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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 102
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 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.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
(g)(1) during the course of an interference conducted under section 135 or section 291, another inventor involved therein establishes, to the extent permitted in section 104, that before such person’s invention thereof the invention was made by such other inventor and not abandoned, suppressed, or concealed, or (2) before such person’s invention thereof, the invention was made in this country by another inventor who had not abandoned, suppressed, or concealed it. In determining priority of invention under this subsection, there shall be considered not only the respective dates of conception and reduction to practice of the invention, but also the reasonable diligence of one who was first to conceive and last to reduce to practice, from a time prior to conception by the other.
A rejection on this statutory basis (35 U.S.C. 102(g) as in force on March 15, 2013) is appropriate in an application or patent that is examined under the first to file provisions of the AIA if it also contains or contained at any time (1) a claim to an invention having an effective filing date as defined in 35 U.S.C. 100(i) that is before March 16, 2013 or (2) a specific reference under 35 U.S.C. 120, 121, or 365(c) to any patent or application that contains or contained at any time such a claim.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claims 19 – 20 are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Glaser (US 20230297886 A1; hereafter referred to as Glaser).
Regarding Claim 19, Glaser teaches:
processing circuitry (Glaser, [0133], “an electronic circuit”) to:
determine, from a dataset used to update one or more machine learning models, one or more images associated with one or more first groups that include one or more first performance scores associated with the one or more machine learning models (Glaser, [0041] “appropriate modifications to training data, an initial learning model LM2 may improve upon the results of an initial learning model LM1 by updating, supplementing, or otherwise augmenting training data TM1 to use the modified training data TM2”; Glaser, [0088] “processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition”);
generate, based at least on the one or more images, one or more augmented images associated with one or more second groups that include one or more second performance scores associated with the one or more machine learning models, the one or more second performance scores being less than the one or more first performance scores (Glaser, [0039] “the method may be used to first detect this cluster of data samples with common false positive events and then augment the training data so that a resulting updated computer vision learning model (e.g., an updated supervised learning model) more accurately interprets data”; Glaser, [0040] “determining a third dataset based in part on the bias assessment S1151 (e.g., the first dataset augmented with supplemental samples)”; Glaser, [0088] A variation used with live data may additionally include processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition. Each cluster, or at least clusters of significance, will preferably have sufficient samples in the training data so that the cluster can be adequately analyzed. As a result, the training data or a subset of training data may be scored based on coverage of all or a select set of clusters”); and
causing an update of the one or more machine learning models using the one or more augmented images (Glaser, [0041] Iterative processing using the model may be used to update the training of a supervised learning model. As shown in FIG. 4, with appropriate modifications to training data, an initial learning model LM2 may improve upon the results of an initial learning model LM1 by updating, supplementing, or otherwise augmenting training data TM1 to use the modified training data TM2”).
Regarding Claim 20, Glaser teaches wherein the one or more processors are 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; 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 Al 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; a system for performing one or more conversational Al 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 (Glaser, [0020] “The system and method may be implemented for any general computer/AI learning implementation. Examples include image processing (e.g., object detection, object classification/identification, navigation/tracking, image-based event detection, image/image synthesis/generation, and the like), audio processing (e.g., sound classification, sound transcription, automated audio generation), natural language processing (e.g., text labeling, test summarization, text synthesis), digital bots, sensor data analysis, analysis of other parametric data (e.g., web or app analytics, financial data, business data), and/or other applications. The system and method may be particularly useful for implementations that include learning models with extremely large or open datasets for learning”).
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.
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 1 – 18 are rejected under 35 U.S.C. 103 as being unpatentable over Thong et al (US 20230419549 A1; hereafter referred to as Thong) in view of Glaser (US 20230297886 A1; hereafter referred to as Glaser).
Regarding Claim 1, Thong teaches:
A method comprising:
associating, based at least on annotation data representative of lighting values and hue values of people as depicted by first images from a first dataset, the first images with groups (Thong, [0042] “aspects of the present invention can measure automatically and quantitatively skin color in a multidimensional manner in images… focus on the perceptual light L*, as a measure of skin tone, and the hue angle h*, as a measure of skin hue, which results in a multidimensional measure for every image”; Thong, [0050] Skin color scores provide a quantitative measure to characterize the appearance of the skin in an image. Extracting such measures helps to identify potential biases towards skin color subgroups in model performance”; [0053] “the face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”);
determining, based at least on one or more machine learning models processing the first dataset, a portion of the first images that the one or more machine learning models accurately processed (Thong, [0050] Skin color scores provide a quantitative measure to characterize the appearance of the skin in an image. Extracting such measures helps to identify potential biases towards skin color subgroups in model performance”; [0077] “images were modified in the CelebAMask-HQ datasets in both directions of L* and h*, for a total of 4×30,000 images, and the binary accuracy of evaluated models on gender and smile classifications was reported);
While Thong teaches identifying potential biases towards skin color subgroups in model performance, it does not explicitly teach:
determining, based at least on the portion of the first images, one or more groups from the groups for which one or more performances of the one or more machine learning models are below a threshold performance; and
causing the one or more machine learning models to be updated based at least on a second dataset that includes one or more second images associated with the one or more groups.
