DETAILED ACTION
This Office action is responsive to the Request for Continued Examination (RCE) filed under 37 CFR §1.53(d) for the instant application on December 22, 2025. The Applicants have properly set forth the RCE, which has been entered into the application, and an examination on the merits follows herewith.
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
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, 3, 9, 11 and 15-20 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, there is no antecedent basis for “the learning output data” recited therein. Claim 1 previously recites “output the machine learning model” and “outputs feature data” but provides no prior recitation of “learning output data.” Also, there is no antecedent basis for “the quality data” recited in claim 1. Claim 1 previously recites “predicting a quality of the product” but does no provide any prior recitation of “quality data” per se.
Claims 3, 9, 11 and 15-18 depend from claim 1 and thereby include all of the limitations of claim 1. Accordingly, claims 3, 9, 11 and 15-18 are considered indefinite for the same reasons as noted above with respect to claim 1.
In each of claims 19 and 20, like with claim 1 described above, there is no antecedent basis for “the learning output data” and “the quality data.”
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 16, 19 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2019/0188840 to Kwon et al. (“Kwon”), and also over the article entitled, “Stacked convolutional sparse denoising auto-encoder for identification of defect patterns in semiconductor wafer map” by Yu et al. (“Yu”).
Regarding claims 1, 19 and 20, Kwon describes a semiconductor defect classification device for classifying a semiconductor defect without depending on a human defect evaluation in the process of manufacturing a semiconductor device (see e.g. paragraph 0006). Like claimed, Kwon particularly teaches:
acquiring multi-dimensional physical-property data representing a physical property of a product (see e.g. paragraphs 0007, 0028, 0036 and 0042: Kwon discloses that the semiconductor defect classification device receives various images of semiconductor patterns on a wafer. The images, individually or in combination, are considered multi-dimensional physical property data representing a physical property of a product, i.e. of a semiconductor.);
deriving learning input data to be input to a machine learning model for predicting a quality of the product from the multi-dimensional physical-property data and deriving, as the learning input data, multi-dimensional physical-property relevance data which is related to the multi-dimensional physical-property data by applying at least part of a neural network to the multi-dimensional physical-property data (see e.g. paragraph 0007: Kwon discloses that the semiconductor defect classification device includes feature extractors configured to extract features from the various received images of the semiconductor patterns. The feature extractors extract features from the images by using, for example, a convolutional neural network (CNN) – see e.g. paragraphs 0063-0067. Kwon further discloses that the semiconductor defect classification device also includes a classifier, e.g. a neural network or support vector machine, that is configured to receive the extracted features and meta information, and that uses machine learning to classify a defect of the semiconductor patterns associated with the images based on the extracted features and the meta information – see e.g. paragraphs 0007, 0045, 0068-0071 and 0109-0116. Kwon suggests that the classifier is trained based on such extracted features and meta information – see e.g. paragraphs 0045-0046, 0049-0050 and 0109-0116. The extracted features are considered “multi-dimensional physical-property relevance data” like claimed, and the combination of the extracted features and meta information is considered “learning input data” like claimed. Kwon thus teaches deriving learning input data to be input to a machine learning model for predicting a quality, i.e. a defect, of the product from the multi-dimensional physical-property data, i.e. from the image(s), and derive, as the learning input data, multi-dimensional physical-property relevance data, i.e. the extracted features, which is related to the multi-dimensional physical-property data by applying at least part of a neural network, e.g. a CNN, to the multi-dimensional physical property data.); and
inputting the learning input data to the machine learning model, performing learning, and outputting the machine learning model as a learned model to be provided for actual operation (see e.g. paragraphs 0045-0046, 0049-0050 and 0109-0116: like noted above, Kwon suggests that the classifier is trained based on the extracted features and meta information. The learning input data, i.e. the extracted features and meta information, is thus understandably input to the machine learning model to perform learning, whereby the machine learning model is output as a learned model to be used for actual operation.);
