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 .
Claim Objections
Claim 1-20 objected to because of the following informalities:
The submitted claims numbering is incorrectly numbered based on the interpretation of the claims. The claim tree below is the method by which the claims are examined and are ordered. Dependent claims considered to depend on the claim shown in “Claims Dependency”. The reason for the objection is to examine the claims properly and avoid the multiple rejections for improper dependency, and 35 USC 112(b) such as lack of antecedent and indefinite issues, and 112(b) issues present in the submitted claim tree. Objections under improper dependency and 35 USC 112 (d) issues would be proper if the issue is not corrected. All claims will be referred to as shown on the claim numbering shown in “Examined Claim Number” Column.
Claim Tree (The claims are examined as shown below)
Submitted Claim Number
Examined Claim Number
Claims Dependency
1
1
Independent Claim
1
2
1
2
3
1
3
4
3
4
5
1
5
6
1
6
7
1
7
8
Independent Claim
8
9
8
9
10
8
10
11
8
11
12
8
12
13
8
13
14
8
14
15
Independent Claim
15
16
15
16
17
16
17
18
15
18
19
15
19
20
15
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) significantly more. The subject matter eligibility test for products and process is describe below for claim 1 in view of dependent claims.
Regarding claim 1:
Step 1: Is the claim to a process machine manufacture or composition of matter?
Yes – Claim 1 recites a method and that falls under the statutory categories.
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“determining, by the device, whether an agreement is achieved between the prediction feedback and the explanation feedback based on a threshold;” - The limitations of claim 1 recites a mental process of determining whether the agreement is reached based on a threshold (see MPEP 2106.04(a)(2)III).
Step 2 Prong 2: Does the claim recite additional elements that integrate the judicial exception into a particular application? No –
The claim includes the additional element(s):
“receiving, by a device and from a user device, a machine learning model, training data utilized to train the machine learning model, and user input for the machine learning model;”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering data from a user, a machine learning model, and training data, and user input. See MPEP 2106.5(g).
“processing, by the device, the training data and the user input, with the machine learning model, to generate a prediction and an explanation of the prediction;”
The additional elements fall under “apply it” as using a generic computer process the data to generate prediction and an explanation of the prediction. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“providing, by the device, the prediction and the explanation to the user device;”
The additional elements fall under “apply it” as using a generic computer to send the prediction and explanation to the user device. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“receiving, by the device and from the user device, prediction feedback for the prediction and explanation feedback for the explanation;”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by receiving the feedback form the user device. See MPEP 2106.5(g).
“updating, by the device, the machine learning model based on the agreement being achieved between the prediction feedback and the explanation feedback, to generate an updated machine learning model;”
The additional elements fall under “apply it” as using a generic computer to update the machine learning model based on the agreement and generate the new model. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“cryptographically protecting, by the device, the updated machine learning model to generate an updated and cryptographically protected machine learning model; and”
The additional elements fall under “apply it” as using a generic computer to cryptographically protect the updated machine learning model and generate the cryptographically protect model. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“performing, by the device, one or more actions based on the updated and cryptographically protected machine learning model.”
The additional elements fall under “apply it” as using a generic computer to perform one or more actions based on the updated cryptographically protected model. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No - The claim does not include additional elements that are sufficient to amount to a significantly more than the judicial exemption. As an order whole, the claim is directed to receiving data and validate predictions to a threshold. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of receiving, processing, updating, cryptographically protecting and performing fall under using generic computer to apply an exemption. The method does not improve on the function of a computer, transforms an article into another article, nor is it applied by a particular machine, making the claim not patent eligible.
Regarding claim 2:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, further comprising: preventing an update of the machine learning model based on the agreement not being achieved between the prediction feedback and the explanation feedback.”
The additional elements fall under “apply it” as using a generic computer to set higher the more strongly the priority of training examples. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Regarding claim 3:
Step 2A Prong 1:
“The method of claim 1, wherein determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on the threshold comprises: determining whether the prediction feedback and the explanation feedback are coherent;” - The limitations recites a mental process of determining whether the agreement is reached based on a threshold based on coherence (see MPEP 2106.04(a)(2)III).
“calculating an agreement level between the prediction feedback and the explanation feedback based on determining that the prediction feedback and the explanation feedback are coherent;” – The limitations recites a mathematical calculation for an agreement level (see MPEP 2106.04(a)(2)I).
“determining whether the agreement level satisfies the threshold; and” The limitations recites a mental process of determining whether the agreement level satisfies the threshold (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application
Regarding claim 4:
Step 2A Prong 1:
“The method of claim 3, wherein calculating the agreement level between the prediction feedback and the explanation feedback comprises: calculating the agreement level between the prediction feedback and the explanation feedback based on weights assigned to domain experts providing the prediction feedback and the explanation feedback.” - The limitations recites a mathematical calculation for an agreement level based on the wights assigned by the domain experts (see MPEP 2106.04(a)(2)I).
Regarding claim 5:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on the threshold comprises: determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on weights assigned to domain experts providing the prediction feedback and the explanation feedback.”