In the same field of endeavor, Glaser teaches:
determining, based at least on the portion of the first images, one or more groups from the groups for which one or more performances of the one or more machine learning models are below a threshold performance (Glaser, [0043] “An undesired condition may be triggered when the instances of flagged data samples within a cluster satisfies some condition such as reaching some threshold of number of occurrences, some percentages of samples within the cluster, occurring at or greater than some frequency. Flagged data samples would be flagged as a result of some output of the supervised learning model potentially having a false positive or false negative result”); and
causing the one or more machine learning models to be updated based at least on a second dataset that includes one or more second images associated with the one or more groups (Glaser, [0039] “determining a bias assessment based on the clustering output S140… Modifications based on the bias assessment may include adjustments to amplify data samples poorly represented, add data samples, added select data samples to supplement categories of data with poor representation; Glaser, [0049] For clusters with false positives or false negatives, the base learning model may be anti-trained within this cluster as it generates the wrong results. Labels may be corrected and/or more examples collected to retrain and improve performance within such a cluster”; Glaser, [0066] “The performance of a trained machine learning model can be evaluated and assed using unsupervised learning (or using other segmentation or classification techniques), and then (optionally) a refined machine learning model can be generated based on such bias assessments (e.g., by collecting supplemental training data, augmenting labeling of data, etc.)”).
Thong and Glaser are considered analogous art as they are reasonably pertinent to the same field of endeavor of image processing. Therefore, it would have been obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Thong with the invention of Glaser to make the invention that determines one or more groups from the groups for which one or more performances of the one or more machine learning models are below a threshold performance and updates the one or more machine learning models based on the second dataset; doing so can result in proper representation of less common situations in training the machine learning models with diverse training datasets which would further result in learning model with better performance (Glaser, [0003] and [0016]); thus one of the ordinary skill in the art would have been motivated to combine the references.
Regarding Claim 2, Thong in view of Glaser teaches the method of claim 1, wherein:
the groups are associated with lighting value ranges and hue value ranges (Thong, [0053] “as some parts of the face can skew the results towards darker values (e.g., facial hair or shaded regions), in some embodiments, some groups which yield a very low L* are excluded. The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”; Thong, [0055] Skin color scores enable the assessment of potential biases in a given image dataset. By extracting them for every sample in the dataset, it is possible to estimate the distribution of every subgroup and characterize how unbalanced the distribution might be”); and
the associating the first images with the groups comprises:
determining, based at least on the annotation data, a respective lighting value of the lighting values and a respective hue value of the hue values associated with an individual image of the first images (Thong, [0052] “To extract skin color scores from a facial image, the below describes exemplary steps to extract a skin color score in human-centric images: (a) an input image of a subject is provided; (b) the skin pixels are segmenting. Segmentation can be done manually by an annotator or predicted by a skin segmentation model; (c) once skin pixels have been identified, they can be converted from the standard RGB space to the target space of the desired scoring values. Aspects of the present invention convert to the CIELAB space to extract the L* component, and further use the a* and b* components to compute the hue angle h*. This results in a point measurement of L* and h* for every skin pixel in the image; and (d) a clustering algorithm is applied, such as KMeans, to group the skin pixels”); and
associating the individual image with a group of the groups based at least on the respective lighting value being within a lighting value range associated with the group and the respective hue value being within a hue value range associated with the group (Thong, [0050] Skin color scores provide a quantitative measure to characterize the appearance of the skin in an image. Extracting such measures helps to identify potential biases towards skin color subgroups in model performance”; Thong, [0052] “extract the L* component, and further use the a* and b* components to compute the hue angle h*. This results in a point measurement of L* and h* for every skin pixel in the image; and (d) a clustering algorithm is applied, such as KMeans, to group the skin pixels”).
