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 .
Status of Claims
The present application is being examined under the claims filed on October 12, 2023.
Claims 1-20 are pending.
Claim Objections
Claims 3, 7, and 10-20 are objected to because of the following informalities:
In claims 3 and 12, “wherein plural data files include clean data files” should read, “where the plurality of data files include clean data files” in view of grammar and antecedent basis.
In claims 7 and 16, “at least two machine learning models of the ensemble machine learning model” should read, “the at least two machine learning models of the ensemble machine learning model” in view of antecedent basis.
In claims 10 and 20, “the at least one training dataset including plural data files” should read, “the at least one training dataset including a plurality of data files” in view of grammar and clarity of antecedent basis.
In claims 10 and 20, “processing the at least one training datasets” should read “processing the at least one training dataset”.
Claims 11-19 are objected to for inheriting the deficiencies of claim 10.
In claim 14, “wherein processing the at least one training dataset with at least two machine learning algorithms for generating at least two machine learning models” should read “wherein the processing of the at least one training dataset with the at least two machine learning algorithms for generating the at least two machine learning models” in view of antecedent basis.
In claim 18, “the plural data files” should read “the plurality of data files” in view of grammar.
In claim 18 “generating at least one adversarial data file” should read, “generating the at least one adversarial data file” in view of antecedent basis.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding Claim 1:
Claim 1 recites “execute at least two machine learning models of an ensemble machine learning model; train at least two machine learning models with the at least one training dataset by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models”
Claim 1 recites executing at least two machine learning models and then training at least two learning models. It is unclear if the models in the training limitation are referring to the models in the execution limitation or if they are referring to another set of models completely. Examiner is interpreting the training limitation to be referring to the execution limitation. Applicant is advised to amend the claim limitation to “train the at least two machine learning models”
Additionally, claim 1 recites “the other of the at least two machine learning models”. There is insufficient antecedent basis for “the other” in the claim. Applicant is advised to amend to clearly delineate between a first model and a second model of the at least two machine learning models.
Regarding Claims 2-9:
Claims 2-9 are rejected to for inheriting the deficiencies of claim 1.
Regarding Claims 2 and 11:
Claims 2 and 11 recite “a runtime input”. Claims 2 and 11 are dependent upon claims 1 and 10 respectively. Claims 1 and 10 also recite “a runtime input”. It is unclear if the runtime inputs from the dependent claims are referring to the runtime input from the independent claims or if they are referring to a separate runtime input entirely. Examiner is interpreting the runtime input from claims 2 and 11 to be referring to the runtime input in claims 1 and 10. Applicant is advised to amend claims 2 and 11 for proper antecedent basis.
Regarding Claim 3:
Claim 3 recites “wherein plural data files include clean data files”. The claim language of claim 3 is written to indicating a narrowing of the “plural data files”, but there is no recitation of “plural data files” earlier in claim 3 or claim 1, which claim 3 depends upon. To further explain, claim 3 corresponds to claim 12, which also recites the same limitation. Claim 12 is dependent upon claim 10, which does recites “the at least one training dataset including plural data files”. On the other hand, claim 3 is dependent upon claim 1, which recites “at least one training dataset”, but does not recite “including plural data files” such as the one found in claim 10. Applicant is advised to amend claim 1 or 3 for proper antecedent basis.
Regarding Claim 4:
Claim 4 recites “wherein at least two machine learning algorithms include convolutional neural networks (CNNs) initialized with random seed weights”. The claim language of claim 4 is written to indicating a narrowing of the “at least two machine learning algorithms”, but there is no recitation of “at least two machine learning algorithms” earlier in claim 4 or claim 1, which claim 4 depends upon. To further explain, claim 4 corresponds to claim 13, which also recites the same limitation. Claim 13 is dependent upon claim 10, which does recites “at least two machine learning algorithms”. On the other hand, claim 4 is dependent upon claim 1, which recites “at least two machine learning models”, but does not recite the algorithms limitation such as the one found in claim 10. Applicant is advised to amend claim 1 or 4 for proper antecedent basis.
Regarding Claim 5:
Claim 5 recites “the at least two machine learning algorithms”. There is insufficient antecedent basis for this limitation in the claim.
Regarding Claim 10:
Claim 10 recites “the other of the at least two machine learning models”. There is insufficient antecedent basis for “the other” in the claim. Applicant is advised to amend to clearly delineate between a first model and a second model of the at least two machine learning models.
Regarding Claims 11-19:
Claims 11-19 are rejected for inheriting the deficiencies of claim 10.
Regarding Claim 18:
Claim 18 recites “
x
’
=
x
-
ϵ
s
i
g
n
∇
x
U
x
,” but
∇
x
is undefined in the claim, making it unclear what
∇
x
is. Although
∇
has a commonly known meaning in mathematics, it is unclear if the claim is adhering to such a meaning or to something else entirely.
Regarding Claim 19:
Claim 19 recites “
L
x
,
x
'
=
L
l
x
+
β
L
U
x
,
x
'
,” but
x
and
x
'
are undefined in the claim, making it unclear what
x
and
x
'
is. In view of claim 18 above,
x
is interpreted as a data file, and
x
'
is interpreted as an adversarial data file.
Claim 19 further recites “
L
U
x
,
x
'
=
∆
-
x
,
x
'
-
ϵ
-
2
,” but
∆
-
is undefined in the claim, making it unclear what
∆
-
is. Although
∆
has a commonly known meaning in mathematics, it is unclear if the claim is adhering to such a meaning or to something else entirely.
Regarding Claim 20:
Claim 20 recites “the other of the at least two machine learning models”. There is insufficient antecedent basis for “the other” in the claim. Applicant is advised to amend to clearly delineate between a first model and a second model of the at least two machine learning models.
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 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1-9 are directed to a system [machine].
Claims 10-19 are directed to a method [process].
Claim 20 is directed to a non-transitory computer readable medium [machine].
