DETAILED ACTION
This nonfinal office action is responsive to claims filed on March 3, 2023. Claims 1-20 are pending. Claims 1, 17, and 19 are independent.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on March 3, 2023 is being considered by the examiner.
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.
Claim 1 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu et al. (CN112541580), hereinafter Zhu, in view of Deng et al., hereinafter Deng.
Regarding claim 1, Zhu teaches the method:
obtaining a first set of labelled training data from a source domain; (Zhu, paragraph n0010: “Use a small number of labeled target domain samples and source samples to train a multiclass discriminator to help the target domain and source domain transfer better.”)
obtaining a second set of labelled training data from a target domain, (Zhu, paragraph n0008: “Select a small number of valuable labels from the target domain samples for labeling through an active learning strategy.”) a number of labelled samples of the first set being greater than a number of labelled samples of the second set; (Zhu, paragraph n0002: “Generally, the domain with a large number of samples and labeled information is called the source domain, and the domain with the fewest samples and lack of labels is called the target domain.” – Since the source domain has a large number of samples and the target domain has the fewest samples this indicates that the number of labelled samples of the first set is greater than the number of labelled samples from the second set.)
training the first machine learning model with the first set and the second set and with a discriminator so that the discriminator is unable to distinguish whether a sample is from the first set or from the second set; and (Zhu, paragraph n0010: “Use a small number of labeled target domain samples and source samples to train a multiclass discriminator to help the target domain and source domain transfer better.” And paragraph n0045: “By combining the source and target domains of multiple classes, a multi-class discriminator is trained to guide the reduction of the distribution distance between the source and target domains.”)
training the first machine learning model … using the first set and the second set. (Zhu discusses training the model with a loss function, see e.g. paragraphs n0016 and n0028.)
Zhu does not explicitly teach:
that the training is done with triplet regularization loss.
However, Deng teaches:
that the training is done with triplet regularization loss. (Deng, page 33, column 2, paragraph 4: “We adopt a two-stage training procedure: we first train the classifier and the feature extractor by minimizing Eq. 2 and Eq. 3, and then train the entire system (SCA) by minimizing Eq. 4.” – Equation 4 on page 32 includes a triplet loss.)
Deng is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu, which already teaches training a machine learning model with a discriminator on training data from a source domain and a target domain but does not explicitly teach that the training is done with a triplet loss regularization, to include the teachings of Deng which does teach that the training is done with a triplet loss regularization so that “images from different domains but of the same class are mapped nearby, and those of different classes are far apart.” (Deng, abstract)
Claims 2, 6-8, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Bekkouch et al. (Triplet Loss Network for Unsupervised Domain Adaptation), hereinafter Bekkouch.
Regarding claim 2, Zhou and Deng teach the method of claim 1, as cited above.
Zhou and Deng do not explicitly teach:
training a classifier of the first machine learning model using classification loss from the first set and the second set.
However, Bekkouch teaches:
training a classifier of the first machine learning model using classification loss from the first set and the second set. (Bekkouch, section 3.3, Classification Loss: “This is the usual cross-entropy loss
H
,
.
,
for the output of source images and their labels and the output of target images (chosen for pseudo-labeling) and their corresponding pseudo-labels, and is computed as follows:
L
C
W
E
,
W
C
=
λ
S
∑
x
S
∈
X
S
H
y
^
S
,
y
S
+
λ
t
∑
x
t
∈
X
t
H
y
^
t
,
y
t
,
where
λ
S
and
λ
t
are used to balance the weighted sum between the source and the target, since we only take a few samples using pseudo-labeling.”)
Bekkouch is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a machine learning model using a first set and a second set but does not explicitly teach that the training includes training a classifier using classification loss from the first set and the second set, to include the teachings of Bekkouch which does teach that the training includes training a classifier using classification loss from the first set and the second set in order to generate “a latent representation that is both domain-invariant and class informative, by pushing samples from the same classes and different domains to share similar distributions.” (Bekkouch, page 3, bullet 2)
Regarding claim 6, Zhu and Deng teach the method of claim 1, as cited above.
Zhu does not explicitly teach:
training a classifier of the first machine learning model using classification loss from the first set and the second set, and
updating a first generator encoder of the first machine learning model based on: domain loss from the training of the discriminator, distance loss from the training with the triplet loss regularization, and classification loss from training a classifier of the first machine learning model using the first set and the second set
wherein the trained first machine learning model for use in an inference phase comprises the updated first generator encoder and the trained classifier.