Regarding Claim 3, Thong in view of Glaser teaches the method of claim 1, wherein the determining the one or more groups for which the one or more performances of the one or more machine learning models are below the threshold performance comprises:
determining, based at least on the portion of the first images, accuracy scores associated with the groups (Thong, [0039] “a multidimensional apparent skin color score is determined, it can be used to reveal biases related to skin color in image datasets and computer vision models, which were previously invisible. The multidimensional skin color score can focus on the perceptual lightless, L*, as a measure of skin tone ranging from light to dark, and the hue angle h*, as a measure of skin hue ranging from red to yellow.”; Thong, [0078] “FIG. 3, it was observed that manipulating the skin color to have a lighter skin tone decreases the gender classification accuracy while the skin hue does not seem to have a large effect. The skin tone bias occurs because models are prone to classify people as feminine when the tone is lighter”);
determining, based at least on the accuracy scores, a threshold accuracy score associated with the groups (Thong, [0010] “ determining the bias in the face image dataset with respect to light and dark skin tones by calculating a number of face images having the L* scalar value greater than and less than a first predetermined threshold; and determining the bias in the face image dataset with respect to skin hue by calculating a number of face images having the h* scalar value greater than and less than a second predetermined threshold”; Thong, [0060] “ a binary thresholding for both perceptual lightness L* and hue angle h*, histograms were plotted of both scores with 20 bins… Individuals with a light skin tone and a red skin hue are over-represented with a much larger count. When considering the skin tone and varying the hue angle thresholding, it was observed that the hue angle has a lower spread for dark skin tones than light skin tones. Conversely, when considering the skin hue and varying the perceptual lightness thresholding, the yellow skin hue tends to have a larger skewness towards light skin tones than the red skin hue.”); and
determining that the one or more groups are associated with one or more accuracy scores, from the accuracy scores, that are less than the threshold accuracy score (Thong, [0060] “a binary thresholding for both perceptual lightness L* and hue angle h*, histograms were plotted of both scores with 20 bins… Individuals with a light skin tone and a red skin hue are over-represented with a much larger count. When considering the skin tone and varying the hue angle thresholding, it was observed that the hue angle has a lower spread for dark skin tones than light skin tones. Conversely, when considering the skin hue and varying the perceptual lightness thresholding, the yellow skin hue tends to have a larger skewness towards light skin tones than the red skin hue.”).
Regarding Claim 4, Thong in view of Glaser teaches the method of claim 1, further comprising:
obtaining a third dataset that includes third images used to update the one or more machine learning models (Glaser, [0039] “training, using a third dataset, a bias assessment modified learning model using supervised learning S150 as shown in FIG. 2”);
determining, based at least on the portion of the first images, one or more second groups for which one or more second performances of the one or more machine learning models are equal or greater than the threshold performance (Glaser, [0040] “ determining a third dataset based in part on the bias assessment S1151 (e.g., the first dataset augmented with supplemental samples); training, using the third dataset, a bias assessment modified learning model using supervised learning S1152; applying the bias assessment modified learning model to the second dataset (or another dataset such as subsequently acquired data) thereby generating a third learning model output S2120”); and
determining the second dataset as including the one or more second images, from the third images, that are associated with the one or more second groups (Glaser, [0107] “Block S150, which includes training, using a third dataset, a bias assessment modified model using supervised learning, functions to train a learning model with the incorporation of the bias assessment. Dependent on implementation, and the desired biases (e.g., added, enhanced, reduced) may be incorporated as rules in the learning model or in the third dataset”; Glaser, [0114] “in many variations the third dataset may be a modified first dataset, and/or the bias assessment modified model is a modified first learning model (e.g., an updated base supervised learning model)”).
Regarding Claim 5, Thong in view of Glaser teaches the method of claim 4, further comprising:
determining one or more accuracy scores associated with the one or more second groups (Thong, [0043] “Given the limitations of conventional skin color scales... deriving quantitative metrics can enable more reliable skin color scores... compute a skin color score from a point measurement”; Thong, [0053] To obtain the final scalar scores for the image, embodiments of the present invention can average the scalar values of every group normalized by their pixel size.... The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”); and
determining one or more weights based at least on the one or more accuracy scores, wherein the determining the second dataset is further based at least on the one or more weights (Glaser, [0030] “the system and method may use upsampling, sample weighting, or other techniques to influence the retraining of a base learning model to address detected issues in select clusters”; Glaser, [0071] “ input training data is used to refine the supervised learning model, which may adjust its weights until a resulting model has been fitted and achieves a desired performance level”)).
Regarding Claim 6, Thong in view of Glaser teaches the method of claim 1, further comprising:
determining one or more accuracy scores associated with the one or more groups (Thong, [0050] Skin color scores provide a quantitative measure to characterize the appearance of the skin in an image. Extracting such measures helps to identify potential biases towards skin color subgroups in model performance”; Thong, [0053] To obtain the final scalar scores for the image, embodiments of the present invention can average the scalar values of every group normalized by their pixel size. However, as some parts of the face can skew the results towards darker values (e.g., facial hair or shaded regions), in some embodiments, some groups which yield a very low L* are excluded. The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”); and
assigning, based at least on the one or more accuracy scores, the one or more second images to the one or more groups, wherein the causing the one or more machine learning models to be updated is based at least on the one or more second images as assigned to the one or more groups (Glaser, [0044] “False positive clusters, false negative clusters, mixed result clusters, and/or inconclusive clusters may have recommendations for additional sample data or receive other training data augmentations when training an updated supervised learning model”; Glaser, [0088] “A variation used with live data may additionally include processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition. Each cluster, or at least clusters of significance, will preferably have sufficient samples in the training data so that the cluster can be adequately analyzed. As a result, the training data or a subset of training data may be scored based on coverage of all or a select set of clusters”; Glaser, [0100] Identification, scoring, and/or labeling of such clusters may be useful in inspecting usage of a learning model and guiding subsequent improvements”).