Regarding Claim 1:
Step 2A, Prong 1: The following limitations are directed to the abstract idea of a mental process [see MPEP 2106.04(a)(2) III. C.]. In particular, the claim recites mental processes that are concepts performed in the human mind or with pen and paper (including an observation, evaluation, judgement, or opinion).
modify at least one training dataset based on mutual information extracted from an ensemble machine learning model to provide at least one adversarial training dataset having a modified measure of epistemic uncertainty
As drafted, under their broadest reasonable interpretation (BRI), in view of the specification, the above limitations cover concepts performed in the human mind (observation, evaluation, judgement, or opinion). Given a sufficiently small set of data, nothing in the claim prohibits this process from being performed mentally or with pen and paper.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
A system configured for automated data processing, that executes a machine learning model configured to detect out-of-distribution samples, the system comprising:
and at least one processor programmed or configured to:
execute at least two machine learning models of an ensemble machine learning model
train at least two machine learning models with the at least one training dataset by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models
train the ensemble machine learning model with the at least one adversarial training dataset to provide a trained ensemble machine learning model
provide the runtime input to the trained ensemble machine learning model to generate a signal output indicating that the runtime input to the trained ensemble machine learning model includes an out-of-distribution sample
The following additional elements are directed to insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(g)].
at least one data storage device configured to store at least one training dataset
receive a runtime input from a client device
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
A system configured for automated data processing, that executes a machine learning model configured to detect out-of-distribution samples, the system comprising:
and at least one processor programmed or configured to:
execute at least two machine learning models of an ensemble machine learning model
train at least two machine learning models with the at least one training dataset by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models
train the ensemble machine learning model with the at least one adversarial training dataset to provide a trained ensemble machine learning model
provide the runtime input to the trained ensemble machine learning model to generate a signal output indicating that the runtime input to the trained ensemble machine learning model includes an out-of-distribution sample
The following additional elements are directed to receiving or transmitting data over a network. The courts (as per Intellectual Ventures v. Symantec, 838 F.3d 1307, 1321; 120 USPQ2d 1353, 1362 (Fed. Cir. 2016)) have recognized receiving or transmitting data over a network as well-understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(d) II.].
receive a runtime input from a client device
The following additional elements are directed to storing and retrieving information in memory. The courts (as per Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015)) have recognized storing and retrieving information in memory as well-understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(d) II.].
at least one data storage device configured to store at least one training dataset
Regarding Claim 2:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements remain directed to insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(g)].
wherein a runtime input includes one or more of an adversarially perturbed sample, a garbage sample, and/or an adversarially perturbed garbage sample
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements remain directed to receiving or transmitting data over a network. The courts (as per Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362) have recognized receiving or transmitting data over a network as well-understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(d) II.].
wherein a runtime input includes one or more of an adversarially perturbed sample, a garbage sample, and/or an adversarially perturbed garbage sample
Regarding Claim 3:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein plural data files include clean data files
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein plural data files include clean data files
Regarding Claim 4:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein at least two machine learning algorithms include convolutional neural networks (CNNs) initialized with random seed weights
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein at least two machine learning algorithms include convolutional neural networks (CNNs) initialized with random seed weights
Regarding Claim 5:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein the at least one processor, as configured to train at least two machine learning models with at least one training dataset, is programmed or configured to: drop out a node in the at least two machine learning algorithms
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein the at least one processor, as configured to train at least two machine learning models with at least one training dataset, is programmed or configured to: drop out a node in the at least two machine learning algorithms
Regarding Claim 6:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein the trained ensemble machine learning model includes a detection rate of false positives no greater than one percent
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein the trained ensemble machine learning model includes a detection rate of false positives no greater than one percent
Regarding Claim 7:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1. Additionally,
The following limitations are/remain directed to the abstract idea of a mental process [see MPEP 2106.04(a)(2) III. C.]. In particular, the claim recites mental processes that are concepts performed in the human mind (including an observation, evaluation, judgement, or opinion).
wherein mutual information…is based on a Jensen-Shannon divergence
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
…between at least two machine learning models of the ensemble machine learning model…
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
…between at least two machine learning models of the ensemble machine learning model…
Regarding Claim 8:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1. Additionally,
The following limitations are/remain directed to the abstract idea of a mental process [see MPEP 2106.04(a)(2) III. C.]. In particular, the claim recites mental processes that are concepts performed in the human mind (including an observation, evaluation, judgement, or opinion).
(a) …modify at least one training dataset based on mutual information…
determining an attack magnitude by randomly sampling an attack magnitude from a uniform distribution
(c) generating at least one adversarial data file based on at least one data file of plural datafiles in at least one training dataset, the attack magnitude, and the mutual information…
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
(a) wherein the at least one processor, as configured to…is programmed or configured to:
(c) …between at least two machine learning models of the ensemble machine learning model
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
…between at least two machine learning models of the ensemble machine learning model…
Regarding Claim 9:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 1. Additionally,
The following limitations are/remain directed to the abstract idea of a mental process [see MPEP 2106.04(a)(2) III. C.]. In particular, the claim recites mental processes that are concepts performed in the human mind (including an observation, evaluation, judgement, or opinion).
determine a cross-entropy loss based on model predictions and ground truth values
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein the at least one processor, as configured to train at least two machine learning models with the at least one training dataset, is programmed or configured to:
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein the at least one processor, as configured to train at least two machine learning models with the at least one training dataset, is programmed or configured to:
Regarding Claim 10:
Step 2A, Prong 1: The following limitations are directed to the abstract idea of a mental process [see MPEP 2106.04(a)(2) III. C.]. In particular, the claim recites mental processes that are concepts performed in the human mind or with pen and paper (including an observation, evaluation, judgement, or opinion).