The additional elements fall under “apply it” as using a generic computer to determine if the agreement is achieved. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Regarding claim 6:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the prediction feedback includes feedback, about the prediction, provided by one or more domain experts, and wherein the explanation feedback includes feedback, about the explanation, provided by one or more domain experts.”
The additional elements fall under “apply it” as using a generic computer to include the feedback by one or more domain experts. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Regarding claim 7:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein updating the machine learning model comprises: updating the machine learning model, to generate the updated machine learning model, based on the prediction feedback and the explanation feedback.”
The additional elements fall under “apply it” as using a generic computer to update the machine learning model based on the feedback. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 8 recites a system and is analogous to the method of claims 1. Therefore, the rejections of claim 1 above applies to claims 8.
Regarding claim 9:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the one or more processors, to cryptographically protect the updated machine learning model to generate the updated and cryptographically protected machine learning model, are configured to: generate a hash based on the prediction feedback and the explanation feedback; and generate a block for the updated machine learning model based on the hash, wherein the block is part of a block chain and corresponds to the updated and cryptographically protected machine learning model.”
The additional elements fall under “apply it” as using a generic computer to generate a hash based on the prediction and explanation feedback and also block for the updated machine learning model based on the hash (see MPEP 2106.05(f)).
The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 10:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the device is a decision support system.”
The additional elements fall under “apply it” as using a generic computer to implement a decision support system. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Regarding claim 11:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: provide the updated and cryptographically protected machine learning model for display; or cause the updated and cryptographically protected machine learning model to be implemented.”
The additional elements fall under “apply it” as using a generic computer to provide the updated cryptographically protected model for display and cause the model to be implemented. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 12:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: retrain the updated and cryptographically protected machine learning model with the training data.”
The additional elements fall under “apply it” as using a generic computer to retrain the model with the training data. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 13:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: receive additional prediction feedback and additional explanation feedback based on the updated and cryptographically protected machine learning model.”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by receiving additional prediction feedback and explanations. See MPEP 2106.5(g).
Regarding claim 14:
Step 2A Prong 2, Step 2B: The additional element(s):
“The device of claim 8, wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: generate a new prediction and a new explanation based on the updated and cryptographically protected machine learning model.”
The additional elements fall under “apply it” as using a generic computer to generate new predictions and explanations based on the cryptographically protected machine learning model. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Claims 15-19 recite a computer readable medium product and is analogous to the method of claims 1-4. Therefore, the rejections of claim 1-4 above applies to claims 15-19.
Claims 20 recite a computer readable medium product and is analogous to the system of claims 9. Therefore, the rejections of claim 9 above applies to claims 20.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 2, 5-11, 13-15, and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Luo et al. (WO2021076139A1) (“Luo”) in view of Belem et al. (US20220114345A1) (“Belem”) and further in view of Nassar, Mohamed, et al. "Blockchain for explainable and trustworthy artificial intelligence." Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 10.1 (2020): e1340, (“Nassar”).
Regarding claim 1 and analogous claims 8 and 15 (Examiner Note: the number is based on the table above), Luo teaches A method, comprising: receiving, by a device and from a user device, [a machine learning model], training data utilized to train the machine learning model, and user input for the machine learning model (Luo para 0035, FIG. 2 illustrates a block diagram of a sample embodiment of a human-model collaborative annotation system 200. In the system 200, unlabeled image data (D1) from a database 210 is provided to experts for labeling using the experts' computer system with annotation software 220. The resulting images labeled by the experts (A1) are provided to a labeled images database 230 to 10 create a database of annotated images [and user input for the machine learning model;]. The labeled images are also used to train a machine learning model of a machine learning system 240 that once trained may, in turn, receive unlabeled images from the database 210 and generate more annotated images for storage in the labeled images database 230. Also, once trained, the collaboratively trained machine learning model of the machine 15 learning system 240 may be deployed for use by non-expert annotators [receiving, by a device and from a user device, training data utilized to train the machine learning model].);
processing, by the device, the training data and the user input, with the machine learning model, to generate a prediction and an explanation of the prediction;
providing, by the device, the prediction and the explanation to the user device (Luo para 0035 line 6-11, The labeled images are also used to train a machine learning model of a machine learning system 240 that once trained may, in turn, receive unlabeled images from the database 210 and generate more annotated images for storage in the labeled images database 230. Also, once trained, the collaboratively trained machine learning model of the machine 15 learning system 240 may be deployed for use by non-expert annotators. para 0036 line 1-12 In sample embodiments, the system 200 further includes an annotation system 250 that selects samples for presentation to non-expert human annotators on display 260 for labeling [processing, by the device, the training data and the user input, with the machine learning model,].