However, Deng further teaches:
updating a first generator encoder of the first machine learning model based on: domain loss from the training of the discriminator, distance loss from the training with the triplet loss regularization, and classification loss from training a classifier of the first machine learning model using the first set and the second set. (Deng, page 32, column 1, paragraph 5: “Finally, the final objective of the collaborative distribution alignment is written as,
L
s
c
a
=
L
c
+
α
L
d
+
β
L
s
where
L
c
is the classification loss,
L
d
is the the domain confusion loss, and
L
s
is the triplet loss. The
α
and the
β
control the relative importance of domain-level alignment and similarity guided constraint, respectively.”)
Zhu and Deng do not explicitly teach:
training a classifier of the first machine learning model using classification loss from the first set and the second set, and
wherein the trained first machine learning model for use in an inference phase comprises the updated first generator encoder and the trained classifier.
However, Bekkouch teaches:
training a classifier of the first machine learning model using classification loss from the first set and the second set, and (Bekkouch, section 3.3, Classification Loss: “This is the usual cross-entropy loss
H
,
.
,
for the output of source images and their labels and the output of target images (chosen for pseudo-labeling) and their corresponding pseudo-labels, and is computed as follows:
L
C
W
E
,
W
C
=
λ
S
∑
x
S
∈
X
S
H
y
^
S
,
y
S
+
λ
t
∑
x
t
∈
X
t
H
y
^
t
,
y
t
,
where
λ
S
and
λ
t
are used to balance the weighted sum between the source and the target, since we only take a few samples using pseudo-labeling. By minimizing this loss, we update the weights of
both the encoder and the classifier,
W
E
and
W
C
.”)
wherein the trained first machine learning model for use in an inference phase comprises the updated first generator encoder and the trained classifier. (Bekkouch, section 5: “Therefore, in this work, we present our model, TripNet, which has a simple model that allows for fast convergence, yet which can achieve good performance on domain adaptation in different image classification problems. TripNet consists of an encoder, a classifier, and a discriminator.” – Using the model in different image classification problems is analogous to using the model in an inference phase.)
Bekkouch is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a machine learning model and updating the first generator encoder based on the losses but does not explicitly teach training a classifier of the first machine learning model using classification loss from the first set and the second set or that the trained model is used in an inference phase, to include the teachings of Bekkouch which does teach training a classifier of the first machine learning model using classification loss from the first set and the second set or that the trained model is used in an inference phase in order to generate “a latent representation that is both domain-invariant and class informative, by pushing samples from the same classes and different domains to share similar distributions.” (Bekkouch, page 3, bullet 2)
Regarding claim 7, Zhu, Deng, and Bekkouch teach the method of claim 6, as cited above.
Zhu and Deng do not explicitly teach:
performing classification with the trained first machine learning model via: inputting a new sample into the updated first generator encoder so that the updated first generator encoder generates an embedding in an embedding space and
inputting the embedding into the trained classifier so that the trained classifier produces a class prediction.
However, Bekkouch further teaches:
performing classification with the trained first machine learning model via: inputting a new sample into the updated first generator encoder so that the updated first generator encoder generates an embedding in an embedding space and (Bekkouch, Fig. 2: “The encoder translates the images (i.e., the X space) to embeddings in the latent space (i.e., the Z space).” – The images to embeddings in the latent space is analogous to inputting new samples into the updated generator encoder to generate an embedding.)
inputting the embedding into the trained classifier so that the trained classifier produces a class prediction. (Bekkouch, Fig. 2: “The latent representation is fed to both the discriminator and the classifier. The discriminator distinguishes if the latent representation is from source or target domain, whereas the classifier finds the suitable label for it.” – The classifier finding the suitable label is analogous to producing a class prediction.)
Regarding claim 8, Zhu and Deng teach the method of claim 1, as cited above.
Zhu does not explicitly teach:
a first generator encoder of the first machine learning model is updated based on:
domain loss from the training with the discriminator,
distance loss from the training with the triplet loss regularization, and
classification loss from training a classifier of the first machine learning model using the first set and the second set.