(a) processing the at least one training datasets…
perturbing at least one data file of the at least one training dataset with mutual information extracted from the ensemble machine learning model to provide an adversarial training dataset
As drafted, under their broadest reasonable interpretation (BRI), in view of the specification, the above limitations cover concepts performed in the human mind (observation, evaluation, judgement, or opinion). Given a sufficiently small set of data, nothing in the claim prohibits this process from being performed mentally or with pen and paper.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
A computer-implemented method for automated data processing, with a machine learning model configured to detect out-of-distribution samples, the method comprising:
(a) …with at least two machine learning algorithms executed by the processor for generating at least two machine learning models
wherein the at least two machine learning models are trained by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models to form an ensemble machine learning model
inputting the adversarial training dataset into the ensemble machine learning model for the processor to train the ensemble machine learning model
generating a signal output by causing the processor to execute a trained ensemble machine learning model with a runtime input, wherein the signal output indicates that the runtime input includes an out-of-distribution sample
The following additional elements are directed to insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(g)].
receiving, as an input to a processor, at least one training dataset, the at least one training dataset including plural data files
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
A computer-implemented method for automated data processing, with a machine learning model configured to detect out-of-distribution samples, the method comprising:
(a) …with at least two machine learning algorithms executed by the processor for generating at least two machine learning models
wherein the at least two machine learning models are trained by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models to form an ensemble machine learning model
inputting the adversarial training dataset into the ensemble machine learning model for the processor to train the ensemble machine learning model
generating a signal output by causing the processor to execute a trained ensemble machine learning model with a runtime input, wherein the signal output indicates that the runtime input includes an out-of-distribution sample
The following additional elements are directed to receiving or transmitting data over a network. The courts (as per Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362) have recognized receiving or transmitting data over a network as well-understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity to the judicial exception [see MPEP 2106.05(d) II.].
receiving, as an input to a processor, at least one training dataset, the at least one training dataset including plural data files
Regarding Claims 11-17:
Claims 11-17 correspond to claims 2-8. In particular, 11:2, 12:3, 13:4, 14:5, 15:6, 16:7, 17:8.
Step 2A, Prong 1: Claims 11-17 recite the same abstract ideas as in claims 2-8.
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application. The analysis of claims 11-17 at this step mirror that of claims 2-8.
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception. The analysis of claims 11-17 at this step mirror that of claims 2-8.
Regarding Claim 18:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 17. Additionally,
The following limitations are/remain directed to the abstract idea of a mathematical concept (i.e., mathematical relationships, mathematical formulas or equations, or mathematical calculations) [see MPEP 2106.04(a)(2) I.C.].
wherein generating at least one adversarial data file is defined by:
x
’
=
x
-
ϵ
s
i
g
n
∇
x
U
x
where
x
’
is the adversarial data file,
x
is a data file of the plural data files,
ϵ
is the attack magnitude, and
U
x
is mutual information between at least two machine learning models of the ensemble machine learning model
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
Regarding Claim 19:
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 10. Additionally,
The following limitations are/remain directed to the abstract idea of a mathematical concept (i.e., mathematical relationships, mathematical formulas or equations, or mathematical calculations) [see MPEP 2106.04(a)(2) I.C.].
an uncertainty loss defined by:
L
x
,
x
'
=
L
l
x
+
β
L
U
x
,
x
'
where
L
x
,
x
'
is the uncertainty loss,
L
l
x
is the cross-entropy loss,
β
is a weighting factor, and
L
U
x
,
x
'
is an uncertainty proportion defined by:
L
U
x
,
x
'
=
∆
-
x
,
x
'
-
ϵ
-
2
ϵ
-
=
ϵ
ϵ
m
a
x
and
ϵ
is an attack magnitude
Step 2A, Prong 2: There are no additional elements in this claim that integrate the judicial exception into a practical application.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
wherein inputting an adversarial training dataset into the ensemble machine learning model includes determining…
Step 2B: There are no additional elements in this claim that amount to significantly more than the judicial exception.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
wherein inputting an adversarial training dataset into the ensemble machine learning model includes determining…
Regarding Claim 20:
Claim 20 corresponds to claim 10.
Step 2A, Prong 1: This claim recites the same abstract ideas as in claim 10.
Step 2A, Prong 2: This claim recites the same additional elements as in claim 10. There are no additional elements in this claim that integrate the judicial exception into a practical application. The analysis of this claim at this step mirror that of claim 10, with the exception the following limitations.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to integrate the judicial exception into a practical application.
A non-transitory computer readable medium encoded with program code for automated data processing with a machine learning model configured to detect out-of- distribution samples, when placed in communicable contact with a computer processor, the program code causing the processor to be configured to perform an operation comprising:
Step 2B: This claim recites the same additional elements as in claim 10. There are no additional elements in this claim that amount to significantly more than the judicial exception. The analysis of this claim at this step mirror that of claim 10, with the exception the following limitations.
The following additional elements are adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea [see MPEP 2106.05(f)] and therefore fails to amount to significantly more than the judicial exception.
A non-transitory computer readable medium encoded with program code for automated data processing with a machine learning model configured to detect out-of- distribution samples, when placed in communicable contact with a computer processor, the program code causing the processor to be configured to perform an operation comprising:
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 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 1-3 and 8-9 are rejected under 35 U.S.C. 103 as being unpatentable over Wu et al. (US 20220245422), hereinafter Wu, in view of Goldenberg et al. (US 11636286), hereinafter Goldenberg, further in view of Chen et al. (“Robust Out-of-distribution Detection for Neural Networks”), hereinafter Chen, and further in view of Munoz et al. (US 20210089895), hereinafter Munoz et al.
Regarding Claim 1:
Wu discloses:
A system configured for automated data processing, that executes a machine learning model configured to detect out-of-distribution samples, the system comprising:
Wu, [0002], “The present disclosure describes systems and methods for machine learning architecture for out-of-distribution data set detection.”
at least one data storage device configured to store at least one training dataset
Wu, [0072], “The system 200 may include a data storage 214. In some embodiments, the data storage 214 may be a secure data store. In some embodiments, the data storage 214 may store input data sets, such as image data, training data sets, or the like.”
and at least one processor programmed or configured to:
Wu, [0067], “The system 200 includes a processor 202 configured to execute processor-readable instructions that, when executed, configure the processor 202 to conduct operations described herein. For example, the system 200 may be configured to conduct operations for out-of-distribution data set detection.”
receive a runtime input from a client device
Wu, [0111], “At operation 502, the processor may receive an input data set. In some embodiments, the input data set may be an image data set. In some embodiments, the input data set may be a data set including alphanumeric data, such as textual data. In some embodiments, the input data set may be received from the client device 210 (FIG. 2).”
provide the runtime input to the trained…machine learning model to generate a signal output indicating that the runtime input to the trained ensemble machine learning model includes an out-of-distribution sample
Wu, [0112], “At operation 504, the processor may generate an out-of-distribution prediction based on the input data set and an auto-encoder model…The auto-encoder model may be trained based on reducing a reconstruction error such that the random transformation may be substantially cancelled by a decoder of the auto-encoder network.”