para 0036 line 1-12, In the sample embodiments, the nonexpert human annotators may receive feedback from an annotator training 20 system 270 that compares the annotated images generated by the non-expert human annotators using the annotation system 250 with the corresponding expert annotated images provided by the labeled image database 230. The feedback includes attention maps that teach the non-expert human annotators to learn from their mistakes [to generate a prediction and an explanation of the prediction;]. The evaluator's performance is evaluated, and the next image for 25 annotation on the annotation system 250 by the non-expert human annotator is selected based on the performance of the non-expert human annotator in labeling a previously presented image [providing, by the device, the prediction and the explanation to the user device;]);
Luo does not teach explicitly [receiving, by a device and from a user device,] a machine learning model
receiving, by the device and from the user device, prediction feedback for the prediction and explanation feedback for the explanation;
determining, by the device, whether an agreement is achieved between the prediction feedback and the explanation feedback based on a threshold;
updating, by the device, the machine learning model based on the agreement being achieved between the prediction feedback and the explanation feedback, to generate an updated machine learning model;
cryptographically protecting, by the device, the updated machine learning model to generate an updated and cryptographically protected machine learning model; and
performing, by the device, one or more actions based on the updated and cryptographically protected machine learning model.
However Belem teaches receiving, by the device and from the user device, prediction feedback for the prediction and explanation feedback for the explanation ((Belem para 0034 line 16-35, After receiving transaction information via input 202, machine learning model 204 may infer predictive scores for both a fraud label ( decision task output 208) and semantic concepts associated with fraud patterns ( explanation task output 210). A fraud analyst (e.g., a domain expert) can review the transaction at expert review 212 and indicate whether the fraud label and semantic concepts have been correctly decided by machine learning model 204. In some embodiments, expert review 212 includes a programmed computer system (e.g., computer system 700 of FIG. 7) that the domain expert utilizes to perform expert review. With respect to fraud detection, examples of human feedback 214 include checks on whether ML model determinations ( e.g., yes or no determinations) and/or associated prediction scores ( e.g., on a scale from 0 to 1) associated with semantic concepts such as suspicious billing address, suspicious customer, suspicious payment, suspicious items, high speed ordering, suspicious email, suspicious IP address, etc. are accurate. For example, the domain expert may select "accurate" or "not accurate" in a user interface [prediction feedback for the prediction and explanation feedback for the explanation;]);
determining, by the device, whether an agreement is achieved between the prediction feedback and the explanation feedback based on a threshold (Belem para 0036 line 15-25, In some embodiments, quality control is incorporated into expert review 212. For example, each human expert may be required to meet a minimum accuracy level ( or other relevant quantitative measure). Additionally, different experts may be utilized to review different semantic concepts ( e.g., experts assigned based on their different areas of expertise). Feedback from specific experts (e.g., with higher accuracy levels or other relevant quantitative measures) may be assigned more weight ( e.g., more training weight for higher impact on training of machine learning model 204).);
updating, by the device, the machine learning model based on the agreement being achieved between the prediction feedback and the explanation feedback, to generate an updated machine learning model (Belem para 0034 line 26-31 With respect to fraud detection, examples of human feedback 214 include checks on whether ML model determinations ( e.g., yes or no determinations) and/or associated prediction scores ( e.g., on a scale from 0to 1) associated with semantic concepts such as suspicious billing address, suspicious customer
para 0036 line 15-25, In some embodiments, quality control is incorporated into expert review 212. For example, each human expert may be required to meet a minimum accuracy level ( or other relevant quantitative measure) [updating, by the device, the machine learning model based on the agreement being achieved between the prediction feedback and the explanation feedback]. Additionally, different experts may be utilized to review different semantic concepts ( e.g., experts assigned based on their different areas of expertise). Feedback from specific experts (e.g., with higher accuracy levels or other relevant quantitative measures) may be assigned more weight ( e.g., more training weight for higher impact on training of machine learning model 204).
Para 0039, FIG. 3 is a diagram illustrating examples of approaches for training a multi-task machine learning model to perform both a decision task and an explanation task. In the example illustrated, golden labels 302 and/or noisy labels 304 are used according to one of a plurality of training strategies 306 to train machine learning model 320 to perform the explanation task. In some embodiments, machine learning model 320 is neural network 100 of FIG. 1A and/or machine learning model 204 of FIG. 2. In various embodiments, machine learning model 320 is configured to perform a detection task as well as the explanation task [to generate an updated machine learning model;]);
Luo and Belem are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Luo in view Belem to determine prediction and explanation feedback to the machine learning model. Doing so to continuously integrate human expert feedback into the learning process (Belem para 0034, An advantage of feedback loop 200 is that machine learning model 204 is able to promptly adapt to human teaching (or tuning), as opposed to a unidirectional ML system that directly infLuonces human decisions but does not allow for the reverse of adapting to human behavior. Oftentimes, unidirectional systems are offline and it is only after a certain period of time that a new model is trained and adapted to collected knowledge. Such limitations are solved by incorporating a human-teaching stage that continuously integrates expert feedback into the learning process.).