However, Deng further teaches:
a first generator encoder of the first machine learning model is updated based on:
domain loss from the training with the discriminator,
distance loss from the training with the triplet loss regularization, and
classification loss from training a classifier of the first machine learning model using the first set and the second set. (Deng, page 32, column 1, paragraph 5: “Finally, the final objective of the collaborative distribution alignment is written as,
L
s
c
a
=
L
c
+
α
L
d
+
β
L
s
where
L
c
is the classification loss,
L
d
is the the domain confusion loss, and
L
s
is the triplet loss. The
α
and the
β
control the relative importance of domain-level alignment and similarity guided constraint, respectively.” – The domain confusion loss is analogous to the domain loss while the triplet loss includes the distance loss, see eq. 1 on page 31.)
Regarding claim 16, Zhu and Deng teach the method of claim 1, as cited above.
Zhu and Deng do not explicitly teach:
the source domain is of a first type and the target domain is of a second type different than the first type.
However, Bekkouch teaches:
the source domain is of a first type and the target domain is of a second type different than the first type. (Bekkouch, abstract: “Domain adaptation is a sub-field of transfer learning that aims at bridging the dissimilarity gap between different domains by transferring and re-using the knowledge obtained in the source domain to the target domain.” – Bridging the dissimilarity gap between different domains is indicative that the source domain is a different type from the target domain.)
Bekkouch is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a machine learning model using data from a source domain and target domain but does not explicitly teach that the source domain and target domain are different types, to include the teachings of Bekkouch which does teach that the source domain and target domain are different types in order to “store the knowledge learned in the primary domain and later transfer that knowledge to a target domain sharing the same tasks but potentially following a different distribution. This can help in reducing the cost of data re-collection and its labelling.” (Bekkouch, Introduction, paragraph 2)
Claims 3-5 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Yin et al. (Metric-learning-assisted domain adaptation), hereinafter Yin, in view of Luo et al. (Deep learning based forecasting of photovoltaic power generation by incorporating domain knowledge), hereinafter Luo.
Regarding claim 3, Zhu and Deng teach the method of claim 1, as cited above.
Zhu and Deng do not explicitly teach:
the training of the first machine learning model with the discriminator and with the triplet loss regularization occurs iteratively based on refining of a sample pool of the first set, wherein the sample pool is refined by evaluating relevancy of the first set in a latent common embedding space between the first set and the second set.
However, Yin teaches:
the training of the first machine learning model with the discriminator and with the triplet loss regularization occurs iteratively based on refining of a sample pool of the first set, (Yin, Fig. 3: “The MLA-DA networks consists of four steps in every iteration. Step1: Fine-tune Feature Extractor F and update Classifier C for samples from source distribution
x
s
,
y
s
. Step2 (Domain Alignment): Align features extracted from source and target data distribution
f
s
and
f
t
by adversarial Discriminator D. Step3: Update margin a with the probability of the second possible label and minimize entropy (E) loss for target samples. Step4 (Metric Learning): Minimize triplet loss with updated a for source metric feature
m
^
s
.” – Four steps in every iteration is indicative that the training occurs iteratively. Fine tuning the feature extractor and Aligning the features extracted is indicative of refining a sample pool based on the first set, i.e. the source data.)
Yin is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training the first machine learning model with the discriminator and triplet loss regularization but does not explicitly teach that it occurs iteratively based on refining a sample pool of the first set, to include the teachings of Yin which does teach that the training occurs iteratively based on refining a sample pool of the first set “for
helping better feature alignment.” (Yin, abstract)
Zhu, Deng, and Yin do not explicitly teach:
wherein the sample pool is refined by evaluating relevancy of the first set in a latent common embedding space between the first set and the second set.
However, Luo teaches:
wherein the sample pool is refined by evaluating relevancy of the first set in a latent common embedding space between the first set and the second set. (Luo, page 3, column 1, paragraph 3: “In the first stage, the filter method examines features based on intrinsic characteristics prior to the learning task. It selects a subset of features from the dataset according to different types of filter criteria, e.g., dependency, information, distance, and consistency [34]. Considering the characteristics of PV data, Pearson correlation coefficient (PCC) [35] is adopted to evaluate the relevance between input variables and the target variable.” – The filter method is analogous to refining the sample pool, where selecting a subset of features according to the filter criteria is analogous to refining by evaluating relevancy.)