[0119], “At operation 506, the processor may generate a signal for providing an indication of whether the input data set is an out-of-distribution data set.”
In para. 112, Wu discloses using the input data with the autoencoder model [provide the runtime input to the trained…machine learning model] to generating a prediction [generate a signal output]. Further in para. 119, Wu discloses the processor generates the signal, interpreted as the model’s prediction, indicating whether the input data is an out-of-distribution data [a signal output indicating that the runtime input to the trained ensemble machine learning model includes an out-of-distribution sample].
Wu does not explicitly disclose:
modify at least one training dataset based on mutual information extracted from an ensemble machine learning model to provide at least one adversarial training dataset having a modified measure of epistemic uncertainty
execute at least two machine learning models of an ensemble machine learning model
train at least two machine learning models with the at least one training dataset by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models
train the ensemble machine learning model with the at least one adversarial training dataset to provide a trained ensemble machine learning model
However, in the same field, analogous art Goldenberg teaches:
execute at least two machine learning models of an ensemble machine learning model
train at least two machine learning models with the at least one training dataset by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models
Goldenberg, [19], “the ML models 102-1 through 102-N of an ensemble 102 are trained to each output a probability vector 103-1, 103-2, 103-3, through 103-N, respectively, indicating the probability of an input data sample 100 belonging to one of K known classes…”
PNG
media_image1.png
631
871
media_image1.png
Greyscale
Goldenberg discloses training an ensemble of models [execute at least two machine learning models of an ensemble machine learning model] using input data sample. FIG. 1 in view of para. 19, Goldenberg teaches that the same input sample is input to each of the models [train at least two machine learning models with the at least one training dataset by feeding an input…of one of the at least two machine learning models to the other of the at least two machine learning models]
Wu, Goldenberg, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu with Goldenberg to use ensembles of machine learning models in order to increase the accuracy predictions. “Machine learning models are being used for various types of projects. Additionally, ensembles of models, which are groups of two or more machine learning models, have been used recently to boost accuracy of a machine learning system. It is generally accepted that diversity between ensemble models is important for leveraging the strength of ensemble approaches” (Goldenberg, [1]).
Wu and Goldenberg do not explicitly disclose:
modify at least one training dataset based on mutual information extracted from an ensemble machine learning model to provide at least one adversarial training dataset…
train the ensemble machine learning model with the at least one adversarial training dataset to provide a trained ensemble machine learning model
However, in the same field, analogous art Chen teaches:
modify at least one training dataset based on mutual information extracted from an ensemble machine learning model to provide at least one adversarial training dataset…
Chen, p. 4, col. 1, “For in-distribution inputs
x
∈
P
X
, ALOE [Adversarial Learning with inliner and Outlier Exposure] creates adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise.”
Chen teaches creating an adversarial inlier for the inputs
x
[modify at least one training dataset]. The inputs are disclosed to be in-distribution inputs [based on mutual information extracted from an ensemble machine learning model to provide at least one adversarial training dataset…].
train the…machine learning model with the at least one adversarial training dataset to provide a…ensemble machine learning model
As cited above, Chen teaches training the model [provide a…ensemble machine learning model] with the perturbed examples from the in-distribution [train the…machine learning model with the at least one adversarial training dataset].
Wu, Goldenberg, Chen, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu and Goldenberg with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Wu, Goldenberg, and Chen do not explicitly disclose:
…having a modified measure of epistemic uncertainty
However, in the same field, analogous art Munoz teaches:
…having a modified measure of epistemic uncertainty
Munoz, [0178], “a candidate counterfactual data sample is generated and modified based on an optimization process of a loss function, wherein the loss function comprises a term to take into account the predictive uncertainty of the neural network, in other words which takes into account the aleatoric and/or epistemic uncertainty contained in the class prediction of the neural network.”
Munoz teaches a term that takes into account the epistemic uncertainty [having a modified measure of epistemic uncertainty].
Wu, Goldenberg, Chen, Munoz, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Munoz to include epistemic uncertainty in order to in order to further increase the robustness of the machine learning system. “According to an example embodiment of the present invention, a counterfactual generation method and system are provided, which allow adversarial effects to be reduced/mitigated by employing an uncertainty regularizing term to a counterfactual generator” (Munoz, [0113]). Munoz discloses that by introducing/employing the uncertainty regularizing term, the system is able to mitigate adversarial effects.
Regarding Claim 2:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system claim 1, and Chen further discloses:
wherein a runtime input includes one or more of an adversarially perturbed sample, a garbage sample, and/or an adversarially perturbed garbage sample
Chen, p. 3, col. 1, “In testing, the detector
G
is evaluated on perturbed inputs drawn from a mixture distribution…”
p. 3, col. 2, “Specifically, we construction adversarial test examples by adding small perturbations in
B
(
x
,
ϵ
)
”
Chen discloses testing their detector on adversarial perturbed test example inputs [wherein a runtime input includes one or more of an adversarially perturbed sample…].
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz, further with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Regarding Claim 3:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system claim 1, and Wu further discloses:
wherein plural data files include clean data files
Wu, [0106], “In scenarios where a set of OOD example data values are available, auto-encoder models may include operations for training a linear classifier (e.g., Logistic Regression) on learned latent representations of both in-distribution and out-of-distribution data to predict in-distribution probability for input data sets.”
Wu discloses input data sets that include in-distribution data [plural data files include clean data files].
Regarding Claim 8:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system claim 1, and Chen further discloses:
wherein the at least one processor, as configured to modify at least one training dataset based on mutual information, is programmed or configured to: determining an attack magnitude by randomly sampling an attack magnitude from a uniform distribution
Chen, p. 4, col. 1, “For in-distribution inputs
x
∈
P
X
, ALOE [Adversarial Learning with inliner and Outlier Exposure] creates adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise.”Chen teaches creating an adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood for the inputs
x
[determining an attack magnitude by randomly sampling an attack magnitude from a uniform distribution].
generating at least one adversarial data file based on at least one data file of plural datafiles in at least one training dataset, the attack magnitude, and the mutual information…
As cited above, Chen teaches the adversarial inlier is created [generating at least one adversarial data file based on] using the input [at least one data file of the plural data files], the
ϵ
-ball [the attack magnitude], and the negative log likelihood [the mutual information].