wherein the learning input data includes the multi-dimensional physical property relevance data that have been derived by applying the at least part of the neural network and production condition data which is set in a production process of the product and to which the neural network has not been applied (as noted above, the combination of the extracted features and meta information used to train the classifier taught by Kwon is considered “learning input data” like claimed. Kwon particularly discloses that the meta information can comprise production condition data that is set in a production process of the semiconductor and to which the feature extractors, e.g. a CNN, used to extract the features from the images has not been applied – see e.g. paragraphs 0031-0033 and 0063-0070. The learning input data thus includes the multi-dimensional physical-property relevance data, i.e. the extracted features, that have been derived by applying at least a part of the neural network, e.g. a CNN, and production condition data, i.e. meta information, which is set in a production process of the product and to which the neural network has not been applied.);
wherein the multi-dimensional physical-property data is input to the neural network, the neural network outputs feature data, and the multi-dimensional physical-property relevance data is derived based on the feature data (see e.g. paragraphs 0007 and 0063-0067: like noted above, Kwon discloses that the semiconductor defect classification device includes feature extractors, e.g. a CNN, configured to extract features from the various received images of the semiconductor patterns. As noted above, the extracted features are considered “multi-dimensional physical-property relevance data” like claimed. Accordingly, Kwon teaches inputting the multi-dimensional physical-property data to the neural network i.e. inputting the image(s) to the CNN, whereby the neural network outputs extracted feature data, and the multi-dimensional physical-property relevance data, i.e. the feature data, is derived based on the feature data.); and
wherein the learning input data is input to the machine learning model, the machine learning model outputs learning output data, an accuracy of prediction of the machine learning model is evaluated, and the machine learning model is updated, while changing the learning input data and quality data (see e.g. paragraphs 0007 and 0070-0071: Kwon discloses that the classifier receives the features extracted from the images and the meta information, and classifies a defect of the semiconductor patterns associated with the images based on the extracted features and meta information. The classifier uses machine learning, e.g. a neural network or support vector machine, to classify the defect of the semiconductor – see e.g. paragraphs 0068-0071 and 0109-0116. The learning input data, i.e. the extracted features and meta information, is thus input to the machine learning model, i.e. to the classifier, which outputs learning output data, i.e. a predicted defect classification. Kwon further suggests that the classifier can be updated with additional learning input data when, e.g., a performance of the classifier decreases – see e.g. paragraphs 0046, 0049-0050, 0149-0150 and 0156-0157. Accordingly, Kwon further teaches evaluating the performance of the classifier, including understandably the accuracy of its predictions, and updating the classifier while changing the learning input data and quality data, i.e. updating the machine learning model based on the additional learning input data and associated ground truths.).
Kwon thus teaches a method similar to that of claim 19. Kwon further teaches that the above-described tasks can be implemented via computer program instructions stored on a non-transitory computer readable storage medium, whereby a computer processor can execute the instructions to perform the tasks (see e.g. paragraphs 0170, and 0172-0173). A computer comprising a processor configured to implement the above-noted tasks taught by Kwon is considered a learning apparatus similar to that of claim 1. The non-transitory computer readable storage medium comprising computer program instructions to implement the above-described tasks taught by Kwon is considered a non-transitory computer readable recording medium similar to that of claim 20. Kwon, however, does not explicitly disclose that the neural network used to derive the multi-dimensional physical-property relevance data from the multi-dimensional physical-property data is an autoencoder like required by claims 1, 19 and 20.