Nassar teaches [receiving, by a device and from a user device,] a machine learning model
cryptographically protecting, by the device, the updated machine learning model to generate an updated and cryptographically protected machine learning model (Page 9 figure 3,
PNG
media_image1.png
888
866
media_image1.png
Greyscale
(i.e. received a model from the user)
Page 11 3.4.3 Aggregated Decisions para 1, AI/XAI predictors hash and store the metadata of decision outcomes on the IPFS. These decision outcomes include types of decisions, vaLuos of evaluation metrics (e.g., levels of accuracy of classifier), confidence vaLuos, explanations about decisions, and types of explanations. The aggregator SC will compare the reported hashes from predictors and it will determine the correct decisions based on the majority. The aggregator SC will also ensure that the predictors are satisfying the SLA requirements set by the user DApp, and it will report the final result to Al-task SC which will report back to frontend DApp for final feedback or approval [cryptographically protecting, by the device, the updated machine learning model to generate an updated and cryptographically protected machine learning model] (i.e. the model is protected by hashing the metadata of the AI/XAI models).);
performing, by the device, one or more actions based on the updated and cryptographically protected machine learning model (Nassar Page 11 3.4.3 Aggregated Decisions para 1, AI/XAI predictors hash and store the metadata of decision outcomes on the IPFS. These decision outcomes include types of decisions, vaLuos of evaluation metrics (e.g., levels of accuracy of classifier), confidence vaLuos, explanations about decisions, and types of explanations. The aggregator SC will compare the reported hashes from predictors and it will determine the correct decisions based on the majority. The aggregator SC will also ensure that the predictors are satisfying the SLA requirements set by the user DApp, and it will report the final result to Al-task SC which will report back to frontend DApp for final feedback or approval [based on the updated and cryptographically protected machine learning model].
page 11 4.1 Medical Image Diagnosis, Recent research works show the emergence of deep medical image analytics for preventive and precision medicines (Lu & Harrison 2018). Our proposed framework can advance this research by enabling a decentralized disease diagnose system for radiologists who can deploy DApps for consensus-based disease detection. The frontend DApps will input the radiology images whereby AI and XAI predictors from different radiology labs can produce decisions and explanations to help radiologists to reach a more trustworthy, explainable, traceable, and unbiased conclusion. All interested parties including physicians, radiologists, insurance companies, patients, and their care takers can run DApps to perform decentralized predictions and access the same results and explanations [one or more actions]).
Luo and Nassar are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Luo in view Nassar to protect the machine learning model using blockchain. Doing so to provide transparency and visibility of AI decision to all participants and refuse or alter decisions (Nassar page 5 2.7 Integrating Blockchain with AI systems, Blockchain augments decentralized AI systems by enabling an open-source and publicly accessible digital ledger which is distributed among AI agents across peer to peer networks (Nebula AI, 2018). It enables AI agents to collaboratively perform consensus and save new decisions on the blocks which could be traced back and difficult to alter. Blockchain provides transparency and visibility of AI decisions to all participating AI agents on the network hence it becomes difficult for AI agents to alter or refuse the decisions (Hasan & Salah, 2019). In addition, the programmable blockchain platforms enable SCs-based programming models for decentralized AI applications which ensure self-execution of AI agents based on predefined terms and conditions.).
Regarding claim 2 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Belem further teaches preventing an update of the machine learning model based on the agreement not being achieved between the prediction feedback and the explanation feedback (Belem para 0036, In various embodiments, prior to utilizing feedback loop 200, machine learning model 204 is trained to perform the explanation task using a bootstrapping technique (also referred to herein as distant supervision, a weakly supervised technique, semi-supervised technique, etc.) that uses an initial concept-based annotated dataset. At this stage, hyperparameters of machine learning model 204 may be tuned and the resulting model is then deployed in a human teaching stage via feedback loop 200 in which machine learning model 204 outputs decisions and explanations and collects human feedback regarding the outputted decisions and explanations. In various embodiments, after a specified number of human feedback instances, parameters of machine learning model 204 are updated through backpropagation. In some embodiments, quality control is incorporated into expert review 212. For example, each human expert may be required to meet a minimum accuracy level ( or other relevant quantitative measure) [preventing an update of the machine learning model] (Examiner Note: The model only updates with human feedback instances that meet the minimum accuracy and effectively preceding an update of ML model)).
Regarding claim 5 and analogous 18 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Belem teaches wherein determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on the threshold comprises: determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on weights assigned to domain experts providing the prediction feedback and the explanation feedback (Para 0027 line 1-5 In various embodiments, neural network 100 is trained using backpropagation and a gradient descent method. In various embodiments, a joint learning approach attempts to minimize both a decision loss, Ln, and an explanation loss, LE.
