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 . This action is made final.
This office action is in response to the amendments filed on April 23, 2026.
Claims 1, 21, and 22 have been amended.
Response to Amendment
The amendments filed April 23, 2026 have been entered. Claims 1, 2, 11-22 remain pending in the case and have been rejected.
Response to Arguments
Regarding the 101 arguments
Applicant’s arguments, see pages 6-8, filed April 23, 2026, with respect to Step 2A Prong 2 have been fully considered and are persuasive. The rejection of December 23, 2025 has been withdrawn.
Regarding the 103 arguments
Applicant’s arguments with respect to claims have been considered but are moot in view of the new grounds of rejection necessitated by the amendment.
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.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claim(s) 1, 2, 12-17, and 20-22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ando (US 12026935 B2), in view of Guo et al. (US 11062180 B2), Campos (US 7069256 B1, referred to as Campos), in view of Salekin et al. (US 20210005067 A1, referred to as Salekin), in view of Longmire et al. (US 11670421 B2, referred to as Longmire).
Regarding claim 1 Ando teaches,
An information processing apparatus comprising:
a processor configured to (Col. 2, lines 33-52: Describes an apparatus/device for performing image processing comprising computer components including a processor.):
obtain a dataset of training data (FIG. 3 and Col. 8, lines 41-58: FIG. 3 shows that an acquisition section 20, takes an image to be processed to produce a trained model; Fig. 4 S102 and Col. 11 lines 59-64: FIG. 4 is a flowchart of the training of the model, and step S102 acquires the input image from the acquisition unit, which takes an image from a plurality of images which corresponds to a dataset of training data; FIG. 4 S111 and Col. 18, lines 50-67: The acquisition section obtains a new input image from the dataset for the iterative process, FIG. 4 starting at step S102 of training the model, at step S111 a decision is made to retrain the model at S102.);
normalizing numerical items of the training data using predetermined conversion functions (Col. 11, lines 63-67, cont. Col 12. Lines 1-35: Describes acquiring input images and associated correct labels which are forms of categorical/numerical information tired to the training samples.; Col. 13, lines 58-67, cont. Col 14, lines 1-5: Describes performing color correction or gamma correction and reconverting RGB values based on a mathematical expression, these expressions are predefined to conversion functions applied to the numerical pixel values computing normalization.; Col. 15, lines 1-67, cont. Col. 16, lines 1-3: Describes applying various image conversion processes (color correction, brightness correction, smoothing, noise addition, affine transformation) with parameter changes, all predetermined transformations that produce normalized or standardized forms of the data; Col. 16, lines 52-67, cont. Col 17, lines 1-67: Describes using cross-entropy or square error for classification or conversion processes, indicating the processed outputs (post-conversion data) are in vector form for model training.);
Although Ando teaches an information processing apparatus comprising a processor configured to obtain a dataset of training data and normalizing numerical items of the training data using predetermined conversion functions. It does not teach it generate a vector conversion model for converting the dataset of training data, wherein the vector conversion model, comprises an input layer , an embedding layer , and an output layer , is updated during training based on the training data , and is trained such that variability in distribution of output vectors is decreased converting categorical items of the training data into vectors using the vector conversion model perform coordinated preprocessing operations on the training data that reduce variability in weight during model training, generate overlapping partial data groups from post-conversion training data based on a time window indicating a predetermined time range.