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz further with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Regarding Claim 9:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system claim 1, and Goldenberg further discloses:
wherein the at least one processor, as configured to train at least two machine learning models with the at least one training dataset, is programmed or configured to: determine a cross-entropy loss based on model predictions and ground truth values
Goldenberg, [36]-[37], “the ground-truth class for a data sample
x
is defined as
q
(
x
)
∈
{
1
,
2
,
.
.
.
,
K
}
. The individual loss (the loss used to train each individual ensemble model
f
i
) may be denoted as
L
(
x
)
=
L
(
f
i
(
x
)
,
q
(
x
)
)
. If ensemble and individual ML model outputs are of the same form, the same loss
L
may be applied to the ensemble output to yield a joint ensemble loss:
L
j
o
i
n
t
(
x
)
=
L
(
M
(
x
)
,
q
(
x
)
)
An example individual loss function for classification problems is the cross-entropy loss:
L
C
E
x
=
-
∑
j
=
1
K
q
(
x
)
log
f
i
j
x
For identification problems, the cross-entropy loss may be combined with additional loss components.”
Goldenberg teaches a cross-entropy loss that uses the ground-truth and the model outputs.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz further with Chen to increase the accuracy predictions. “Machine learning models are being used for various types of projects. Additionally, ensembles of models, which are groups of two or more machine learning models, have been used recently to boost accuracy of a machine learning system. It is generally accepted that diversity between ensemble models is important for leveraging the strength of ensemble approaches” (Goldenberg, [1]).
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Goldenberg, and Chen in view of Munoz as applied to claim 1 above, and further in view of Ma et al. (US 20230306723), hereinafter Ma.
Regarding Claim 4:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system claim 1, and Goldenberg further discloses:
wherein at least two machine learning algorithms include convolutional neural networks (CNNs)…
Goldenberg, [86], “Each server may also include one or more machine learning models 1170 of an ensemble of machine learning models, such as a CNN.”
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz further with Goldenberg to use CNNs in their ensembles of machine learning models in order to increase the accuracy predictions. “Instead, during the training process of a machine learning model, the machine learning model adaptively makes its own determinations as to how to modify each computation, convolution or transformation of a given processing node to produce better and/or superior results from the input values it receives” (Goldenberg, [32]).
Wu, Goldenberg, Chen, and Munoz do not explicitly disclose:
…initialized with random seed weights
However, in the same field, analogous art Ma teaches:
…initialized with random seed weights
Ma, [0051], “Preliminary analysis shows that in random initialization settings (e.g., initializing from scratch), transformers lag behind CNNs on medical target tasks with limited annotated data.”
Ma teaches randomly initializing CNNs [initialized with random seed weights].
Wu, Goldenberg, Chen, Munoz, Ma, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz with Ma to use random initialization in order to achieve better performance. “Firstly, in random initialization (scratch) settings (horizontal lines), transformers (e.g., ViT-B and/or Swin-B) cannot compete with CNNs (e.g., such as ResNet-50) in medical applications, as they offer performance equal to or even worse than CNNs. This inferior performance is attributable to the respective transformer’s lack of desirable inductive bias in comparison to CNNs, which has a negative impact on transformer performance on medical target tasks with limited annotated data” (Ma, [0053]).
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Goldenberg, and Chen in view of Munoz as applied to claim 1 above, and further in view of Georgescu et al. (US 20230099938), hereinafter Georgescu.
Regarding Claim 5:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system of claim 1, but do not explicitly disclose:
wherein the at least one processor, as configured to train at least two machine learning models with at least one training dataset, is programmed or configured to: drop out a node in the at least two machine learning algorithms
However, in the same field, analogous art Georgescu teaches:
wherein the at least one processor, as configured to train at least two machine learning models with at least one training dataset, is programmed or configured to: drop out a node in the at least two machine learning algorithms
Georgescu, [0082], “In particular, convolutional neural networks 900 can be trained based on the backpropagation algorithm. For preventing overfitting, methods of regularization can be used, e.g. dropout of nodes 912-920, stochastic pooling, use of artificial data, weight decay based on the L1 or the L2 norm, or max norm constraints.”
Georgescu teaches that during training of the neural networks [train at least two machine learning models with at least one training dataset], dropout of nodes is used to prevent overfitting [drop out a node in the at least two machine learning algorithms].
Wu, Goldenberg, Chen, Munoz, Georgescu, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz with Georgescu to use dropout in order to increase the robustness of the model. “In particular, convolutional neural networks 900 can be trained based on the backpropagation algorithm. For preventing overfitting, methods of regularization can be used, e.g. dropout of nodes 912-920, stochastic pooling, use of artificial data, weight decay based on the L1 or the L2 norm, or max norm constraints” (Georgescu, [0082]). Georgescu discloses that dropout prevents overfitting of the neural network, allowing for a more robustness neural network that can handle more generalized data.
Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Goldenberg, and Chen in view of Munoz as applied to claim 1 above, and further in view of Taniguchi et al. (US 20220131884), hereinafter Taniguchi.
Regarding Claim 6:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system of claim 1, but do not explicitly disclose:
wherein the trained ensemble machine learning model includes a detection rate of false positives no greater than one percent
However, in the same field, analogous art Taniguchi teaches:
wherein the trained…machine learning model includes a detection rate of false positives no greater than one percent
Taniguchi, [0226], “In addition, as illustrated in the detection result management table 541, the false positive rate for the legitimate domain is equal to or less than the false positive threshold value even when the feature ‘freshness’ is utilized. The false positive threshold value is, for example, 1%.”