Para 0028, Various types of loss functions can be used. Which loss functions to use depends on the nature of the task. As the semantic task corresponds to a multi-labeling task, in some embodiments, a sigmoid function is used and applied to each individual entry of the output before using it in the loss function. To find a mapping that simultaneously satisfies satisfy h:X----;, Y and h:X----;,S for a given input vector, xEX, in some embodiments, the (categorical) cross-entropy is mutually minimized for both tasks. Thus, for an input vector, xEX, a set of domain concepts, sES, and decision labels, yEY, decision task and explanation task loss functions can be formulated as:
PNG
media_image2.png
28
410
media_image2.png
Greyscale
PNG
media_image3.png
34
589
media_image3.png
Greyscale
(Equation 4), respectively. Decision task loss and explanation task loss can be combined into a combined loss in which weights for decision task loss and explanation task loss can be adjusted:
PNG
media_image4.png
29
331
media_image4.png
Greyscale
,
Para 0029 FIG. 1B is a flow diagram illustrating an embodiment of a process for training a machine learning model using distant supervision. In some embodiments, the process of FIG.1B is utilized to train neural network 100 of FIG. 1A, machine learning model 204 of FIG. 2, and/or machine learning model 320 of FIG. 3. In some embodiments, the process of FIG. 1B is performed by computer system 700 of FIG. 7.
Para line 18-25, Additionally, different experts may be utilized to review different semantic concepts ( e.g., experts assigned based on their different areas of expertise). Feedback from specific experts (e.g., with higher accuracy levels or other relevant quantitative measures) may be assigned more weight ( e.g., more training weight for higher impact on training of machine learning model 204). [determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on weights assigned to domain experts providing the prediction feedback and the explanation feedback]).
Regarding claim 6 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Belem further teaches wherein the prediction feedback includes feedback, about the prediction, provided by one or more domain experts, and wherein the explanation feedback includes feedback, about the explanation, provided by one or more domain experts (Belem Fig.2,
PNG
media_image5.png
751
1143
media_image5.png
Greyscale
[wherein the prediction feedback includes feedback, about the prediction,] para 0034 line 16-35, After receiving transaction information via input 202, machine learning model 204 may infer predictive scores for both a fraud label ( decision task output 208) and semantic concepts associated with fraud patterns ( explanation task output 210). A fraud analyst (e.g., a domain expert) can review the transaction at expert review 212 and indicate whether the fraud label and semantic concepts have been correctly decided by machine learning model 204. In some embodiments, expert review 212 includes a programmed computer system (e.g., computer system 700 of FIG. 7) that the domain expert utilizes to perform expert review. With respect to fraud detection, examples of human feedback 214 include checks on whether ML model determinations ( e.g., yes or no determinations) and/or associated prediction scores ( e.g., on a scale from 0 to 1) associated with semantic concepts such as suspicious billing address, suspicious customer, suspicious payment, suspicious items, high speed ordering, suspicious email, suspicious IP address, etc. are accurate. For example, the domain expert may select "accurate" or "not accurate" in a user interface [provided by one or more domain experts, and wherein the explanation feedback includes feedback, about the explanation, provided by one or more domain experts]).
Regarding claim 7 and analogous 19 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Belem teaches wherein updating the machine learning model comprises: updating the machine learning model, to generate the updated machine learning model, based on the prediction feedback and the explanation feedback (Belem para 0034 line 26-31 With respect to fraud detection, examples of human feedback 214 include checks on whether ML model determinations ( e.g., yes or no determinations) and/or associated prediction scores ( e.g., on a scale from 0 to 1) associated with semantic concepts such as suspicious billing address, suspicious customer
para 0036 line 15-25, In some embodiments, quality control is incorporated into expert review 212. For example, each human expert may be required to meet a minimum accuracy level ( or other relevant quantitative measure). Additionally, different experts may be utilized to review different semantic concepts ( e.g., experts assigned based on their different areas of expertise). Feedback from specific experts (e.g., with higher accuracy levels or other relevant quantitative measures) may be assigned more weight ( e.g., more training weight for higher impact on training of machine learning model 204) [based on the prediction feedback and the explanation feedback].
Para 0039, FIG. 3 is a diagram illustrating examples of approaches for training a multi-task machine learning model to perform both a decision task and an explanation task. In the example illustrated, golden labels 302 and/or noisy labels 304 are used according to one of a plurality of training strategies 306 to train machine learning model 320 to perform the explanation task. In some embodiments, machine learning model 320 is neural network 100 of FIG. 1A and/or machine learning model 204 of FIG. 2. In various embodiments, machine learning model 320 is configured to perform a detection task as well as the explanation task [wherein updating the machine learning model comprises: updating the machine learning model, to generate the updated machine learning model,]).
Regarding claim 9 and analogous 20 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claim 8 and 15.
Nassar teaches wherein the one or more processors, to cryptographically protect the updated machine learning model to generate the updated and cryptographically protected machine learning model, are configured to: generate a hash based on the prediction feedback and the explanation feedback; and generate a block for the updated machine learning model based on the hash, wherein the block is part of a block chain and corresponds to the updated and cryptographically protected machine learning model (Nassar page 7,
PNG
media_image6.png
711
995
media_image6.png
Greyscale
[wherein the block is part of a block chain and corresponds to the updated and cryptographically protected machine learning model.)