Regarding Claim 7, Thong in view of Glaser teaches the method of claim 6, further comprising:
determining, based at least on the one or more accuracy scores, one or more weights associated with the one or more groups, wherein the assigning the one or more second images to the one or more groups is based at least on the one or more weights (Glaser, [0030] “the system and method may use upsampling, sample weighting, or other techniques to influence the retraining of a base learning model to address detected issues in select clusters”; Glaser, [0071] “input training data is used to refine the supervised learning model, which may adjust its weights until a resulting model has been fitted and achieves a desired performance level”; Glaser, [0088] “A variation used with live data may additionally include processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition. Each cluster, or at least clusters of significance, will preferably have sufficient samples in the training data so that the cluster can be adequately analyzed. As a result, the training data or a subset of training data may be scored based on coverage of all or a select set of clusters”).
Regarding Claim 8, Thong in view of Glaser teaches the method of claim 1, wherein:
the one or more second images are associated with one or more first lighting values and one or more first hue values (Thong, [0039] “biases related to skin color in image datasets and computer vision models, which were previously invisible. The multidimensional skin color score can focus on the perceptual lightless, L*, as a measure of skin tone ranging from light to dark, and the hue angle h*, as a measure of skin hue ranging from red to yellow”; Thong, [0063] “Given a dataset {x.sub.i,a.sub.1}.sub.i=1.sup.N of N facial images x with their associated race label a. Skin color scales are extracted corresponding to perceptual lightness y.sup.L* and the hue angle y.sup.h* for every facial image”);
the method further comprising generating, based at least on the one or more second images, an augmented dataset that includes one or more augmented images (Glaser, [0037] “the related data samples can then be mapped to a cluster using the unsupervised learning model, which may then be used to augment a new version of a learning model/system”; Glaser, [0039] “the method may be used to first detect this cluster of data samples with common false positive events and then augment the training data so that a resulting updated computer vision learning model”),
the one or more augmented images being associated with one or more second lighting values and one or more second hue values corresponding to the one or more groups (Thong, [0010] “ determining the bias in the face image dataset with respect to light and dark skin tones by calculating a number of face images having the L* scalar value greater than and less than a first predetermined threshold; and determining the bias in the face image dataset with respect to skin hue by calculating a number of face images having the h* scalar value greater than and less than a second predetermined threshold”; Thong, [0039] “ biases related to skin color in image datasets and computer vision models, which were previously invisible. The multidimensional skin color score can focus on the perceptual lightless, L*, as a measure of skin tone ranging from light to dark, and the hue angle h*, as a measure of skin hue ranging from red to yellow”); and
the causing the one or more machine learning models to be updated is based at least on the augmented dataset (Glaser, [0034] “Through multi-stage learning, learning trends may be identified, which can then be used to augment training data and/or the training process to provide (or alter) subjective interpretations in the constructed learning model”).
Regarding Claim 9, Thong teaches:
one or more processors to:
obtain output data associated with one or more machine learning models processing a first dataset including first images associated with groups (Thong, [0051] “provide a framework that starts from a facial image x and outputs a final scalar scoring value y or a set of scalar scoring values y={y.sup.1, . . . ,y.sup.y}”);
determine, based at least on the output data, one or more first groups from the groups for which one or more first performances of the one or more machine learning models are below a threshold performance and one or more second groups from the groups for which one or more second performances of the one or more machine learning models are equal to or greater than the threshold performance (Thong, [0052] “to extract skin color scores from a facial image, the below describes exemplary steps to extract a skin color score in human-centric images: (a) an input image of a subject is provided; (b) the skin pixels are segmenting. Segmentation can be done manually by an annotator or predicted by a skin segmentation model… extract the L* component, and further use the a* and b* components to compute the hue angle h*. This results in a point measurement of L* and h* for every skin pixel in the image; and (d) a clustering algorithm is applied, such as KMeans, to group the skin pixels”; Thong, [0053] “some parts of the face can skew the results towards darker values (e.g., facial hair or shaded regions), in some embodiments, some groups which yield a very low L* are excluded. The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”; Thong, [0056] “when lightness L* is over a value of 60, it corresponds to a light skin tone (and conversely for a dark skin tone). When hue angle h* is over 55°, it corresponds to a skin turning towards yellow (and conversely for a skin turning towards red). These threshold values were taken from literature references”);
While Thong teaches identifying potential biases towards skin color subgroups in model performance, it does not explicitly teach:
determine, based at least on the one or more second groups, a second dataset that includes one or more second images; and
generate, based at least on the one or more second images, an augmented dataset for updating the one or more machine learning modes, the augmented dataset including one or more augmented images associated with the one or more first groups.