Yu nevertheless teaches using an autoencoder (i.e. a stacked convolutional sparse denoising auto-encoder) to derive multi-dimensional physical-property relevance data (i.e. to extract features) from multi-dimensional physical-property data (i.e. from an image) (see e.g. the Abstract, section 1 “Introduction,” and section 3 “Stacked convolutional sparse denoising auto-encoder”). Particularly, like claimed, Yu teaches deriving learning input data to be input to a machine learning model (i.e. to an SVM classifier) for predicting a quality (e.g. defect pattern) of a product (i.e. a wafer map) from multi-dimensional physical property data (i.e. from an image) representing a physical property of the product (see e.g. the Abstract, section 1 “Introduction,” and section 3 “Stacked convolutional sparse denoising auto-encoder”). The learning input data is derived as the multi-dimensional physical-property relevance data (i.e. pooled features), which is related to the multi-dimensional physical-property data, by applying at least a part of an autoencoder to the multi-dimensional physical-property data (see e.g. the Abstract, section 1 “Introduction,” and section 3 “Stacked convolutional sparse denoising auto-encoder”). Specifically, the multi-dimensional physical-property data is input to an autoencoder, which outputs feature data from an encoder network of the autoencoder, and whereby the multi-dimensional physical-property relevance data is derived based on the feature data (see e.g. the Abstract, section 1 “Introduction,” and section 3 “Stacked convolutional sparse denoising auto-encoder”).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon and Yu before the effective filing date of the claimed invention, to modify the apparatus, method and non-transitory computer readable recording medium taught by Kwon so as to utilize an autoencoder like taught by Yu to derive the multi-dimensional physical-property relevance data from the multi-dimensional physical-property data. It would have been advantageous to one of ordinary skill to utilize such a combination because it can obtain much more effective feature representations, as is taught by Yu (see e.g. section 1 “Introduction”). Accordingly, Kwon and Yu are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 1, an operation method like that of claim 19, and a non-transitory computer readable recording medium like that of claim 20.
As per claim 16, Kwon further teaches that the multi-dimensional physical-property data includes image data obtained by imaging the product (see e.g. paragraphs 0007, 0028 and 0036). Yu provides a similar teaching (see e.g. section 3 “Stacked convolutional space denoising auto-encoder”). Accordingly, the above-described combination of Kwon and Yu is further considered to teach a learning apparatus like that of claim 16.
Claims 3, 9 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Kwon and Yu, which is described above, and also over Japanese Patent Publication No. JP 2016091359 A to Kasahara et al. (“Kasahara”), which is cited in Applicant’s information disclosure statement filed on January 14, 2022. Reference is made herein to the previously-provided machine-translation of Kasahara.
Regarding claim 3, Kwon and Yu teach a learning apparatus like that of claim 1, as is described above, and which comprises an autoencoder for deriving multi-dimensional physical-property relevance data from multi-dimensional physical property data (i.e. an image) representing a physical property of a product. Kwon and Yu, however, do not disclose that the autoencoder is learned by inputting the multi-dimensional physical-property data of the product of which the quality is higher than a preset level, and wherein the first processor inputs the multi-dimensional physical-property data to the autoencoder, outputs output data, and derives the multi-dimensional physical property relevance data based on difference data between the multi-dimensional physical property data which is input to the autoencoder and the output data, as is required by claim 3.
Similar to Kwon and Yu, Kasahara teaches deriving learning input data to be input to a machine learning model for predicting a quality of a product from multi-dimensional physical property data representing a physical property of the product, and particularly teaches deriving, as the learning input data, multi-dimensional physical-property relevance data which is related to the multi-dimensional physical property data by applying at least a part of an autoencoder to the multi-dimensional physical-property data (see e.g. paragraphs 0010-0011: Kasahara describes an information processing device that acquires multidimensional data relating to the physical properties of an object and that identifies whether or not the object is, for example, a product that meets a standard. In particular, the processing device reduces the number of dimensions of the multidimensional data using a dimension reduction means trained in advance by semi-supervised learning, and then restores the number of dimensions to the original number of dimensions using a dimension restoration means also trained in advance using semi-supervised learning – see e.g. paragraphs 0028-0029. The dimension reduction means and the dimension restoration means can be implemented via stacked autoencoders – see e.g. paragraphs 0050-0051. The information processing device next