Guo teaches generate a vector conversion model for converting the dataset of training data, wherein the vector conversion model (Col. 8, 37-59: Describes that the embedding model is trained to take the dataset and produce vectors representing its content “an embedding model may be initially trained on the entirety of the noisy set of training images for the image classification”. The embedding model takes in image categories and converts them into vector representations in a structured feature space corresponding to a vector conversion model.):
comprises an input layer , an embedding layer , and an output layer , is updated during training based on the training data , and is trained such that variability in distribution of output vectors is decreased (FIG. 1, Col. 11, lines 5-67, cont. Col. 12, lines 1-67, cont. Col.13, lines 1-2: Describes a vector conversion model implemented as a deep convolutional neural network (CNN) that generates vector representations (embeddings) of training data. The training engine TE !@) receives labeled training images and “trains an embedding model” to map each image into a high-dimensional feature space via deep convolutional features, which requires an input layer, intermediate embedding layers, and an output layer that produces the embedding vector. The embedding model and the downstream MVM are iteratively updated during training through supervised or weakly-supervised backpropagation, where curriculum trainer 128 iteratively updates the MVM’s weights such that the loss function is minimized. The embedding process places visually similar images close together in the feature space, producing clean subsets with tightly grouped embeddings and progressively nosier subsets with greater dispersion. The curriculum learning framework begins training with the cleanest, tightly clustered subsets and gradually incorporates nosier ones, thereby reducing the variability of the embedding output vectors and stabilizing the learned representations.);
converting categorical items of the training data into vectors using the vector conversion model (Col. 8, 37-59: Describes using image categories to convert into vector representations, which corresponds to categorical items.; Col. 11, lines 16-35: Describes further the creation of a conversion model and uses image categories for generation of that model.; Col. 13, lines 63-67, cont. Col. 14, lines 1-9: Describes that the model outputs a vector for each image based on its category.);
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando by incorporating the teachings of Guo to include converting training data into vectors and generates a vector conversion model based ion the training data. Doing so would improve computational efficiency and reduce model complexity by transforming high-dimensional data into compact vector representations. This allows neural networks to process categorized training inputs more efficiently in a lower-dimensional space. Enabling the system to represent the training data in a structured format for machine learning, such as a vectorized feature space. This structured representation allows for more efficient processing, easier comparison between data points, and improve accuracy when training classification or prediction models, which is a known benefit in the flied of machine learning.
Although Guo teaches an generating a vector conversion model for converting the dataset of training data, wherein the vector conversion model that comprises an input layer , an embedding layer , and an output layer , is updated during training based on the training data , and is trained such that variability in distribution of output vectors is decreased and converting categorical items of the training data into vectors using the vector conversion model. It does not teach perform coordinated preprocessing operations on the training data that reduce variability in weight during model training.
Campos teaches, perform coordinated preprocessing operations on the training data that reduce variability in weight during model training by (Col. 4 lines, 60-67 cont. Col. 5, lines 1-39: Describes A data preprocessing block which receives training data and processes the training data before the data is used to build a model. Numeric calms within the training data may be normalized to restrict the range of data or eliminate outliers in default or predefined processing may be performed.);
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando’s reduced variation weight updating model training, in view of Guo’s vector conversion representations, to incorporate Campos’s preprocessing operations. Doing so would improve the training data for model input by providing cleaner, consistently scaled inputs to the model and downstream network.
Although Campos teaches an perform coordinated preprocessing operations on the training data that reduce variability in weight during model training. It does not teach generate overlapping partial data groups from post-conversion training data based on a time window indicating a predetermined time range.
Salekin teaches, generate overlapping partial data groups from post-conversion training data based on a time window indicating a predetermined time range (FIG. 2, and [0034-0038]: Describes segmenting an input clip into overlapping window segments of a predetermined time length with a predetermined temporal overlap, and then computing feature vectors for each segment that form a matrix representing the states of those windows. This generates overlapping partial data groups form post-conversion training data based on a time window.); and
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando, in view of Guo, in view of Campos’s system by incorporating the teachings of Salekin to include overlapping partial data groups from post-conversion training data based on a time window indicating a predetermined time range. Doing so would allow the system to segment abrupt transitions between training samples and promote more stable weight updates.