Luo is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu, Deng, and Yin, which already teaches refining the sample pool but does not explicitly teach that the refining is done by evaluating relevancy of the first set in the common embedding space, to include the teachings of Luo which does teach refining the feature subset by filtering according to different criteria to evaluate the relevance in order to “identify which ones are important to the model and which ones are irrelevant, thus selecting an appropriate feature subset as the model input.” (Luo, page 3, column 1, paragraph 2)
Regarding claim 4, Zhu, Deng, Yin, and Luo teach the method of claim 3, as cited above.
Zhu and Deng do not explicitly teach:
the refining comprises comparing a first distance between a matching pair to a second distance between a non-matching pair to determine a triplet function value, wherein the matching pair and the non-matching pair each belongs to a triplet and comprises an anchor sample from the second set and a respective additional sample from the first set.
However, Yin further teaches:
the refining comprises comparing a first distance between a matching pair to a second distance between a non-matching pair to determine a triplet function value, wherein the matching pair and the non-matching pair each belongs to a triplet and comprises an anchor sample from the second set and a respective additional sample from the first set. (Yin, page 271, first paragraph: “
PNG
media_image1.png
150
611
media_image1.png
Greyscale
Where
M
:
X
→
R
m
is the metric function learning by embedding function
F
:
X
→
R
n
and metric generator
G
:
R
n
→
R
m
,
b
is the batch size,
D
s
b
a
t
c
h
is a batch of samples drawn from source distribution.” – Where the first term is the distance between positive pairs, the second term is the distance between negative pairs and the third term is the class margin. Thus, there is a comparison between a first distance and a second distance.)
Regarding claim 5, Zhu, Deng, Yin, and Luo teach the method of claim 4, as cited above.
Zhu, Deng, and Yin do not explicitly teach:
the refining further comprises discarding a first triplet for the iterative training in response to the triplet function value for the first triplet not exceeding a threshold value.
However, Luo further teaches:
the refining further comprises discarding a first triplet for the iterative training in response to the triplet function value for the first triplet not exceeding a threshold value. (Luo, page 3, column 2, paragraph 2: “In the first stage, correlations between variables are assessed individually via filter criteria, and feature variables can be filtered accordingly by establishing proper thresholds.” – Luo teaches the refining above, the assessment being individual via filter criterion and being filtered accordingly by the proper threshold is analogous to discarding a first triplet in response to the triplet value not exceeding a threshold value.)
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Bekkouch in view of Zonoozi et al. (A Survey on Adversarial Domain Adaptation), hereinafter Zonoozi.
Regarding claim 9, Zhu, Deng, and Bekkouch teach the method of claim 8, as cited above.
Zhu does not explicitly teach:
updating a second generator encoder based on:
domain loss from the training with the discriminator, and
classification loss from training the classifier of the first machine learning model using the first set and the second set;
wherein the first generator encoder and the second generator encoder are used for the training with the discriminator and for the training of the classifier.
Deng further teaches:
updating a … generator encoder based on: domain loss from the training with the discriminator, and classification loss from training the classifier of the first machine learning model using the first set and the second set; (Deng, page 32, column 1, paragraph 5: “Finally, the final objective of the collaborative distribution alignment is written as,
L
s
c
a
=
L
c
+
α
L
d
+
β
L
s
where
L
c
is the classification loss,
L
d
is the the domain confusion loss, and
L
s
is the triplet loss. The
α
and the
β
control the relative importance of domain-level alignment and similarity guided constraint, respectively.”)
Zhu, Deng, and Bekkouch do not explicitly teach:
That the updating a generator encoder is updating a second generator encoder.
wherein the first generator encoder and the second generator encoder are used for the training with the discriminator and for the training of the classifier.
However, Zonoozi teaches:
That the updating a generator encoder is updating a second generator encoder. (Zonoozi, Fig. 12 - shows there is a generator encoder for both the source and target and, therefore, teaches a second generator encoder while fig. 8 shows the training that is already taught by Deng above.)
wherein the first generator encoder and the second generator encoder are used for the training with the discriminator and for the training of the classifier. (Zonoozi, Fig. 8 - shows the feature extractor, e.g. the generator encoder, of both the source and target is used to train the classifier and the discriminator.)