calculates an error based on the original multidimensional data and the restored multidimensional data, and then uses a classification means trained in advance by semi-supervised learning to identify, based on the error, whether the object is a product that meets the standard – see e.g. paragraphs 0030-0031. Going into more detail, Kasahara discloses that the dimension reduction means and the dimension restoration means are trained in advance using multidimensional data for a product that meets the predetermined standard – see e.g. paragraphs 0140-0143. Kasahara discloses that the dimension reduction means and the dimension restoration means, when implemented using stacked autoencoders, are trained by inputting the multi-dimensional physical-property data to the autoencoders and adjusting the weights of the autoencoders until the input data and the output data are the same – see e.g. paragraphs 0153-0163. In addition, an error calculation unit calculates a multidimensional error vector based on the difference between the multi-dimensional physical-property data input to the autoencoders and the multidimensional data output from the autoencoders – see e.g. paragraphs 0165-0167. The classification means is then trained using this error data – see e.g. paragraphs 0168-0172. Accordingly, the error data is considered “learning input data” to be input to a machine learning model, i.e. into the classification means, for predicting a quality of the product from the multidimensional physical property data, wherein the learning input data is derived as multidimensional physical property-relevance data, i.e. as the error data, which is related to the multidimensional physical property data by applying at least a part of an autoencoder to the multidimensional physical property data.). Like in claim 3, Kasahara further teaches that the autoencoder is learned by inputting the multi-dimensional physical-property data of the product of which the quality is higher than a preset level, and that a processor inputs the multi-dimensional physical property data to the autoencoder, outputs output data, and derives the multi-dimensional physical property relevance data based on difference data between the multi-dimensional physical property data which is input to the autoencoder and the output data (see e.g. paragraphs 0031-0032, 0139-0140, and 0153-0163: like noted above, Kasahara teaches that multidimensional physical property data for a product that meets a predetermined standard is used to train the dimension reduction means and the dimension restoration means, i.e. the autoencoders. The autoencoder is thus learned by inputting multidimensional physical-property data of a product of which the quality is higher than a preset level, i.e. that meets the predetermined standard. Like further noted above, Kasahara also teaches that an error calculation unit calculates a multidimensional error vector based on the difference between the multi-dimensional physical-property data input to the autoencoders and the multidimensional data output from the autoencoders – see e.g. paragraphs 0165-0167. As noted above, this error data is considered multidimensional physical property-relevance data/learning input data like claimed. Kasahara thus further teaches inputting the multi-dimensional physical property data to the autoencoder, and deriving the multi-dimensional physical property relevance data, i.e. the error data, based on difference data between the multi-dimensional physical property data which is input to the autoencoder and the output data output from the autoencoder.).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu and Kasahara before the effective filing date of the claimed invention, to modify the apparatus taught by Kwon and Yu so as to derive the multi-dimensional physical-property relevance data from the multi-dimensional physical-property data using the autoencoder in a manner like taught by Kasahara, whereby the autoencoder is learned by inputting the multi-dimensional physical-property data of a product of which the quality is higher than a preset level, and wherein a processor inputs the multi-dimensional physical property data to the autoencoder, outputs output data, and derives the multi-dimensional physical property relevance data based on difference data between the multi-dimensional physical property data which is input to the autoencoder and the output data. It would have been advantageous to one of ordinary skill to utilize such a combination because it would enable the evaluation of products without requiring training data from defective products, and would enable the evaluation of products having unexpected defects, as is suggested by Kasahara (see e.g. paragraphs 0004-0006 and 0172). Accordingly, Kwon, Yu and Kasahara are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 3.
Regarding claim 9, Kwon and Yu teach a learning apparatus like that of claim 1, as is described above, which comprises an autoencoder for deriving multi-dimensional physical-property relevance data from multi-dimensional physical property data (i.e. an image) representing a physical property of a product. Kwon and Yu, however, do not disclose that the multi-dimensional physical-property data includes image data of a spectrum which is represented by spectrum data detected by performing spectroscopic analysis on the product, as is required by claim 9.