Ando in view of Guo, in view of Campos, in view of Salekin teaches, generate a model with decreased variability in weight based on the overlapping partial data groups (Ando Col. 20, lines 8-18: Describes calculating a loss function based on the difference between outputs of differently augmented versions of the same image, and minimizing that difference; Col. 20, lines 29-67, cont. Col. 21, lines 1-25: Describes that this training produces a model which is robust to changes in image input, with parameters adjusted to minimize variation in the model’s response, indicating that the training process produces stable parameters that are less affected by variations in the input data, corresponding to reduced variability in weight values.; As described above Salekin describes overlapping partial data groups.);
Although Ando, in view of Guo, in view of Campos, in view of Salekin teaches An information processing apparatus comprising: a processor configured to: obtain a dataset of training data; generate a vector conversion model for converting the dataset of training data, wherein the vector conversion model: comprises an input layer, an embedding layer, and an output layer, is updated during training based on the training data, and is trained such that variability in distribution of output vectors is decreased; perform coordinated preprocessing operations on the training data that reduce variability in weight during model training by: normalizing numerical items of the training data using predetermined conversion functions; and converting categorical items of the training data into vectors using the vector conversion model; generate overlapping partial data groups from post-conversion training data based on a time window indicating a predetermined time range; and generate a model with decreased variability in weight based on the overlapping partial data groups. They do not teach apply the model to input data to generate an inference result, wherein a decrease in the variability in weight enhances an accuracy of the inference results relative to a model generated without the coordinated preprocessing operations.
Longmire teaches, apply the model to input data to generate an inference result(Col. 12, lines 50-67 cont. Col. 13, lines 1-5 and FIGS 15-16:Describes combining normalized features into a single feature vector per patient per unit of time, creating a time series for downstream ECOG analysis, and classifying or regressing ECOG ratings from input features or latent representation suing prediction algorithms including neural networks, which are trained with stochastic gradient descent.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando, in view of Guo, in view of Campos, in view of Salekin’s system by incorporating the teachings of Longmire to include the coordinated preprocessing inference method. Doing so would provide normalized numerical data and vectorized categorical data in a common feature vector format, improving training stability and downstream prediction/classification accuracy.
Ando, in view of Guo, in view of Campos, in view of Salekin, in view of Longmire teaches wherein a decrease in the variability in weight (As described by Ando above.) enhances an accuracy of the inference results relative to a model (Longmire Col. 9, lines 12-48: Describes evaluating neural network performance using metrics including accuracy, precision, recall, and F1-score and measuring delta improvement with respect to baseline performance.; Col. 10, lines 23-41: Describes that co-analyzing multimodal data can significantly improve the accuracy of recognition and prediction.) generated without the coordinated preprocessing operations (As described above by Campos and Guo, where Campos is relied upon for numerical preprocessing/normalization, and Guo is relied upon for vector conversion/embedding.)
Regarding claim 2, Ando in view of Guo, in view of Campos, in view of Salekin teaches the information processing apparatus according to claim 1. Ando further teaches wherein the processor is further configured to generate the model in such a way that there is a decrease in standard deviation or dispersion of the weight (FIG. 5A and Col. 10, lines 61-67 cont. Col. 11, lines 1-9: The training process includes adjusting weight coefficients between layers of a neural network. Specifically, each node’s output is calculated using weights and biases, and backpropagation is applied to refine those weights throughout the training.; FIG. 4 S109-S111 and Col. 17 lines 56-67 cont. Col. 18 line 1-67: The iterative training process in which weight coefficients update each layer of the neural network through backpropagation (S109-S110) and refined through repeated training cycles (S102-S111) until a condition is met. This process adjusts weights towards convergence to reduce prediction error. As the model converges, the dispersion of weight values is decreased as the weight values are optimized throughout the training process.).
Regarding claim 11, Ando in view of Guo, in view of Campos, in view of Salekin teaches an information processing apparatus of Claim 1. Guo further teaches, wherein the processor is configured to generate the vector conversion model by performing a training operation (FIG. 1 and Col. 10, lines 5-20, FIG. 9 and Col. 21, lines 23-43: The training engine (TE 120) which is implemented and/or hosted by training engine computing device 102, which can be implemented by computing device 900. Thus, the learning unit is, for example, a processor, integrated circuit, or equivalent thereof. This can implement the training engine computing device 102.; FIG. 1, and Col. 11, lines 16-18: “TE 120 employs the set of training images labelled with the associated image category to train an embedding model for all images in the whole training set.” This corresponds to a learning unit (TE120) generating the vector conversion model by performing a training operation.).