Wu, Goldenberg, Chen, Munoz, Taniguchi, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz with Taniguchi to use a false positive threshold value rate in order to increase the robustness of the model. “The false positive threshold value is, for example, 1%. Therefore, even if the feature ‘freshness’ is utilized to detect the behavior of the wide-area malicious domain, it is considered possible to avoid an incident in which the behavior of the legitimate domain is erroneously determined as the behavior of the malicious domain.” (Taniguchi, [0226]). Taniguchi discloses that the false positive threshold value makes it possible to avoid scenarios where a non-malicious prediction is predicted to be malicious.
Claim 7 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Goldenberg, and Chen in view of Munoz as applied to claim 1 above, and further in view of Al-Turki et al. (US 10839269), hereinafter Al-Turki.
Regarding Claim 7:
As discussed above, Wu, Goldenberg, Chen, and Munoz teach [the] system of claim 1, but do not explicitly disclose:
wherein mutual information between at least two machine learning models of the ensemble machine learning model is based on a Jensen-Shannon divergence
However, in the same field, analogous art Al-Turki teaches:
wherein mutual information between at least two machine learning models of the…machine learning model is based on a Jensen-Shannon divergence
Al-Turki, [33], “JSD is the Jensen-Shannon divergence describing the distance between two distributions. It is in the range of 0 and 1, and equals 0 when two distributions are same.”
Wu, Goldenberg, Chen, Munoz, Al-Turki, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, Chen, and Munoz with Al-Turki to use Jensen-Shannon divergence in order to increase the robustness of the model by achieving adversarial domain adapation. “JSD is the Jensen-Shannon divergence describing the distance between two distributions. It is in the range of 0 and 1, and equals 0 when two distributions are same. Namely, when the distributions of source and target data are identical, l*=−2 log 2 is the global optimum of l(X.sub.S, X.sub.T, M.sub.T). In this case, the goal of adversarial domain adaptation is well achieved.” (Al-Turki, [33]).
Claims 10-12, 17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Wu in view of Goldenberg, and further in view of Chen.
Regarding Claim 10:
Wu discloses:
A computer-implemented method for automated data processing, with a machine learning model configured to detect out-of-distribution samples, the method comprising:
Wu, [0002], “The present disclosure describes systems and methods for machine learning architecture for out-of-distribution data set detection.”
receiving, as an input to a processor, at least one training dataset, the at least one training dataset including plural data files
Wu, [0111], “At operation 502, the processor may receive an input data set. In some embodiments, the input data set may be an image data set. In some embodiments, the input data set may be a data set including alphanumeric data, such as textual data.”
Wu discloses the processor [as an input to a processor] receiving an input data set [receiving…at least one training dataset], and that the data set may include images of textual data [the at least one training dataset including plural data files].
generating a signal output by causing the processor to execute a trained…machine learning model with a runtime input, wherein the signal output indicates that the runtime input includes an out-of-distribution sample
Wu, [0112], “At operation 504, the processor may generate an out-of-distribution prediction based on the input data set and an auto-encoder model…The auto-encoder model may be trained based on reducing a reconstruction error such that the random transformation may be substantially cancelled by a decoder of the auto-encoder network.”
[0119], “At operation 506, the processor may generate a signal for providing an indication of whether the input data set is an out-of-distribution data set.”
In para. 112, Wu discloses generating a prediction [generating a signal output] using an autoencoder model [causing the processor to execute a trained…machine learning model] and an input data set [with a runtime input]. Further in para. 119, Wu discloses the processor generates the signal, interpreted as the model’s prediction, indicating whether the input data is an out-of-distribution data [wherein the signal output indicates that the runtime input includes an out-of-distribution sample].
Wu does not explicitly disclose:
processing the at least one training datasets with at least two machine learning algorithms executed by the processor for generating at least two machine learning models, wherein the at least two machine learning models are trained by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models to form an ensemble machine learning model
perturbing at least one data file of the at least one training dataset with mutual information extracted from the ensemble machine learning model to provide an adversarial training dataset
inputting the adversarial training dataset into the ensemble machine learning model for the processor to train the ensemble machine learning model
However, in the same field, analogous art Goldenberg teaches:
processing the at least one training datasets with at least two machine learning algorithms executed by the processor for generating at least two machine learning models, wherein the at least two machine learning models are trained by feeding an input or output of one of the at least two machine learning models to the other of the at least two machine learning models to form an ensemble machine learning model
Goldenberg, [19], “the ML models 102-1 through 102-N of an ensemble 102 are trained to each output a probability vector 103-1, 103-2, 103-3, through 103-N, respectively, indicating the probability of an input data sample 100 belonging to one of K known classes…”
PNG
media_image1.png
631
871
media_image1.png
Greyscale
Goldenberg discloses training an ensemble of models [with at least two machine learning algorithms executed by the processor for generating at least two machine learning models] using input data sample [processing the at least one training datasets]. FIG. 1 in view of para. 19, Goldenberg teaches that the same input sample is input to each of the models [wherein the at least two machine learning models are trained by feeding an input…of the least two machine learning models to the other of the at least two machine learning models to form an ensemble learning model.]
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu with Goldenberg to use ensembles of machine learning models in order to increase the accuracy predictions. “Machine learning models are being used for various types of projects. Additionally, ensembles of models, which are groups of two or more machine learning models, have been used recently to boost accuracy of a machine learning system. It is generally accepted that diversity between ensemble models is important for leveraging the strength of ensemble approaches” (Goldenberg, [1]).
Wu and Goldenberg do not explicitly disclose:
perturbing at least one data file of the at least one training dataset with mutual information extracted from the ensemble machine learning model to provide an adversarial training dataset
inputting the adversarial training dataset into the ensemble machine learning model for the processor to train the ensemble machine learning model
However, in the same field, analogous art Chen teaches:
perturbing at least one data file of the at least one training dataset with mutual information1 extracted from the…machine learning model to provide an adversarial training dataset
Chen, p. 4, col. 1, “For in-distribution inputs
x
∈
P
X
, ALOE [Adversarial Learning with inliner and Outlier Exposure] creates adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise.”
Chen teaches creating an adversarial inlier for the inputs
x
[perturbing at least one data file of the at least one training dataset]. The inputs are disclosed to be in-distribution inputs [with mutual information extracted from the ensemble machine learning model].
inputting the adversarial training dataset into the…machine learning model for the processor to train the…machine learning model
As cited above, Chen teaches training the model [for the processor to train…the machine learning model] with the perturbed examples from the in-distribution [inputting the adversarial training dataset into the…machine learning model].
