(IV)(C), which states that “selection of any order of performing process steps is prima facie obvious in the absence of new or unexpected results.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Chen by implementing the use of zero-padding for the perturbed first data sequence, so as to arrive at the claimed invention of this dependent claim. The motivation would have been to enable a smaller input to be compatible with a neural network, particularly in a manner that has “no effect on the classification accuracy but considerably reduced the training time,” as suggested by Hashemi, abstract (see parts quoted above and the entirety of the abstract)..
As to claims 12 and 19, the further limitations of these claims are the same or substantially the same as those of claim 5. Therefore, the rejection made to claim 5 is applied to claims 12 and 19.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. The following references depict the state of the art.
Chen et al., “A Simple Framework for Contrastive Learning of Visual Representations,” arXiv:2002.05709v3 [cs.LG] 1 Jul 2020 is the paper corresponding to the Chen patent document discussed in this action, and provides additional explanations and examples for the techniques in Chen.
Zheng et al., “ReSSL: Relational Self-Supervised Learning with Weak Augmentation,” arXiv:2107.09282v2 [cs.CV] 23 Jul 2021 teaches contrastive learning with the momentum weight transfer technique.
Xiao et al., US 20230074706 A1 teaches the use of contrastive loss.
Bunel et al., US 12288393 B2 teaches the use of contrastive loss
Any inquiry concerning this communication or earlier communications from the examiner should be directed to YAO DAVID HUANG whose telephone number is (571)270-1764. The examiner can normally be reached Monday - Friday 9:00 am - 5:30 pm.
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, Miranda Huang can be reached at (571) 270-7092. 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.
/Y.D.H./Examiner, Art Unit 2124
/MIRANDA M HUANG/ Supervisory Patent Examiner, Art Unit 2124