Page 9 figure 3,
PNG
media_image1.png
888
866
media_image1.png
Greyscale
[and generate a block for the updated machine learning model based on the hash,]
Page 11 3.4.3 Aggregated Decisions para 1, AI/XAI predictors hash and store the metadata of decision outcomes on the IPFS. These decision outcomes include types of decisions, vaLuos of evaluation metrics (e.g., levels of accuracy of classifier), confidence vaLuos, explanations about decisions, and types of explanations. The aggregator SC will compare the reported hashes from predictors and it will determine the correct decisions based on the majority. The aggregator SC will also ensure that the predictors are satisfying the SLA requirements set by the user DApp, and it will report the final result to Al-task SC which will report back to frontend DApp for final feedback or approval [generate a hash based on the prediction feedback and the explanation feedback]).
Regarding claim 10 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo teaches wherein the device is a decision support system (Luo Para 0036 In sample embodiments, the system 200 further includes an annotation system 250 that selects samples for presentation to non-expert human annotators on display 260 for labeling. In the sample embodiments, the nonexpert human annotators may receive feedback from an annotator training system 270 that compares the annotated images generated by the non-expert human annotators using the annotation system 250 with the corresponding expert annotated images provided by the labeled image database 230. The feedback includes attention maps that teach the non-expert human annotators to learn from their mistakes. The evaluator's performance is evaluated, and the next image for annotation on the annotation system 250 by the non-expert human annotator is selected based on the performance of the non-expert human annotator in labeling a previously presented image. Once the non-expert human annotator receives consistently good ratings from the performance evaluation system, the nonexpert human annotator may be added to the pool of expert annotators and permitted to annotate additional images that may be added to the pool of labeled samples in the labeled images database 230. The added images may be weighted with weightings based on the accuracy of the annotations by the non-expert human annotator, thus improving the training model (i.e. a decision support system as it helps a human make better decision for annotations)).
Regarding claim 11 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claim 8 (“7”) and 14 (“15”)
Nassar teaches wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: provide the updated and cryptographically protected machine learning model for display; or cause the updated and cryptographically protected machine learning model to be implemented (Nassar page 7 Figure 2,
PNG
media_image7.png
506
849
media_image7.png
Greyscale
3.3 Blockchain design and framework, In this section, we present our blockchain-based design and framework to achieve explainable and trustworthy AI. The framework is leveraging blockchain SCs to record, govern interactions, and provide consensus for AI predictions and outcomes among AI and XAI oracles. The framework includes also decentralized storage, registration, and reputation support services. Figure 2 presents the overall system architecture and design of our proposed framework. The framework components are grouped into two distinct subsystems: (1) fronted decentralized applications (DApps) and (2) backend decentralized services and platforms.
Page 8 AI layer, AI layer is the primary layer of our proposed framework whereby all data processing and knowledge discovery operations to produce trustworthy, collaborated, and agree-upon decisions are performed. The layer is composed of two different types of predictors; namely: AI predictors (P AU to P AI_n) and XAI predictors (PXAI_l to PXAI_m), as shown in Figure 2. Depending on the configuration received from frontend DApps, AI and XAI nodes either run on raw data and perform all prepossessing operations (such as data cleaning, noise removal, outliers' detection, feature extraction, dimensionality reduction, etc.), or the nodes directly perform decisions on already processed data by inputting into learning models and generating decisions accordingly [or cause the updated and cryptographically protected machine learning model to be implemented]).
Regarding claim 13 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claim 8 (“7”) and 14 (“15”)
Nassar further teaches wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: receive additional prediction feedback and additional explanation feedback based on the updated and cryptographically protected machine learning model ((Page 7 3.2 I Participants, The participants or stakeholders in our proposed solution primary include the following three types:
• Frontend users who are interested in performing a task requiring AI prediction or computation on data (e.g., classification, clustering, regression, etc.). The user may also require explanations about some decisions and provide positive or negative feedback about the decisions.
PNG
media_image8.png
746
1070
media_image8.png
Greyscale
[receive additional prediction feedback and additional explanation feedback based on the updated and cryptographically protected machine learning model] (Examiner Note: The model continuously updates the blockchain as new request)).
Regarding claim 14 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claim 8 (“7”) and 14 (“15”)
Nassar further teaches wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: generate a new prediction and a new explanation based on the updated and cryptographically protected machine learning model (Nassar Page 7 3.2 I Participants, The participants or stakeholders in our proposed solution primary include the following three types:
• Frontend users who are interested in performing a task requiring AI prediction or computation on data (e.g., classification, clustering, regression, etc.). The user may also require explanations about some decisions and provide positive or negative feedback about the decisions.
PNG
media_image8.png
746
1070
media_image8.png
Greyscale
[generate a new prediction and a new explanation based on the updated and cryptographically protected machine learning model]
Page 7 3.3 I Blockchain design and framework
In this section, we present our blockchain-based design and framework to achieve explainable and trustworthy AI. The framework is leveraging blockchain SCs to record, govern interactions, and provide consensus for AI predictions and outcomes among AI and XAI oracles. The framework includes also decentralized storage, registration, and reputation support services. Figure 2 presents the overall system architecture and design of our proposed framework. The framework components are grouped into two distinct subsystems: (1) fronted decentralized applications (DApps) and (2) backend decentralized services and platforms. (Examiner Note: The model continuously updates the blockchain that forms the machine learning model and generates new predictions.).