Claims 10 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Bekkouch in view of Zonoozi in view of Gong et al. (Domain Adaptation with Conditional Transferable Components), hereinafter Gong, in view of Dash et al. (A review of some techniques for inclusion of domain-knowledge into deep neural networks), hereinafter Dash.
Regarding claim 10, Zhu, Deng, Bekkouch, and Zonoozi teach the method of claim 9, as cited above.
Zhu, Deng, Bekkouch, and Zonoozi do not explicitly teach:
comparing at least one of a shape and a dimension of the labelled training data of the second set to a corresponding at least one of a shape and a dimension of the labelled training data of the first set; and
initializing weights of the first generator encoder based on the comparison.
However, Gong teaches:
comparing at least one of a shape and a dimension of the labelled training data of the second set to a corresponding at least one of a shape and a dimension of the labelled training data of the first set; and (Gong, page 1, column 2, paragraph 2: “For instance, many existing domain adaptation methods consider the covariate shift situation where the distributions on two domains only differ in the marginal distribution of the features P(X), while the conditional distribution of the target given the features P(Y |X) does not change. In this case, one can match the feature distribution P(X) on source and target domains by importance reweighting methods if the source domain is richer than the target domain (Shimodaira, 2000; Sugiyama et al., 2008; Huang et al., 2007). The weights are defined as the density ratio between the source and target domain features and can be efficiently estimated by various methods such as the kernel mean matching procedure (KMM) (Huang et al., 2007).” – Matching feature distribution on source and target domains by importance reweighting methods is analogous to comparing the shape and dimension of the training data of the first and second set.)
Gong is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu, Deng, Bekkouch, and Zonoozi, which already teaches updating the first and second generator encoders but does not explicitly teach comparing the shape and dimension of the training data, to include the teachings of Gong which does teach comparing the shape and dimension of the training data in order to "make use of these components for domain adaptation." (Gong, page 2, column 1, paragraph 1)
Zhu, Deng, Bekkouch, Zonoozi, and Gong do not explicitly teach:
initializing weights of the first generator encoder based on the comparison.
However, Dash teaches:
initializing weights of the first generator encoder based on the comparison. (Dash, page 9, paragraph 1: “Similar to transfer learning, the knowledge is transferred from a source domain to a target domain in the form of a prior distribution over the model parameters.” – transferring knowledge from a source to a target domain is analogous to initializing the weights when the distributions are close enough.)
Dash is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu, Deng, Bekkouch, Zonoozi, and Gong, which already teaches comparing the shape and dimension of the training data sets but does not explicitly teach initializing the weights based off the comparison, to include the teachings of Dash which does teach initializing the weights based off the comparison in order to "boost performance significantly." (Dash, page 8, last paragraph)
Regarding claim 11, Zhu, Deng, Bekkouch, Zonoozi, Gong, and Dash teach the method of claim 10, as cited above.
Zhu, Deng, Bekkouch, and Zonoozi do not explicitly teach:
wherein via the comparison the at least one of the shape and the dimension of the labelled training data of the second set passes a similarity threshold with the corresponding at least one of a shape and a dimension of the labelled training data of the first set; and
wherein the initialization comprises weights of the second generator encoder being implemented as the weights for the first generator encoder.
However, Gong teaches:
wherein via the comparison the at least one of the shape and the dimension of the labelled training data of the second set passes a similarity threshold with the corresponding at least one of a shape and a dimension of the labelled training data of the first set; and (Gong, page 1, column 2, paragraph 2: “For instance, many existing domain adaptation methods consider the covariate shift situation where the distributions on two domains only differ in the marginal distribution of the features P(X), while the conditional distribution of the target given the features P(Y |X) does not change. In this case, one can match the feature distribution P(X) on source and target domains by importance reweighting methods if the source domain is richer than the target domain (Shimodaira, 2000; Sugiyama et al., 2008; Huang et al., 2007). The weights are defined as the density ratio between the source and target domain features and can be efficiently estimated by various methods such as the kernel mean matching procedure (KMM) (Huang et al., 2007).” – Matching feature distribution on source and target domains by importance reweighting methods is analogous to comparing the shape and dimension of the training data of the first and second set.)
Zhu, Deng, Bekkouch, Zonoozi, and Gong do not explicitly teach:
wherein the initialization comprises weights of the second generator encoder being implemented as the weights for the first generator encoder.