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu and Goldenberg with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Regarding Claim 11:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, and Chen further discloses:
wherein a runtime input includes one or more of an adversarially perturbed sample, a garbage sample, or an adversarially perturbed garbage sample
Chen, p. 3, col. 1, “In testing, the detector
G
is evaluated on perturbed inputs drawn from a mixture distribution…”
p. 3, col. 2, “Specifically, we construction adversarial test examples by adding small perturbations in
B
(
x
,
ϵ
)
”
Chen discloses testing their detector on adversarial perturbed test example inputs [wherein a runtime input includes one or more of an adversarially perturbed sample…].
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen, further with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Regarding Claim 12:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, and Wu further discloses:
wherein plural data files include clean data files
Wu, [0106], “In scenarios where a set of OOD example data values are available, auto-encoder models may include operations for training a linear classifier (e.g., Logistic Regression) on learned latent representations of both in-distribution and out-of-distribution data to predict in-distribution probability for input data sets.”
Wu discloses input data sets that include in-distribution data [plural data files include clean data files].
Regarding Claim 17:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, and Chen further discloses:
wherein perturbing at least one data file of at least one dataset includes: determining an attack magnitude by randomly sampling an attack magnitude from a uniform distribution
Chen, p. 4, col. 1, “For in-distribution inputs
x
∈
P
X
, ALOE [Adversarial Learning with inliner and Outlier Exposure] creates adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise.”
Chen teaches creating an adversarial inlier within the
ϵ
-ball that maximize the negative log likelihood for the inputs
x
[determining an attack magnitude by randomly sampling an attack magnitude from a uniform distribution].
generating at least one adversarial data file based on at least one data file of the plural data files, the attack magnitude, and the mutual information…
As cited above, Chen teaches the adversarial inlier is created [generating at least one adversarial data file based on] using the input [at least one data file of the plural data files], the
ϵ
-ball [the attack magnitude], and the negative log likelihood [the mutual information].
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen, further with Chen to perturb data for training in order to increase the robustness of the model. “For in-distribution inputs
x
∈
P
X
, ALOE creates adversarial inlier within the-ball that maximize the negative log likelihood. Training with perturbed examples from the in-distribution helps calibrate the error on inliers, and make the model more invariant to the additive noise” (Chen, p. 4, col. 1). Chen discloses that training the model with perturbed examples increases the robustness of the model by making it more invariant to additive noise.
Regarding Claim 20:
Claim 20 corresponds to claim 10 and is rejected for at least the same reasons as given in the rejection of claim 10, with the exception of the following limitations.
Wu discloses:
A non-transitory computer readable medium encoded with program code for automated data processing with a machine learning model configured to detect out-of- distribution samples, when placed in communicable contact with a computer processor, the program code causing the processor to be configured to perform an operation comprising:
Wu, [0002], “The present disclosure describes systems and methods for machine learning architecture for out-of-distribution data set detection.”
[0017], “In another aspect, a non-transitory computer-readable medium or media having stored thereon machine interpretable instructions which, when executed by a processor may cause the processor to perform one or more methods described herein.”
Claim 13 is rejected under 35 U.S.C. 103 as being unpatentable over Wu and Goldenberg in view of Chen as applied to claim 10 above, and further in view of Ma et al. (US 20230306723), hereinafter Ma.
Regarding Claim 13:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, and Goldenberg further discloses:
wherein at least two machine learning algorithms include convolutional neural networks (CNNs)…
Goldenberg, [86], “Each server may also include one or more machine learning models 1170 of an ensemble of machine learning models, such as a CNN.”
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen, further with Goldenberg to use CNNs in their ensembles of machine learning models in order to increase the accuracy predictions. “Instead, during the training process of a machine learning model, the machine learning model adaptively makes its own determinations as to how to modify each computation, convolution or transformation of a given processing node to produce better and/or superior results from the input values it receives” (Goldenberg, [32]).
Wu, Goldenberg, and Chen do not explicitly disclose:
…initialized with random seed weights
However, in the same field, analogous art Ma teaches:
…initialized with random seed weights
Ma, [0051], “Preliminary analysis shows that in random initialization settings (e.g., initializing from scratch), transformers lag behind CNNs on medical target tasks with limited annotated data.”
Ma teaches randomly initializing CNNs [initialized with random seed weights].
Wu, Goldenberg, Chen, Ma, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Ma to use random initialization in order to achieve better performance. “Firstly, in random initialization (scratch) settings (horizontal lines), transformers (e.g., ViT-B and/or Swin-B) cannot compete with CNNs (e.g., such as ResNet-50) in medical applications, as they offer performance equal to or even worse than CNNs. This inferior performance is attributable to the respective transformer’s lack of desirable inductive bias in comparison to CNNs, which has a negative impact on transformer performance on medical target tasks with limited annotated data” (Ma, [0053]).
Claim 14 is rejected under 35 U.S.C. 103 as being unpatentable over Wu and Goldenberg in view of Chen as applied to claim 10 above, and further in view of Georgescu et al. (US 20230099938), hereinafter Georgescu.
Regarding Claim 14:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, but do not explicitly disclose:
wherein processing the at least one training dataset with at least two machine learning algorithms for generating at least two machine learning models includes dropping out a node in the at least two machine learning algorithms
However, in the same field, analogous art Georgescu teaches:
wherein processing the at least one training dataset with at least two machine learning algorithms for generating at least two machine learning models includes dropping out a node in the at least two machine learning algorithms
Georgescu, [0082], “In particular, convolutional neural networks 900 can be trained based on the backpropagation algorithm. For preventing overfitting, methods of regularization can be used, e.g. dropout of nodes 912-920, stochastic pooling, use of artificial data, weight decay based on the L1 or the L2 norm, or max norm constraints.”
Georgescu teaches that during training of the neural networks [processing the at least one training dataset with at least two machine learning algorithms for generating at least two machine learning models includes], dropout of nodes is used to prevent overfitting [includes dropping out a node in the at least two machine learning algorithms].