Like noted above (see the rejection for claim 3), Kasahara similarly teaches deriving multi-dimensional physical-property relevance data by applying at least a part of an autoencoder to multi-dimensional physical-property data representing a physical property of a product. Like in claim 9, Kasahara further teaches that the multi-dimensional physical-property data includes image data of a spectrum which is represented by spectrum data detected by performing spectroscopic analysis on the product (see e.g. paragraphs 0011, 0013-0014, 0019-0020, 0108 and 0138-0140).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu and Kasahara before the effective filing date of the claimed invention, to modify the apparatus taught by Kwon and Yu such that the multi-dimensional physical-property data additionally or alternatively includes image data of a spectrum like taught by Kasahara, which is represented by spectrum data detected by performing spectroscopic analysis on the product. It would have been advantageous to one of ordinary skill to utilize such spectrum data because it can aid in identifying whether or not a product meets a predetermined standard, as is suggested by Kasahara (see e.g. paragraph 0011). Accordingly, Kwon, Yu and Kasahara are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 9.
Regarding claim 11, Kwon, Yu and Kasahara teach a learning apparatus like that of claim 3, as is described above, which comprises an autoencoder for deriving multi-dimensional physical-property relevance data from multi-dimensional physical property data (i.e. an image) representing a physical property of a product. Kwon and Yu do not disclose that the multi-dimensional physical-property data includes image data of a spectrum which is represented by spectrum data detected by performing spectroscopic analysis on the product, as is required by claim 11. Nevertheless, Kasahara teaches multi-dimensional physical-property data that includes image data of a spectrum which is represented by spectrum data detected by performing spectroscopic analysis on a product (see e.g. paragraphs 0011, 0013-0014, 0019-0020, 0108 and 0138-0140). It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu and Kasahara before the effective filing date of the claimed invention, to further modify the apparatus taught by Kwon, Yu and Kasahara such that the multi-dimensional physical-property data additionally or alternatively includes image data of a spectrum like taught by Kasahara, which is represented by spectrum data detected by performing spectroscopic analysis on the product. It would have been advantageous to one of ordinary skill to utilize such spectrum data because it can aid in identifying whether or not a product meets a predetermined standard, as is suggested by Kasahara (see e.g. paragraph 0011). Accordingly, Kwon, Yu and Kasahara are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 11.
Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Kwon, Yu and Kasahara, which is described above, and also over the article entitled, “Novel segmented stacked autoencoder for effective dimensionality reduction and feature extraction in hyperspectral imaging” by Zabalza et al. (“Zabalza”).
Regarding claim 15, Kwon, Yu and Kasahara teach a learning apparatus like that of claim 9, as is described above, which comprises a first processor configured to derive multi-dimensional physical-property relevance data that is related to multi-dimensional physical-property data, wherein the multi-dimensional physical-property data includes image data of a spectrum which is represented by spectrum data detected by performing spectroscopic analysis on a product. Kwon, Yu and Kasahara, however, do not explicitly disclose that the first processor derives the multi-dimensional physical-property relevance data for each of a plurality of intervals obtained by dividing the spectrum data, as is required by claim 15.
Similar to Kwon, Yu and Kasahara, Zabalza teaches deriving multi-dimensional physical property relevance data (i.e. features) that is related to multi-dimensional physical property data (i.e. a hyperspectral image) by applying at least a part of an autoencoder to the multi-dimensional physical-property data, wherein the multi-dimensional physical-property data includes image data of a spectrum which is represented by spectrum data:
Hyperspectral imaging (HSI) is a very motivating field dealing with several different challenges in the last decade. The HSI cameras and devices provide a spatial 2-D image in hundreds of different wavelengths from the electromagnetic spectrum in nature (spectral bands). As a result, a 3-D structure called hypercube is obtained, where each pixel in the 2-D image is represented by an array of spectral values. Obviously, with such amount of information, the use of HSI data for applications including remote classification of image pixels is proving promising, although it demands advanced signal processing applied to stages such as feature extraction or data reduction.
In the last 2–3 decades, a number of methods have been proposed for feature extraction and data reduction in HSI, including both well-known classical techniques and new approaches. These feature extraction and data reduction techniques aim to boost the general data analysis procedures by improving the characterization of features (efficacy) and/or relieving computational complexity (efficiency). For instance, features containing adequate information usually lead to higher classification accuracy of pixels and, in many cases, this can be done along with a reduction in the number of features (feature dimensionality), which in turn increases the overall efficiency. Although there are many methodologies, in this paper we focus on a particular approach related to a new and really promising field, the deep learning (DL) framework, in particular with the study of stacked autoencoders (SAEs).