Regarding claim 12, Ando in view of Guo, in view of Campos, in view of Salekin teaches an information processing apparatus of Claim 11. Guo further teaches, wherein the vector conversion model that is trained in features of the training data ([Col. 8, lines 37-59: Describes how the embedding model is trained to imbed each image into a vector in a feature space, this corresponds to a vector conversion model trained in features of the training data.; FIG 5. Blocks 502, 504 and 506, and Col. 17, lines 62-67 cont. Col. 18, lines 1-8: Describes receiving a set of training images (block 502), then an embedding model is generated based on the set of training images (block 504), each training image is embedded into a vector representation based on that category (block 506). This transformation corresponds to generating a vector conversion model trained based on the set if training images, which corresponds to features of the training data.).
Regarding claim 13, Ando in view of Guo, in view of Campos, in view of Salekin the information processing apparatus of Claim 12. Guo further teaches the learning unit generates the vector conversion model in such a way that there is a decrease in variability in distribution of vectors output by the vector conversion model (FIG. 1 and Col. 12 lines 66-67 cont. Col. 13, lines 1-2: “The curriculum trainer 128 may iteratively updated the MVM's weights via various backpropagation methods, such that the loss function for the trained MVM is minimized, or at least decreased.” The TE training engine 120 from FIG. 1 corresponds to the learning unit. It incorporates the curriculum trainer which iteratively updates weights of the vector conversion model using backpropagation. This minimizes the loss function by decreasing the output error of the training engine, which converges the output of the model, reducing variability).
Regarding claim 14, Ando in view of Guo, in view of Campos, in view of Salekin teaches the information processing apparatus according to claim 1.
Salekin teaches wherein the processor is configured to generate the model, with data corresponding to each of the overlapping partial data groups serving as data to be input to the model (FIG. 2 and [0034-0038]: Describes that the processor generates the model using data corresponding to each of the overlapping partial data groups. The segmenting of input audio clip into a sequence of overlapping window segments each defined by a predetermined time length and predetermined temporal overlap. For each overlapping window segment, the system extracts a corresponds HLD feature vector HLDi and the collection of feature vectors HLD1 to HLDn forms the feature matrix used as the input to the model for training and classification).
Regarding claim 15, Ando in view of Guo, in view of Campos, in view of Salekin teaches an information processing apparatus of Claim 14. Salekin further teaches wherein the processor is configured to generate the model using a partial data group that is generated from the dataset, in which sets of training data are associated to time, based on a time window indicating a predetermined time range ([0026-0033]: [0026] describes receiving audio signals to detect certain target events, by segmenting an audio stream into audio clips based on a predetermined range (e.g. 30 seconds), corresponding to training data associated to time, based on a time window indicating a predetermined time range.).
Regarding claim 16, Ando in view of Guo, in view of Campos, in view of Salekin teaches an information processing apparatus of Claim 15. Salekin further teaches wherein the processor is configured to generate the model using the partial data group in which a plurality of sets of partial data overlappingly contains a single set of the post-conversion training data (FIG. 2 and [0034]: Discloses that an individual audio clip 102, represents a single set of training data. This is divided into a sequence of overlapping window segments 104, these correspond to a plurality of partial data group which overlaps the single dataset.).
Regarding claim 17, Ando in view of Guo, in view of Campos, in view of Salekin teaches an information processing apparatus of Claim 15. Salekin further teaches wherein the processor is configured to generate the model, with data corresponding to each of the partial data group serving as data to be input to the model (FIG. 2, and [0026-0039]: shows the creation of the partial data groups; FIG. 3, and [0040-0045]: Describes the input of the partial data groups for the generation of the model).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando in view of Guo in view of Salekin by incorporating the further teachings of Salekin to include generates the model using the partial data group. Doing so would enable modular and parallel training, improving scalability and processing of the model. Allowing each partial data group to be process individually avoids overburdening the model with excessive amounts of data.
Regarding claim 20 Ando in view of Guo, in view of Campos, in view of Salekin teaches the information processing apparatus according to claim 1. Ando further teaches wherein the generating unit generates the model by
sending data to be used in generation of the model to an external model generation server, requesting the external model generation server to learn the model, and
receiving the model learnt by the external model generation server from the external model generation server
(FIG. 13, and Col. 27, lines 38-52: Teaches generating the model by sending data from a terminal device to an external server system, performing training of the model and returning the trained model to the terminal device.).