However, Dash teaches:
wherein the initialization comprises weights of the second generator encoder being implemented as the weights for the first generator encoder. (Dash, page 9, paragraph 1: “Similar to transfer learning, the knowledge is transferred from a source domain to a target domain in the form of a prior distribution over the model parameters. This form of domainadaptation uses the same model structure as the source, along with an initial set of parameter values obtained from the source model” .” – transferring knowledge from a source to a target domain is analogous to initializing the weights when the distributions are close enough.)
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Bekkouch in view of Zonoozi in view of Gong in view of Dash in view of Muhamedrahimov et al. (US20220318567), hereinafter Muhamedrahimov.
Regarding claim 12, Zhu, Deng, Bekkouch, Zonoozi, Gong, and Dash teach the method of claim 10, as cited above.
Zhu, Deng, Bekkouch, and Zonoozi do not explicitly teach:
wherein via the comparison the at least one of the shape and the dimension of the labelled training data of the second set fails a similarity threshold with the corresponding at least one of a shape and a dimension of the labelled training data of the first set; and
wherein the initialization comprises randomly initializing the weights of the first generator encoder.
However, Gong teaches:
wherein via the comparison the at least one of the shape and the dimension of the labelled training data of the second set fails a similarity threshold with the corresponding at least one of a shape and a dimension of the labelled training data of the first set; and (Gong, page 1, column 2, paragraph 2: “For instance, many existing domain adaptation methods consider the covariate shift situation where the distributions on two domains only differ in the marginal distribution of the features P(X), while the conditional distribution of the target given the features P(Y |X) does not change. In this case, one can match the feature distribution P(X) on source and target domains by importance reweighting methods if the source domain is richer than the target domain (Shimodaira, 2000; Sugiyama et al., 2008; Huang et al., 2007). The weights are defined as the density ratio between the source and target domain features and can be efficiently estimated by various methods such as the kernel mean matching procedure (KMM) (Huang et al., 2007).” – Matching feature distribution on source and target domains by importance reweighting methods is analogous to comparing the shape and dimension of the training data of the first and second set.)
Zhu, Deng, Bekkouch, Zonoozi, Gong, and Dash do not explicitly teach:
wherein the initialization comprises randomly initializing the weights of the first generator encoder.
However, Muhamedrahimov teaches:
wherein the initialization comprises randomly initializing the weights of the first generator encoder. (Muhamedrahimov, paragraph 0163: “First, an attempt was made to directly predict the contrast phase in chest scans using an ML model trained only on abdomen CT samples, to explore whether similarities in feature distributions would allow the phase to be directly inferred in the new domain. Performance was compared between chest contrast phase classifiers trained with random weight initialization and with weights initialized from the trained abdomen ML model, incrementing the number of training samples available.”)
Muhamedrahimov is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu, Deng, Bekkouch, Zonoozi, Gong, and Dash, which already teaches comparing the shape and dimension of the training datasets and initializing the weights of the generator encoder based on the comparison but does not explicitly teach that the weights are randomly initialized, to include the teachings of Muhamedrahimov which teaches an approach in the state of the art which contemplates random weight initialization compared to weight initialization from a model which has a different domain. Thus, we have a method of initializing weights both randomly and by using the weights of the other model and comparing those results, it would then be obvious to compare the data from both domains first to determine if using the weights of the existing model or randomly initializing the weights would be more appropriate based off that comparison as the results show that transferring weights from sufficiently related tasks/body parts shows better results but that this may not extend to different tasks/body parts. (Muhamedrahimov, paragraph 0185)
Claims 13-15 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Shuang Li et al. (Domain Invariant and Class Discriminative Feature Learning for Visual Domain Adaptation), hereinafter S. Li.
Regarding claim 13, Zhu and Deng teach the method of claim 1, as cited above.
Zhu and Deng do not explicitly teach:
the triplet loss regularization comprises penalization in response to samples from the second set being mapped at a distance greater than a distance threshold from samples from the first set and from the second set having the same class labels as the samples from the second set.
However, S. Li teaches:
the triplet loss regularization comprises penalization in response to samples from the second set being mapped at a distance greater than a distance threshold from samples from the first set and from the second set having the same class labels as the samples from the second set. (S. Li, page 4263, column 2, last paragraph: “To develop an effective loss term to improve the discriminative power of the learned features, we first expect to reduce the intra-class variation by minimizing the distance between the same class instances in the source and target domains, respectively.” – The loss term being the triplet loss is taught by Deng above. Reducing the intra-class variation by minimizing the distance is analogous to penalizing the triplet loss based on a distance being too large when they are from the same class.)