Based on neural network architectures, SAEs are able to reduce feature dimensionality to few elements contained in the deep layers of those networks. In SAEs, an input pixel of the HSI image is introduced in the network by the first layer (or input layer), with as many nodes as original features (spectral bands) in the pixel. Then, the pixel information travels the network through subsequent layers with reduced number of nodes or units, to finally achieve a reconstructed pixel at the output matching the original one. Therefore, SAEs can be employed effectively for feature extraction, where the abstraction level achieved in deep layers leads to representative reduced features. In that sense, the powerful capabilities from machine learning can be exploited to perform data reduction in such context that seems promising and needs proper investigation.
(Section 1 “Introduction”; emphasis added and internal citations omitted).
In hyperspectral remote sensing, SAEs can be used for feature reduction in the spectral domain of pixels, in an unsupervised manner. After training the SAE with a representative portion of samples, every pixel can then be reduced to the output values (y) of the deepest layer.
The training process in SAEs consists of an iterative update of the multiple internal coefficients w and b, an update by which the error between the input pixel and the reconstructed one at the output of the network is progressively reduced until it is below some value or threshold. An effective training translates into a reduced error as expressed in Eq. (2), which ensures appropriate internal features. Fig. 3 shows both the original spectral data and the reconstructed profile after an appropriate training of the SAE, where the similarity between both profiles is clear. This similarity proves that the SAE network is able to reconstruct the input pixel from internal layers with reduced number of nodes, i.e., the reduced features F from the internal layer are representative and contain the main information from the pixel by high abstraction, being possible to employ them for feature extraction.
(Section 3 “Stacked autoencoders”).
Regarding the claimed invention, Zabalza further teaches deriving the multi-dimensional physical-property relevance data for each of a plurality of intervals (i.e. segments/regions) obtained by dividing the spectrum data:
However, the use of SAEs with HSI data can be complex, due to the hundreds of spectral bands available in the hypercubes. Hidden units in the SAEs layers are required to evaluate the input and derived values from all the spectral bands simultaneously in the same activation functions, and this complexity makes more difficult to find appropriate abstraction. As a result, the main motivation of the present work is to evaluate the SAEs and to propose an alternative solution to address these two main problems: the computational complexity in the implementations, and the lack of proper abstraction in the features, i.e., the limited accuracy in classification analysis.
To this end, we propose a spectral segmentation in the pixels or samples that can divide the complexity and also allow local extraction of features, eventually providing better extraction capability. In this paper, the segmented SAE (S-SAE) method is introduced, where local SAEs are applied to different segments of the spectrum. By locally working in spectral regions, the computational complexity is reduced and, at the same time, the resulting features are improved thus better classification accuracy is obtained thanks to local extraction of information. From our results it is found that, yet with reduced complexity, S-SAE performs better than the conventional SAE implementation and also other state-of-the-art methods in land-cover analysis, which leaves an open door for future investigation and related ideas.
(Section 1 “Introduction”; emphasis added).
The conventional application of SAEs treats equally and simultaneously all spectral bands. This yields complexity because hidden nodes in the first layer deal directly with the original feature dimension, which seems excessive. In addition, there are no considerations with relation to the different spectral regions of the data, while it is usual to find particular local regions with more information than others.
For all that, SAEs application can be implemented by parts, into different segments of the spectrum. This concept was already introduced for other feature extraction methods such as principal component analysis (PCA), segmented PCA and other similar variants.
Fig. 4 presents the generic structure of our proposed S-SAE, where the spectral domain of samples p is segmented into K different regions pk;
k
∈
[
1
,
K
]
to which the SAE technique is applied individually.