Regarding claim 21, recites substantially the same limitations as claim 1. Claim 21 further recites an information processing method implemented in an information processing apparatus, comprising (Ando, Col. 6, lines 11-26: Describes the image processing apparatus executes the information processing method.) steps corresponding to the information processing apparatus of claim 1 is therefore rejected on the same premise.
Regarding claim 22, recites substantially the same limitations as claim 1. Claim 21 further recites a non-transitory computer-readable storage medium having stored therein an information processing program that causes a computer to execute (Ando, Col. 10, lines 13-36: Describes the computer-readable storage medium which stores the information processing method program within.) operations corresponding to the information processing apparatus of claim 1 is therefore rejected on the same premise.
Claim(s) 18, and 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ando (US 12026935 B2), in view of Guo et al. (US 11062180 B2), Campos (US 7069256 B1, referred to as Campos), in view of Salekin et al. (US 20210005067 A1, referred to as Salekin), in view of Longmire et al. (US 11670421 B2, referred to as Longmire), in view of Mayer et al. (US 20210287089 A1, referred to as Mayer).
Regarding claim 18, Ando in view of Guo, in view of Campos, in view of Salekin, in view of Longmire teaches an information processing apparatus of Claim 1. However, Ando in view of Guo, in view of Campos, in view of Salekin does not explicitly teach wherein the processor is configured to generate the model using batch normalization.
Mayer teaches wherein the processor is configured to generate the model using batch normalization(FIG.2 and [0127]: Explains that batch normalization is used when the Initial Hyperparameter Module 218 determines it appropriate.; FIG. 5 and [0114]: Shows that the network 500 utilizes batch normalization following activation of the layers in FIG. 5.; FIG. 2, FIG.7, and [0131-0132]: FIG. 7 describes the training of a neural network using batch normalization in step 718. FIG. 2, Initial Hyperparameter Module 218, is the module for implementing the batch normalization in the model, within the Training Module 214, which corresponds to a generating unit the generates the model using batch normalization.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando in view of Guo, in view of Campos, in view of Salekin apparatus by incorporating the teachings of Mayer to include Batch Normalization. Doing so would allow the model to converge faster, reduce training time, and produce more stable and generalizable models. Batch normalization minimizes internal covariate shift stabilizing learning dynamics.
Regarding claim 19, Ando in view of Guo, in view of Campos, in view of Salekin, in view of Longmire, in view of Mayer teaches an information processing apparatus of Claim 18.
Mayer further teaches wherein the processor is configured to generate the model using the batch normalization in which input of each layer of the model is normalized (FIG. 5 and [0114]: Shows that the network 500 includes a series of layers that are each followed by batch normalization layers BN1-BN3, corresponding to each layer of the model is normalized.; FIG.2, and [0127]: Explains that batch normalization is applied to the output of each layer, and that the Initial Hyperparameter Module 218 determines how it is applied.; FIG. 2, and [0082]: Training Module 214 generates the model using batch normalization in the Initial Hyperparameter Module 218. The Model Construction Module 216 creates the layers of the model to be trained by the Training Module 214, this corresponds to each layer of the model being normalized using batch normalization by the generating unit.).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Ando in view of Guo, in view of Campos, in view of Salekin in view of Mayer by incorporating the further teachings of Mayer to include Batch Normalization. Doing so would reduce internal covariate shift to stabilize training distribution, leading to faster convergence and improved model generalization. This is known to improve deep neural network training efficiency, especially in models with many layers.
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 DONALD T RODEN whose telephone number is (571)272-6441. The examiner can normally be reached Mon-Thur 8:00-5:00 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, Omar Fernandez Rivas can be reached at (571) 272-2589. 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.
/D.T.R./Examiner, Art Unit 2128
/OMAR F FERNANDEZ RIVAS/Supervisory Patent Examiner, Art Unit 2128
Prosecution Insights