In the same field of endeavor, Glaser teaches:
determine, based at least on the one or more second groups, a second dataset that includes one or more second images (Glaser, [0043] “An undesired condition may be triggered when the instances of flagged data samples within a cluster satisfies some condition such as reaching some threshold of number of occurrences, some percentages of samples within the cluster, occurring at or greater than some frequency. Flagged data samples would be flagged as a result of some output of the supervised learning model potentially having a false positive or false negative result”; [0053] “a cluster may be inspected where a number of representative image data samples are shown”; [0058] “the first dataset and the second dataset include image data”); and
generate, based at least on the one or more second images, an augmented dataset for updating the one or more machine learning modes, the augmented dataset including one or more augmented images associated with the one or more first groups (Glaser, [0037]” a machine learning service may offer a way for users to flag results (e.g., flagging as problematic or wrong), the related data samples can then be mapped to a cluster using the unsupervised learning model, which may then be used to augment a new version of a learning model/system”; Glaser, [0039] “the method may be used to first detect this cluster of data samples with common false positive events and then augment the training data so that a resulting updated computer vision learning model (e.g., an updated supervised learning model) more accurately interprets data”; Glaser, [0041] “Iterative processing using the model may be used to update the training of a supervised learning model. As shown in FIG. 4, with appropriate modifications to training data, an initial learning model LM2 may improve upon the results of an initial learning model LM1 by updating, supplementing, or otherwise augmenting training data TM1 to use the modified training data TM2. This enhancement process may be repeated any number of times by applying processes S130, S140, and S150 for further iterations of the process”).
Thong and Glaser are considered analogous art as they are reasonably pertinent to the same field of endeavor of image processing. Therefore, it would have been obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Thong with the invention of Glaser to make the invention that determines a second dataset that includes one or more second images and generates, based at least on the one or more second images, an augmented dataset for updating the one or more machine learning modes; doing so can result in proper representation of less common situations in training the machine learning models with diverse training datasets which would further result in learning model with better performance (Glaser, [0003] and [0016]); thus one of the ordinary skill in the art would have been motivated to combine the references.
Regarding Claim 10, Thong in view of Glaser teaches the system of claim 9, wherein:
the one or more second images are associated with one or more first values of a first color attribute and one or more first values of a second color attribute corresponding to one or more subjects as depicted by the one or more second images (Thong, [0042] “uses of this multidimensional measurement of skin color by (i) quantifying to what extent common image datasets are skewed towards light-red skin color and under-represent dark-yellow skin color, and how generative models trained on these datasets reproduce a similar bias; (ii) revealing multidimensional skin color biases in saliency-based image cropping and face verification models; and (iii) measuring the causal effect of skin color in attribute prediction in multiple commercial and non-commercial models”; Thong, [0072] Skin color causal effect in models. Given a facial image x, the objective is to predict the presence or absence of an attribute”); and
the generation of the augmented dataset comprises:
determining one or more value ranges of the first color attribute and one or more value ranges of the second color attribute associated with the one or more first groups (Thong, [0045] “ the CIE L*a*b* (CIELAB) color space, which correlates with the response of the human eye by covering its entire range of color perception, where the L* component corresponds to the perceptual lightness and ranges from black at value 0 to white at value 100; the a* component describes the green-red opponent colors, with negative values corresponding to green and positive values to red; and the b* component refers to the blue-yellow opponent colors, with negative values corresponding to blue and positive values to yellow”); and
generating the one or more augmenting images by at least augmenting the one or more second images such that the one or more subjects depicted by the one or more second images (Glaser, [0034] “Through multi-stage learning, learning trends may be identified, which can then be used to augment training data and/or the training process to provide (or alter) subjective interpretations in the constructed learning model”; Glaser, [0037] “a machine learning service may offer a way for users to flag results (e.g., flagging as problematic or wrong), the related data samples can then be mapped to a cluster using the unsupervised learning model, which may then be used to augment a new version of a learning model/system”; Glaser, [0041] “Iterative processing using the model may be used to update the training of a supervised learning model. As shown in FIG. 4, with appropriate modifications to training data, an initial learning model LM2 may improve upon the results of an initial learning model LM1 by updating, supplementing, or otherwise augmenting training data TM1 to use the modified training data TM2. This enhancement process may be repeated any number of times by applying processes S130, S140, and S150 for further iterations of the process”) correspond to one or more second values of the first color attribute that are within the one or more value ranges of the first color attribute and one or more second values of the second color attribute that are within the one or more value ranges of the second color attribute (Thong, [0045] “ the CIE L*a*b* (CIELAB) color space, which correlates with the response of the human eye by covering its entire range of color perception, where the L* component corresponds to the perceptual lightness and ranges from black at value 0 to white at value 100; the a* component describes the green-red opponent colors, with negative values corresponding to green and positive values to red; and the b* component refers to the blue-yellow opponent colors, with negative values corresponding to blue and positive values to yellow”).