Claim(s) 3, 4, 16, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Luo in view of Belem and Nassar and further in view of Lee et al., "Towards Efficient Annotations for a Human-AI Collaborative, Clinical Decision Support System: A Case Study on Physical Stroke Rehabilitation Assessment," IUI '22: 27th International Conference on Intelligent User Interfaces, March 2022, 8 Pages (“Lee”).
Regarding claim 3 and analogous 16 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 3 (“2”) and analogous claim 8 (“7”) and 14 (“15”).
Luo does not explicitly teach wherein determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on the threshold comprises:
determining whether the prediction feedback and the explanation feedback are coherent;
calculating an agreement level between the prediction feedback and the explanation feedback based on determining that the prediction feedback and the explanation feedback are coherent;
determining whether the agreement level satisfies the threshold;
and determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on determining whether the agreement level satisfies the threshold.
Belem teaches determining whether the prediction feedback and the explanation feedback are coherent (Belem para 0036, In various embodiments, prior to utilizing feedback loop 200, machine learning model 204 is trained to perform the explanation task using a bootstrapping technique (also referred to herein as distant supervision, a weakly supervised technique, semi-supervised technique, etc.) that uses an initial concept-based annotated dataset. At this stage, hyperparameters of machine learning model 204 may be tuned and the resulting model is then deployed in a human teaching stage via feedback loop 200 in which machine learning model 204 outputs decisions and explanations and collects human feedback regarding the outputted decisions and explanations. In various embodiments, after a specified number of human feedback instances, parameters of machine learning model 204 are updated through backpropagation. In some embodiments, quality control is incorporated into expert review 212. For example, each human expert may be required to meet a minimum accuracy level ( or other relevant quantitative measure) (Examiner Note: The experts must achieve certain level of accuracy for the system to consider their feedback thus they must be coherent));
Lee teaches calculating an agreement level between the prediction feedback and the explanation feedback based on determining that the prediction feedback and the explanation feedback are coherent; determining whether the agreement level satisfies the threshold (Lee page 6,
PNG
media_image9.png
566
1162
media_image9.png
Greyscale
Page 2.2 Technology-Assisted Rehabilitation Assessment para 3, Instead of relying on either a rule-based (RB) model or a machine learning (ML) model alone, we present a human-AI collaborative, clinical decision support system that combines an ML model with an RB model to assess the quality of post-stroke patient's rehabilitation exercises. Initially, when an annotated dataset is not available, this system operates with an RB model to predict the assessment on the quality of motion and identify samples with low confidence scores. Therapists then leveraged these confidence scores to prioritize which samples to annotate. When the annotated dataset is collected, our system trains an ML model and integrates it with the RB model to exploit the strength of both ML and RB models and support human-AI collaborative decision-making [ 42].
Page 8 4.2 Rule-Based Model, For assessing "ROM" performance component, our rules describe whether the estimated target position of each exercise is achieved or not. For example, the assessment of the ROM component for Exercise 1 was specified as follows:
PNG
media_image10.png
86
565
media_image10.png
Greyscale
where Y denotes the predicted label on a performance component. pmax (j, c) indicates the maximum joint position with a joint j (e.g. the wrist ( wr) and the spine shoulder, the top of spine, (spsh)) and the coordinate of a joint position, c in the set C E { Cx, cy, Cz }. This rule monitors whether the maximum position of the post-stroke survivor's wrist joint (pmax ( wr, cy )) exceeds that of the post-stroke survivor's spine shoulder joint (pmax(spsh)) [based on determining that the prediction feedback and the explanation feedback are coherent] in they-coordinate to roughly assess whether a post-stroke survivor achieves the target position of Exercise 1 that requires a post-stroke survivor to bring the post-stroke survivor's wrist to the mouth as if drinking water [calculating an agreement level between the prediction feedback and the explanation feedback]);
and determining whether the agreement is achieved between the prediction feedback and the explanation feedback based on determining whether the agreement level satisfies the threshold ((Page 8 4.2 Rule-Based Model, For assessing "ROM" performance component, our rules describe whether the estimated target position of each exercise is achieved or not. For example, the assessment of the ROM component for Exercise 1 was specified as follows:
PNG
media_image10.png
86
565
media_image10.png
Greyscale
where Y denotes the predicted label on a performance component. pmax (j, c) indicates the maximum joint position with a joint j (e.g. the wrist ( wr) and the spine shoulder, the top of spine, (spsh)) and the coordinate of a joint position, c in the set C E { Cx, cy, Cz }. This rule monitors whether the maximum position of the post-stroke survivor's wrist joint (pmax ( wr, cy )) exceeds that of the post-stroke survivor's spine shoulder joint (pmax(spsh in they-coordinate to roughly assess whether a post-stroke survivor achieves the target position of Exercise 1 that requires a post-stroke survivor to bring the post-stroke survivor's wrist to the mouth as if drinking water.
Page 9
PNG
media_image11.png
618
1149
media_image11.png
Greyscale
[based on determining whether the agreement level satisfies the threshold])).