S. Li is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a model using triplet loss regularization but does not explicitly teach a penalization in response to the distances being further apart, to include the teachings of S. Li which does teach a penalization in response to the distances being further apart as “the class discriminative properties will dramatically alleviate the cross-domain distribution inconsistency, which further boosts the classification performance.” (S. Li, abstract)
Regarding claim 14, Zhu and Deng teach the method of claim 1, as cited above.
Zhu and Deng do not explicitly teach:
the triplet loss regularization comprises penalization in response to samples from the second set being mapped at a distance less than a distance threshold from samples from the first set and from the second set having different class labels as the samples from the second set.
However, S. Li teaches:
the triplet loss regularization comprises penalization in response to samples from the second set being mapped at a distance less than a distance threshold from samples from the first set and from the second set having different class labels as the samples from the second set. (S. Li, page 4264, column 2, first paragraph: “We propose to let the distance between instances with non-matching labels in the latent space as large as possible, which will pull different clusters away to improve the discriminativeness of the learned feature representations.” – letting the distance between instances with non-matching labels be as large as possible is analogous to a penalization in response to samples being mapped at a distanced smaller than a threshold when they are from different classes.)
S. Li is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a model using triplet loss regularization but does not explicitly teach a penalization in response to the distances being closer together, to include the teachings of S. Li which does teach a penalization in response to the distances being closer together as “the class discriminative properties will dramatically alleviate the cross-domain distribution inconsistency, which further boosts the classification performance.” (S. Li, abstract)
Regarding claim 15, Zhu and Deng teach the method of claim 1, as cited above.
Zhu and Deng do not explicitly teach:
the triplet loss regularization comprises:
penalization in response to samples from the second set being mapped at a first distance greater than a first distance threshold from samples from the first set and from the second set having the same class labels as the samples from the second set and
penalization in response to samples from the second set being mapped at a second distance less than a second distance threshold from samples from the first set and from the second set having different class labels as the samples from the second set.
However, S. Li teaches:
the triplet loss regularization comprises:
penalization in response to samples from the second set being mapped at a first distance greater than a first distance threshold from samples from the first set and from the second set having the same class labels as the samples from the second set and (S. Li, page 4263, column 2, last paragraph: “To develop an effective loss term to improve the discriminative power of the learned features, we first expect to reduce the intra-class variation by minimizing the distance between the same class instances in the source and target domains, respectively.” – The loss term being the triplet loss is taught by Deng above. Reducing the intra-class variation by minimizing the distance is analogous to penalizing the triplet loss based on a distance being too large when they are from the same class.)
penalization in response to samples from the second set being mapped at a second distance less than a second distance threshold from samples from the first set and from the second set having different class labels as the samples from the second set. (S. Li, page 4264, column 2, first paragraph: “We propose to let the distance between instances with non-matching labels in the latent space as large as possible, which will pull different clusters away to improve the discriminativeness of the learned feature representations.” – letting the distance between instances with non-matching labels be as large as possible is analogous to a penalization in response to samples being mapped at a distanced smaller than a threshold when they are from different classes.)
S. Li is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a model using triplet loss regularization but does not explicitly teach a penalization in response to the distances of matching class labels being further apart and the distances of non-matching class labels being closer together, to include the teachings of S. Li which does teach a penalization in response to the distances of matching class labels being further apart and the distances of non-matching class labels being closer together as “the class discriminative properties will dramatically alleviate the cross-domain distribution inconsistency, which further boosts the classification performance.” (S. Li, abstract)
Claims 17 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Chidlovskii (US20210019629), hereinafter Chidlovskii.
Regarding claim 17, Claim 17 has all the same limitations of claim 1 which are taught by Zhu and Deng – see claim 1 above.