Regarding Claim 11, Thong in view of Glaser teaches the system of claim 9, wherein the one or more processors are further to:
update, using a third dataset that includes third images, the one or more machine learning models during a first updating process, where the second dataset includes a portion of the third dataset (Glaser, [0039] “The third dataset may be a dataset that may be based on the first dataset, modified based on the bias assessment. Modifications based on the bias assessment may include adjustments to amplify data samples poorly represented, add data samples, added select data samples to supplement categories of data with poor representation, add synthesized data samples, and/or modified in other ways”); and
update, using the augmented dataset, the one or more machine learning models during a second updating process (Glaser, [0039] “The third dataset may be a dataset that may be based on the first dataset, modified based on the bias assessment. Modifications based on the bias assessment may include adjustments to amplify data samples poorly represented, add data samples, added select data samples to supplement categories of data with poor representation, add synthesized data samples, and/or modified in other ways”; Glaser, [0058] “determining a bias assessment for at least one cluster in the clustering of the second set of image data (S140); and optionally training an updated supervised learning CV model with an updated training set of data (S150), wherein the updated training set of data is generated based on the bias assessment of the at least one cluster”).
Regarding Claim 12, Thong in view of Glaser teaches the system of claim 9, wherein the determination of the one or more first groups for which the one or more first performances are less than the threshold performance and the one or more second groups for which the one or more second performances are equal to or greater than the threshold performance comprises:
determining, based at least the output data, one or more first accuracy scores associated with the one or more first groups and one or more second accuracy scores associated with the one or more second groups (Thong, [0039] “a multidimensional apparent skin color score is determined, it can be used to reveal biases related to skin color in image datasets and computer vision models, which were previously invisible. The multidimensional skin color score can focus on the perceptual lightless, L*, as a measure of skin tone ranging from light to dark, and the hue angle h*, as a measure of skin hue ranging from red to yellow.”; Thong, [0078] “FIG. 3, it was observed that manipulating the skin color to have a lighter skin tone decreases the gender classification accuracy while the skin hue does not seem to have a large effect. The skin tone bias occurs because models are prone to classify people as feminine when the tone is lighter”);
determining, based at least on the one or more first accuracy scores and the one or more second accuracy scores, a threshold accuracy score associated with the groups (Thong, [0010] “ determining the bias in the face image dataset with respect to light and dark skin tones by calculating a number of face images having the L* scalar value greater than and less than a first predetermined threshold; and determining the bias in the face image dataset with respect to skin hue by calculating a number of face images having the h* scalar value greater than and less than a second predetermined threshold”; Thong, [0060] “ a binary thresholding for both perceptual lightness L* and hue angle h*, histograms were plotted of both scores with 20 bins… Individuals with a light skin tone and a red skin hue are over-represented with a much larger count. When considering the skin tone and varying the hue angle thresholding, it was observed that the hue angle has a lower spread for dark skin tones than light skin tones. Conversely, when considering the skin hue and varying the perceptual lightness thresholding, the yellow skin hue tends to have a larger skewness towards light skin tones than the red skin hue.”);
determining that the one or more first accuracy scores are less than the threshold accuracy score (Thong, [0060] “a binary thresholding for both perceptual lightness L* and hue angle h*, histograms were plotted of both scores with 20 bins… Individuals with a light skin tone and a red skin hue are over-represented with a much larger count. When considering the skin tone and varying the hue angle thresholding, it was observed that the hue angle has a lower spread for dark skin tones than light skin tones. Conversely, when considering the skin hue and varying the perceptual lightness thresholding, the yellow skin hue tends to have a larger skewness towards light skin tones than the red skin hue.”); and
determining the one or more second accuracy scores are equal to or greater than the threshold accuracy score (Thong, [0010] “determining the bias in the face image dataset with respect to light and dark skin tones by calculating a number of face images having the L* scalar value greater than and less than a first predetermined threshold; and determining the bias in the face image dataset with respect to skin hue by calculating a number of face images having the h* scalar value greater than and less than a second predetermined threshold”; Thong, [0056] “when lightness L* is over a value of 60, it corresponds to a light skin tone (and conversely for a dark skin tone). When hue angle h* is over 55°, it corresponds to a skin turning towards yellow (and conversely for a skin turning towards red). These threshold values were taken from literature references”).