Luo and Lee are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Luo in view Lee to use threshold defined by the expert to determine an agreement level. Doing so leverage confidence scores to prioritize samples to annotate and exploit the strength of both ML and RB models (Instead of relying on either a rule-based (RB) model or a machine learning (ML) model alone, we present a human-AI collaborative, clinical decision support system that combines an ML model with an RB model to assess the quality of post-stroke patient's rehabilitation exercises. Initially, when an annotated dataset is not available, this system operates with an RB model to predict the assessment on the quality of motion and identify samples with low confidence scores. Therapists then leveraged these confidence scores to prioritize which samples to annotate. When the annotated dataset is collected, our system trains an ML model and integrates it with the RB model to exploit the strength of both ML and RB models and support human-AI collaborative decision-making [ 42].).
Regarding claim 4 and analogous 17 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 (“7”) and 14 (“15”).
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 3 and analogous claim 8 and 15.
Luo and Lee are combine in the same rational as set forth above with respect to claim 3 and analogous claim 16.
Lee teaches calculating the agreement level between the prediction feedback and the explanation feedback based on weights assigned to domain experts providing the prediction feedback and the explanation feedback (Page 6,
PNG
media_image12.png
561
1171
media_image12.png
Greyscale
[calculating the agreement level between the prediction feedback and the explanation feedback]
4.3 Hybrid Model.
A hybrid model integrates two perspectives on assessment using a weighted average ensemble technique [4, 41]: a machine learning model that discovers how to assess the quality of motion from data and a rule-based model from therapists. For the assessment of the quality of motion, a hybrid model (HM) computes a weighted average of prediction scores from two models, in which the contribution of each model is weighted by the performance of a model (i.e. the Fl-score of each model in the range of [0, 1]). The equation to compute the prediction score of an HM (PHM) model is described as follows:
PNG
media_image13.png
59
459
media_image13.png
Greyscale
where PML and PRB indicate the predicted scores of a machine learning (ML) model and a rule-based (RB) model, and Pml and Prb describe the performance, Fl-score of an ML model and an RB model respectively [based on weights assigned to domain experts providing the prediction feedback and the explanation feedback].).
Claim(s) 12 are rejected under 35 U.S.C. 103 as being unpatentable over Luo in view Belem and Nassar and further in view of H. Zhang, P. Gao, J. Yu, J. Lin and N. N. Xiong, "Machine Learning on Cloud With Blockchain: A Secure, Verifiable and Fair Approach to Outsource the Linear Regression," in IEEE Transactions on Network Science and Engineering, vol. 9, no. 6, pp. 3956-3967, 1 Nov.-Dec. 2022 (“Zhang”).
Regarding claim 12 (Examiner Note: The claim number is based on the table above), Luo in view of Belem and Nassar teach the method of claim 1 and analogous claim 8 and 15.
Luo, Belem, Nassar are combine in the same rational as set forth above with respect to claim 1 and analogous claim 8 and 15.
Luo does not explicitly teach wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: retrain the updated and cryptographically protected machine learning model with the training data.
However Zhang teaches wherein the one or more processors, to perform the one or more actions based on the updated and cryptographically protected machine learning model, are configured to: retrain the updated and cryptographically protected machine learning model with the training data (Zhang page 3965, C. Blockchain, Blockchain technology enables secure, trusted, and decentralized autonomous ecosystems for various scenarios. The advanced blockchain technology has been widely leveraged in machine learning to securely and efficiently collect, organzine and audit the extensive quantities of data for model building and accurate prediction [41]–[49]. Li et al. in [42] presented a security mechanics for distributed cloud storage based on blockchain. In their framework, client could distribute all their data into encrypted data blocks and send these data blocks randomly to blockchain network. Juneja et al. in [46] implemented blockchain technology to develop an access control system in which classifier can safely store and access data during retraining in real-time using Stacked Denoising Autoencoders (SDA) networks. Kurtulmus et al. in [48] proposed a blockchainbased model for exchanging machine learning models [retrain the updated and cryptographically protected machine learning model with the training data]).
Luo and Zhang are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Luo in view Zhang to protect the machine learning model using blockchain. Doing so to provide secure, trusted, and decentralize autonomous ecosystems for Machine learning (Zhang page 3965, C. Blockchain, Blockchain technology enables secure, trusted, and decentralized autonomous ecosystems for various scenarios. The advanced blockchain technology has been widely leveraged in machine learning to securely and efficiently collect, organzine and audit the extensive quantities of data for model building and accurate prediction [41]–[49]. Li et al. in [42] presented a security mechanics for distributed cloud storage based on blockchain. In their framework, client could distribute all their data into encrypted data blocks and send these data blocks randomly to blockchain network. Juneja et al. in [46] implemented blockchain technology to develop an access control system in which classifier can safely store and access data during retraining in real-time using Stacked Denoising Autoencoders (SDA) networks. Kurtulmus et al. in [48] proposed a blockchainbased model for exchanging machine learning models).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
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, Michael J. Huntley can be reached at (303) 297-4307. 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.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129