Zhu and Deng do not explicitly teach:
A computer system for training a first machine learning model, the computer system comprising: one or more processors, one or more computer-readable tangible storage media, and program instructions stored on at least one of the one or more computer-readable tangible storage media for execution by at least one of the one or more processors to cause the computer system to:
However, Chidlovskii teaches:
A computer system for training a first machine learning model, the computer system comprising: (Chidlovskii, paragraph 0065: “In adversarial domain adaptation (ADDA) a domain classifier is combined with domain representation learning to form an adversarial domain adaptation network.” And paragraph 0145: “Although the above embodiments have been described in the context of method steps, they also represent a description of a corresponding component, module or feature of a corresponding apparatus or system.”)
one or more processors, one or more computer-readable tangible storage media, and program instructions stored on at least one of the one or more computer-readable tangible storage media for execution by at least one of the one or more processors to cause the computer system to: (Chidlovskii, paragraphs 0146-0147: “Some or all of the method steps may be implemented by a computer in that they are executed by (or using) a processor, a microprocessor, an electronic circuit or processing circuitry. The embodiments described above may be implemented in hardware or in software.”)
Chidlovskii is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a machine learning model using a triplet loss so that a discriminator is unable to distinguish if the sample is from the first set or the second set but does not explicitly teach a system with one or more processors, one or more storage media, and program instructions stored to carry out the method, to include the teachings of Chidlovskii which does teach a system with one or more processors, one or more storage media, and program instructions stored to carry out the method. Storing instructions to train a machine learning model in memory is a known technique and, therefore, it would have been obvious to store and execute the method, which is taught by Zhu and Deng, on the system of Chidlovskii.
Regarding claim 19, claim 19 has all the same limitations of claim 1 which are taught by Zhu and Deng – see claim 1 above.
Zhu and Deng do not explicitly teach:
A computer program product for training a first machine learning model, the computer program product comprising a computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executable by a processor to cause the processor to:
However, Chidlovskii teaches:
A computer program product for training a first machine learning model, the computer program product comprising a computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executable by a processor to cause the processor to: (Chidlovskii, paragraph 0065: “In adversarial domain adaptation (ADDA) a domain classifier is combined with domain representation learning to form an adversarial domain adaptation network.” And paragraph 0148: “The program code or the computer-executable instructions may, for example, be stored on a computer-readable storage medium.”)
Chidlovskii is considered analogous to the claimed invention as it is in the same field of endeavor, machine learning. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date to have modified Zhu and Deng, which already teaches training a machine learning model using a triplet loss so that a discriminator is unable to distinguish if the sample is from the first set or the second set but does not explicitly teach a computer program product for training the machine learning model comprising a storage medium having the instructions stored thereon, to include the teachings of Chidlovskii which does teach a computer program product for training the machine learning model comprising a storage medium having the instructions stored thereon. Storing instructions to train a machine learning model in memory is a known technique and, therefore, it would have been obvious to store the method, which is taught by Zhu and Deng, as a computer program product of Chidlovskii.
Claims 18 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Deng in view of Chidlovskii in view of Yin in view of Luo.
Regarding claim 18, Zhu, Deng, and Chidlovskii teach the computer system of claim 17, as cited above.
Claim 18 additionally has the same limitations of claim 3 which are taught by Yin and Luo – see claim 3 above.
The teachings of Chidlovskii does not substantially change the method of claim 3, thus the motivation to combine both Yin and Luo with the teachings of Zhu, Deng, and Chidlovskii is the same as the motivation to combine Yin and Luo with the teachings of Zhu and Deng.
Regarding claim 20, Zhu, Deng, and Chidlovskii teach the computer program product of claim 19, as cited above.
Claim 20 additionally has the same limitations of claim 3 which are taught by Yin and Luo – see claim 3 above.
The teachings of Chidlovskii does not substantially change the method of claim 3, thus the motivation to combine both Yin and Luo with the teachings of Zhu, Deng, and Chidlovskii is the same as the motivation to combine Yin and Luo with the teachings of Zhu and Deng.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Lei Li et al. (Unsupervised Domain Adaptation via Discriminative Feature Learning and Classifier Adaptation from Center-Based Distances)
Tsai et al. (US20190354807)
Lagunes-Fortiz et al. (Learning Discriminative Embeddings for Object Recognition on-the fly)
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JACQUELINE MEYER whose telephone number is (703)756-5676. The examiner can normally be reached M-F 8:00 am - 4:30 pm EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tamara Kyle can be reached at 571-272-4241. 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.
/J.C.M./Examiner, Art Unit 2144
/TAMARA T KYLE/Supervisory Patent Examiner, Art Unit 2144