Wu, Goldenberg, Chen, Georgescu, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Georgescu to use dropout in order to increase the robustness of the model. “In particular, convolutional neural networks 900 can be trained based on the backpropagation algorithm. For preventing overfitting, methods of regularization can be used, e.g. dropout of nodes 912-920, stochastic pooling, use of artificial data, weight decay based on the L1 or the L2 norm, or max norm constraints” (Georgescu, [0082]). Georgescu discloses that dropout prevents overfitting of the neural network, allowing for a more robustness neural network that can handle more generalized data.
Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Wu and Goldenberg in view of Chen as applied to claim 10 above, and further in view of Taniguchi et al. (US 20220131884), hereinafter Taniguchi.
Regarding Claim 15:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, but do not explicitly disclose:
wherein the trained ensemble machine learning model includes a detection rate of false positives no greater than one percent
However, in the same field, analogous art Taniguchi teaches:
wherein the trained…machine learning model includes a detection rate of false positives no greater than one percent
Taniguchi, [0226], “In addition, as illustrated in the detection result management table 541, the false positive rate for the legitimate domain is equal to or less than the false positive threshold value even when the feature ‘freshness’ is utilized. The false positive threshold value is, for example, 1%.”
Wu, Goldenberg, Chen, Taniguchi, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Taniguchi to use a false positive threshold value rate in order to increase the robustness of the model. “The false positive threshold value is, for example, 1%. Therefore, even if the feature ‘freshness’ is utilized to detect the behavior of the wide-area malicious domain, it is considered possible to avoid an incident in which the behavior of the legitimate domain is erroneously determined as the behavior of the malicious domain.” (Taniguchi, [0226]). Taniguchi discloses that the false positive threshold value makes it possible to avoid scenarios where a non-malicious prediction is predicted to be malicious.
Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Wu and Goldenberg in view of Chen as applied to claim 10 above, and further in view of Al-Turki et al. (US 10839269), hereinafter Al-Turki.
Regarding Claim 16:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, but do not explicitly disclose:
wherein mutual information between at least two machine learning models of the ensemble machine learning model is based on a Jensen-Shannon divergence
However, in the same field, analogous art Al-Turki teaches:
wherein mutual information between at least two machine learning models of the…machine learning model is based on a Jensen-Shannon divergence
Al-Turki, [33], “JSD is the Jensen-Shannon divergence describing the distance between two distributions. It is in the range of 0 and 1, and equals 0 when two distributions are same.”
Wu, Goldenberg, Chen, Al-Turki, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Al-Turki to use Jensen-Shannon divergence in order to increase the robustness of the model by achieving adversarial domain adapation. “JSD is the Jensen-Shannon divergence describing the distance between two distributions. It is in the range of 0 and 1, and equals 0 when two distributions are same. Namely, when the distributions of source and target data are identical, l*=−2 log 2 is the global optimum of l(X.sub.S, X.sub.T, M.sub.T). In this case, the goal of adversarial domain adaptation is well achieved.” (Al-Turki, [33]).
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over Wu and Goldenberg in view of Chen as applied to claim 10 above, and further in view of Kariyappa et al. (“Improving Adversarial Robustness of Ensembles with Diversity Training”), hereinafter Kariyappa.
Regarding Claim 18:
As discussed above, Wu, Goldenberg, and Chen teach [the] computer-implemented method of claim 10, but do not explicitly disclose:
wherein generating at least one adversarial data file is defined by:
x
’
=
x
-
ϵ
s
i
g
n
∇
x
U
x
where
x
’
is the adversarial data file,
x
is a data file of the plural data files,
ϵ
is the attack magnitude, and
U
x
is mutual information between at least two machine learning models of the ensemble machine learning model
However, in the same field, analogous art Kariyappa teaches:
wherein generating at least one adversarial data file is defined by:
x
’
=
x
-
ϵ
s
i
g
n
∇
x
U
x
where
x
’
is the adversarial data file,
x
is a data file of the plural data files,
ϵ
is the attack magnitude, and
U
x
is mutual information between at least two machine learning models of the ensemble machine learning model
Kariyappa, p. 3, col. 1, “Let
J
θ
,
x
,
y
denote the loss function of the model
f
, where
θ
represents the model parameters,
x
is the benign input and
y
is the label…Fast Gradient Sign Method (FGSM): FGSM (Goodfellow et al., 2015) uses a linear approximation of the loss function to find an adversarial perturbation that causes the loss function to increase. Let
∇
x
J
θ
,
x
,
y
denote the gradient of the loss function with respect to
x
. The input is modified by adding a perturbation of size
ϵ
in the direction of the gradient vector.
x
’
=
x
+
ϵ
∙
s
i
g
n
∇
x
J
θ
,
x
,
y
Several variants of FGSM have been proposed in recent literature with the goal of improving the effectiveness of the perturbation in increasing the loss function.”
Kariyappa defines an adversarial perturbation defined using a benign input file [a data file of the plural data files], perturbation size
ϵ
[attack magnitude], and model parameters [mutual information].
Wu, Goldenberg, Chen, Kariyappa, and the instant application are analogous art because they are all directed to neural networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify Wu, Goldenberg, and Chen with Kariyappa to perturb data for training in order to find effective adversarial perturbations. “Fast Gradient Sign Method (FGSM): FGSM (Goodfellow et al., 2015) uses a linear approximation of the loss function to find an adversarial perturbation that causes the loss function to increase…Several variants of FGSM have been proposed in recent literature with the goal of improving the effectiveness of the perturbation in increasing the loss function” (Kariyappa, p. 3, col. 1).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to STEVEN PHUNG whose telephone number is (703) 756-1499. The examiner can normally be reached Monday-Thursday: 9:00AM-4:00PM ET.
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, KAMRAN AFSHAR can be reached at (571) 272-7796. 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.
/STEVEN PHUNG/Examiner, Art Unit 2125
/KAMRAN AFSHAR/Supervisory Patent Examiner, Art Unit 2125
1 In view of the Specification filed October 12, 2023, para. [0016], “mutual information (e.g., information based on in-distribution training)”. Examiner is interpreting mutual information to be data related to in-distribution data.
Prosecution Insights