Since local SAEs have a small region of the spectrum as input, they present reduced number of hidden nodes (Lk, Fk), i.e., S-SAE needs several SAEs but they are simpler than the one employed in the conventional case. In addition, abstraction from the deep layers is achieved in an easier way. Finally, reduced features from local regions yk;
k
∈
[
1
,
K
]
are concatenated
∑
k
=
1
K
F
k
=
F
to form a reduced feature vector.
(Section 4 “Segmented SAE”; internal citations omitted and emphasis added).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu, Kasahara and Zabalza before the effective filing date of the claimed invention, to modify the learning apparatus taught by Kwon, Yu and Kasahara such that the first processor derives the multi-dimensional physical-property relevance data for each of a plurality of intervals obtained by dividing the spectrum data, as is taught by Zabalza. It would have been advantageous to one of ordinary skill to utilize such a combination because it can result in “reduced complexity but improved efficacy of data abstraction and accuracy of data classification” as is taught by Zabalza (see the Abstract). Accordingly, Kwon, Yu, Kasahara and Zabalza are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 15.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Kwon and Yu, which is described above, and also over the article entitled “Flow Synthesis of Functional Materials” by Sebastian et al. (“Sebastian”).
Regarding claim 17, Kwon and Yu teach a learning apparatus like that of claim 1, as is described above, which is configured to derive learning input data to be input to a machine learning model for predicting a quality of a product. Kwon and Yu, however, do not explicitly disclose that the product is produced using a flow synthesis method, as is required by claim 17.
Sebastian nevertheless suggests that flow synthesis can be employed to produce a variety of different products:
Advanced functional materials occupy a prominent place in the day-to-day life of a significant portion of the global population. The impact of these materials on the quality of human life is seen either in the direct form (viz., communication devices, energy storage devices, light-emitting devices, reflecting mirrors, currency notes, precious metals and their various forms and compositions, cosmetics, diagnostics, paints and coatings, stain resistant fabric, biomimetic colorants, etc.) or indirectly (viz,. catalysts, display technologies, toughened surfaces of high speed machining components, drag-reducing lubricants, functionalized nanosilica for precise chemical separations, light-weight materials, nanocomposites in conducing inks, etc.). In general, functional materials can be classified based on their functionality and quality of performance. A variety of these functional materials is used over a very wide range of quantities (from a few mg to few tons).
The functionality of these materials depends on the dimensions of the material, viz., size and shape for particulate matters, pore size and surface area in the case of porous materials, and thickness for films. This implies that attaining the desired dimensions through controlled synthesis is the key to retain the properties that make these materials functional in true sense. A few such examples that highlight the impact of size on the specific property and hence application are given in Figure 1. The only way to achieve such consistency in the properties is through wet chemical synthesis or through biological routes. Among the two, the later approach is precise yet unreliable due to the scarcity and purity of specific biological moieties needed for certain activity and the former option of chemical synthesis becomes a more reliable approach, albeit only if the synthetic recipe allows consistent product quality at all scales of production. Continuous-flow synthesis of the functional materials not only paves the way to achieve consistency in properties but it is also scalable and yet decentralized, giving it a flexibility of on-site-on-demand production. This article aims to provide developing a broader perspective of the utility and relevance of flow synthesis for the manufacture of a variety of functional materials and also gazes in to the crystal ball, developing a map for exploration in as yet new areas, which might rise to prominence in the coming years.
(Section 1 “Introduction”; emphasis added).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu and Sebastian before the effective filing date of the claimed invention, to apply the learning apparatus taught by Kwon and Yu to a product produced using a flow synthesis method like taught by Sebastian. It would have been advantageous to one of ordinary skill to utilize such a combination because flow synthesis “not only paves the way to achieve consistency in properties but it is also scalable and yet decentralized, giving it a flexibility of on-site-on-demand production,” as is taught by Sebastian (see Section 1 “Introduction,” which is excerpted above). Accordingly, Kwon, Yu and Sebastian are considered to teach, to one of ordinary skill in the art, a learning apparatus like that of claim 17.
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Kwon and Yu, which is described above, and also over U.S. Patent Application Publication No. 2017/0220951 to Chidlovskii et al. (“Chidlovskii”).