Regarding Claim 13, Thong in view of Glaser teaches the system of claim 9, wherein the determination of the second dataset comprises:
obtaining a third dataset that includes third images used to update the one or more machine learning models (Glaser, [0039] “training, using a third dataset, a bias assessment modified learning model using supervised learning S150 as shown in FIG. 2”); and
determining the second dataset as including the one or more second images, from the third images, that are associated with the one or more second groups (Glaser, [0107] “Block S150, which includes training, using a third dataset, a bias assessment modified model using supervised learning, functions to train a learning model with the incorporation of the bias assessment. Dependent on implementation, and the desired biases (e.g., added, enhanced, reduced) may be incorporated as rules in the learning model or in the third dataset”; Glaser, [0114] “in many variations the third dataset may be a modified first dataset, and/or the bias assessment modified model is a modified first learning model (e.g., an updated base supervised learning model)”).
Regarding Claim 14, Thong in view of Glaser teaches the system of claim 9, wherein the one or more processors are further to:
determining one or more accuracy scores associated with the one or more second groups (Thong, [0043] “Given the limitations of conventional skin color scales... deriving quantitative metrics can enable more reliable skin color scores... compute a skin color score from a point measurement”; Thong, [0053] To obtain the final scalar scores for the image, embodiments of the present invention can average the scalar values of every group normalized by their pixel size.... The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”); and
determining one or more weights based at least on the one or more accuracy scores, wherein the determining the second dataset is further based at least on the one or more weights (Glaser, [0030] “the system and method may use upsampling, sample weighting, or other techniques to influence the retraining of a base learning model to address detected issues in select clusters”; Glaser, [0071] “ input training data is used to refine the supervised learning model, which may adjust its weights until a resulting model has been fitted and achieves a desired performance level”)).
Regarding Claim 15, Thong in view of Glaser teaches the system of claim 9, further comprising:
determining one or more accuracy scores associated with the one or more groups (Thong, [0050] Skin color scores provide a quantitative measure to characterize the appearance of the skin in an image. Extracting such measures helps to identify potential biases towards skin color subgroups in model performance”; Thong, [0053] To obtain the final scalar scores for the image, embodiments of the present invention can average the scalar values of every group normalized by their pixel size. However, as some parts of the face can skew the results towards darker values (e.g., facial hair or shaded regions), in some embodiments, some groups which yield a very low L* are excluded. The face skin can be clustered into five groups and the top-3 groups with the highest L* are kept to compute the final L* and h* scalar scores for the image”); and
assign, based at least on the one or more accuracy scores, the one or more second images to the one or more groups, wherein the causing the one or more machine learning models to be updated is based at least on the one or more second images as assigned to the one or more groups (Glaser, [0044] “False positive clusters, false negative clusters, mixed result clusters, and/or inconclusive clusters may have recommendations for additional sample data or receive other training data augmentations when training an updated supervised learning model”; Glaser, [0088] “A variation used with live data may additionally include processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition. Each cluster, or at least clusters of significance, will preferably have sufficient samples in the training data so that the cluster can be adequately analyzed. As a result, the training data or a subset of training data may be scored based on coverage of all or a select set of clusters”; Glaser, [0100] Identification, scoring, and/or labeling of such clusters may be useful in inspecting usage of a learning model and guiding subsequent improvements”).
Regarding Claim 16, Thong in view of Glaser teaches the system of claim 15, further comprising:
determining, based at least on the one or more accuracy scores, one or more weights associated with the one or more groups, wherein the assigning the one or more second images to the one or more groups is based at least on the one or more weights (Glaser, [0030] “the system and method may use upsampling, sample weighting, or other techniques to influence the retraining of a base learning model to address detected issues in select clusters”; Glaser, [0071] “input training data is used to refine the supervised learning model, which may adjust its weights until a resulting model has been fitted and achieves a desired performance level”; Glaser, [0088] “A variation used with live data may additionally include processing the first dataset with the second learning model and evaluating/scoring clusters based on mapping of training data to clusters. This may be used, for example, to identify clusters where the training data does not satisfy a sample count condition. Each cluster, or at least clusters of significance, will preferably have sufficient samples in the training data so that the cluster can be adequately analyzed. As a result, the training data or a subset of training data may be scored based on coverage of all or a select set of clusters”).
Regarding Claim 17, Thong in view of Glaser teaches the system of claim 9, wherein:
the groups are associated with value ranges of the first color attribute and value ranges of the second color attribute (Thong, [0042] “uses of this multidimensional measurement of skin color by (i) quantifying to what extent common image datasets are skewed towards light-red skin color and under-represent dark-yellow skin color, and how generative models trained on these datasets reproduce a similar bias; (ii) revealing multidimensional skin color biases in saliency-based image cropping and face verification models; and (iii) measuring the causal effect of skin color in attribute prediction in multiple commercial and non-commercial models”; [0072] Skin color causal effect in models. Given a facial image x, the objective is to predict the presence or absence of an attribute”); and
the first images are associated with the groups by:
determining, based at least on annotation data associated with the first dataset, a respective value of the first color attribute and a respective value of the second color attribute associated with an individual image of the first images (Thong, [0045] “ the CIE L*a*b* (CIELAB) color space, which correlates with the response of the human eye by covering its entire range of color perception, where the L* component corresponds to the perceptual lightness and ranges from black at value 0 to white at value 100; the a* component describes the green-red opponent colors, with negative values corresponding to green and positive values to red; and the b* component refers to the blue-yellow opponent colors, with negative values corresponding to blue and positive values to yellow”); and
associating the individual image with a group of the groups based at least on the respective value of the first color attribute being within a value range of the first color attribute associated with the group and the respective value of the second color attribute being within a value range of the second color attribute associated with the group (Thong, [0063] “a quantitative score is proposed to measure the probability of a face to be selected by the cropping algorithm. Given a dataset {x.sub.i,a.sub.1}.sub.i=1.sup.N of N facial images x with their associated race label a. Skin color scales are extracted corresponding to perceptual lightness y.sup.L* and the hue angle y.sup.h* for every facial image. To evaluate the cropping algorithm, images were built to include a unique pair of two facial images x.sub.i and x.sub.j with i≠j”; Glaser, [0047] “the method may be used to appropriately address clusters according to the results within a cluster and/or the associated training data related to the cluster”; Glaser, [0048] For clusters of true positive or true negatives, the base learning model may be determined to be well trained within this cluster as they generate appropriate outputs for a samples within the cluster. The method may additionally verify that the associated training data associated with this cluster also meets conditions to qualify as a well trained cluster”).
Regarding Claim 18, Thong in view of Glaser teaches the system of claim 9, 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; 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 (Glaser, [0018] “the system and method can be used in automating the identification and detection of different output scenarios of a machine learning model including various edge cases”); a system implemented using a robot; a system for performing one or more generative Al operations (Thong, [0058] “images were generated with a generative adversarial network (StyleGAN3) and a diffusion model”); 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; a system for performing one or more conversational Al 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 (Glaser, [0020] “The system and method may be implemented for any general computer/AI learning implementation. Examples include image processing (e.g., object detection, object classification/identification, navigation/tracking, image-based event detection, image/image synthesis/generation, and the like), audio processing (e.g., sound classification, sound transcription, automated audio generation), natural language processing (e.g., text labeling, test summarization, text synthesis), digital bots, sensor data analysis, analysis of other parametric data (e.g., web or app analytics, financial data, business data), and/or other applications. The system and method may be particularly useful for implementations that include learning models with extremely large or open datasets for learning”).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
US 20260065146 A1 BIAS MITIGATION METHOD AND SYSTEM FOR AI SYSTEMS A computer-implemented method for supporting bias mitigation in an artificial intelligence (AI) system includes determining a set of sensitive attributes and providing a dataset including a number of data elements. Each data element is labelled with sensitive attributes. The AI system runs on the dataset and determines whether a prediction for an element is correct. Upon checking whether a bias with regard to a sensitive attribute is present, for each sensitive attribute that exhibits a bias, a model is trained for an attribute-based global explanation for each class of correct and incorrect predictions. For each incorrectly predicted data element based on the trained model for the at least one attribute-based global explanation, a counterfactual data element is generated that leads to a correct classification.
WO 2025019335 A2 BIAS MITIGATION OF MACHINE LEARNING MODELS Techniques for performing image analysis and/or classification using a trained machine learning model, and mitigating biases in said trained machine learning model, are provided. Global importance analysis techniques described herein enable the quantitative assessment of image feature(s) used by the trained machine learning model in making predictions such that unwanted biases may be iteratively unlearned in successive versions of the trained machine learning model. The trained machine learning model may be assessed using a set of composite or transformed images generated by combining segmented first image portions with second image portions and/or by applying transformations to the first and/or second image portions. A quantitative assessment of feature importance is then performed using global importance analysis techniques, and the trained machine learning model may be newly trained or retrained using the generated set of images to mitigate bias in the model.
Contact Information
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VAISALI RAO KOPPOLU whose telephone number is (571)270-0273. The examiner can normally be reached Monday - Friday 8:30 - 5.
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, Jennifer Mehmood can be reached at (571) 272-2976. 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.
VAISALI RAO. KOPPOLU
Examiner
Art Unit 2664
/VAISALI RAO KOPPOLU/Examiner of Art Unit 2664