Regarding claim 18, Kwon and Yu teach a learning apparatus like that of claim 1, as is described above, which comprise a first processor configured to derive learning input data, input the learning input data to a machine learning model, perform learning, and output the machine learning model as a learned model to be provided for actual operation. Like in claim 18, Kwon further teaches: (i) acquiring the learned model (i.e. a classifier); (ii) acquiring multi-dimensional physical-property relevance data (e.g. extracting features) for prediction which is data of a product of which a quality is unknown; (iii) inputting the multi-dimensional physical-property relevance data for prediction to the learned model and predicting the quality (i.e. classifying any defects); and (iv) controlling outputting of a prediction result of the quality by the learned model (see e.g. paragraphs 0007, 0046 and 0063-0071). Kwon and Yu, however, do not disclose that these tasks are implemented on an operating apparatus comprising a second processor, which acquires the learned model from the first processor of the learning apparatus, as is required by claim 18.
Training a machine learning model on a first device and then providing the machine learning model to a second device to perform classification on input data is nevertheless taught in the art. Chidlovskii for example describes a system comprising a first apparatus (i.e. a machine learning device) that acquires learning input data (i.e. training instances), inputs the learning input data to a machine learning model (i.e. a classifier), performs learning, and then outputs the machine learning model as a learned model to be provided for actual operation (see e.g. paragraphs 0015-0020). Chidlovskii further teaches that a second apparatus comprising a second processor can then: (i) acquire the learned model which is output by the processor of the first apparatus; (ii) acquire multi-dimensional data for prediction which is data of which a classification is unknown (i.e. acquire an unlabeled input instance comprising a feature vector); (iii) input the multi-dimensional data for prediction to the learned model and predict the classification (i.e. a label); and (iv) output the prediction result by the learned model (see e.g. paragraphs 0015 and 0020-0021).
It would have been obvious to one of ordinary skill in the art, having the teachings of Kwon, Yu and Chidlovskii before the effective filing date of the claimed invention, to modify the learning apparatus taught by Kwon and Yu such that it similarly provides the learned model to a second apparatus comprising a second processor, which is configured to: (i) acquire the learned model which is output by the processor of the learning apparatus; (ii) acquire multi-dimensional data for prediction which is data of which a classification is unknown (i.e. acquire the multi-dimensional physical-property relevance data for prediction which is data of a product of which a quality is unknown); (iii) input the multi-dimensional data for prediction to the learned model and predict the classification (i.e. input the multi-dimensional physical-property relevance data for prediction to the learned model and predict the quality); and (iv) output the prediction result (i.e. of the quality) by the learned model, as is taught by Chidlovskii. It would have been advantageous to one of ordinary skill to utilize such a combination, because this would enable more suitable computers to perform the learning and classification tasks, as is suggested by Chidlovskii (see e.g. paragraph 0015). Accordingly, Kwon, Yu and Chidlovskii are considered to teach, to one of ordinary skill in the art, a learning input apparatus like that of claim 18.
Response to Arguments
The Examiner acknowledges the Applicant’s amendments to claims 1, 19 and 20. The Applicant’s arguments regarding the pending claims have been considered, but are moot in view of the new grounds of rejection presented above.
Conclusion
The prior art made of record on form PTO-892 and not relied upon is considered pertinent to applicant’s disclosure. The applicant is required under 37 C.F.R. §1.111(C) to consider these references fully when responding to this action. In particular, the article by Ning et al. cited therein (“Optical Emission Spectrum Processing Using Wavelet Compression During Wafer Fabrication”) generally teaches obtaining spectrum data by performing spectroscopic analysis on a wafer, wherein the spectrum data can be used to identify faults during wafer fabrication.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BLAINE T BASOM whose telephone number is (571)272-4044. The examiner can normally be reached Monday-Friday, 9:00 am - 5:30 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, Kieu Vu can be reached at (571)272-4057. 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.
/BTB/
3/30/2026
/Mariela Reyes/Supervisory Patent Examiner, Art Unit 2142
Prosecution Insights