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 .
Claims 1-4, 6-19, and 21 are pending in this application. Claims 1, 6, 8, and 21 are amended and claims 5 and 20 are canceled by applicant’s amendment filed 23 February 2026.
Claim Objections
Claims 8-19 and 21 are objected to because of the following informalities:
Regarding Claim 8, it recites “the integrated ciruit” (third-to-last line). This appears to contain a typographical error, and should recite “the integrated circuit.”
Regarding Claim 21, it also recites “the integrated ciruit” (last line). This appears to contain a typographical error, and should recite “the integrated circuit.”
Regarding Claims 9-19, they are objected to as being dependent on an objected base claim.
Appropriate correction is required.
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 1-4, 7, and 21 are rejected under 35 U.S.C. 103 as being unpatentable over van der Made et al. (U.S. 2017/0229117, hereinafter “van der Made”) in view of Shafiee, Ali, et al. (“ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars,” ACM SIGARCH Computer Architecture News 44.3 (2016): 14-26; hereinafter “Shafiee”).
Regarding Claim 1, van der Made teaches an integrated circuit for signal detection in an offline state (fig. 2; ¶ [0024]), comprising:
a host processor coupled with a co-processor and configured to receive a signal stream (fig. 2, Microprocessor 204 and spiking neuron adaptive processor 208; ¶ [0026] – [0028]);
a neuromorphic the co-processor including a neural network that is configured to identify one or more target signals from the received signal stream (fig. 2, spiking neuron adaptive processor 208; ¶ [0027] – [0028]—the spiking neural networks identify target signals in the received audio stream);
a communications interface between the host processor and the co-processor configured to transmit information therebetween (¶ [0028]—commands sent between the neuromorphic co-processor and the host processor indicate a communication interface); and
wherein, in the absence of the cloud connection, the co-processor operates to locally classify and differentiate the received signal stream, including distinguishing recognized keywords, identifying specific speakers, and categorizing image classes (¶ [0023] – [0024]—the co-processor differentiates the received signal stream to distinguish recognized keywords. van der Made is directed towards a system that operates in small, low-power, portable devices, as described in ¶ [0010] and [0023] – [0024]; there is no mention of a cloud connection or other type of server because the system functions in a standalone manner).
van der Made does not specifically teach:
the neural network disposed in a multiplier array;
wherein a set of weights produced by training the neural network are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and
wherein the weights of the neural network are stored in a non-volatile memory accessible to the neuromorphic co-processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection; and
wherein, in the absence of the cloud connection, the co-processor accesses the stored weights to locally classify and differentiate the received signal stream.
However, Shafiee teaches:
a neural network disposed in a multiplier array (section II. D. and fig. 1—a neural network is implemented in a crossbar array that performs multiply operations. A system including the neural network is shown in section III and fig. 2);
wherein a set of weights produced by training the neural network are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network (section II. D. and fig. 1—the memristor array is a core that performs multiplication of input current values by weights to provide output current values); and
wherein the weights of the neural network are stored in a non-volatile memory accessible to the neuromorphic co-processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection (sections II. D and IV, and fig. 2—the ISAAC system stores weights in a memristor array, which is a non-volatile memory accessible to the neuromorphic co-processor. Section III explains that the weights are loaded into the memristor array after training. The training is not performed in-the-field, but it is clear that inference is performed in-the-field by performing in-situ multiply-accumulate operations using only the weights stored in the memristor, i.e. it is performed in an offline state without an active cloud connection); and
wherein, in the absence of the cloud connection, the co-processor accesses the stored weights to locally classify and differentiate the received signal stream, including distinguishing recognized keywords, identifying specific speakers, and categorizing image classes (p. 17—weights are stored in the memristor cells for performing inference locally, without a cloud connection. The inference may categorize image classes, such as the face detection problem described on p. 16, second paragraph and in the experiments using the DeepFace dataset in section VIII).
All of the claimed elements were known in van der Made and Shafiee and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the multiplier array and input and output currents of Shafiee with the co-processor and neural network of van der Made to yield the predictable result of the co-processor including a neural network disposed in a multiplier array that is configured to identify one or more target signals from the received signal stream; wherein a set of weights produced by training the neural network are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and wherein the weights of the neural network are stored in a non-volatile memory accessible to the neuromorphic co-processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection; and wherein, in the absence of the cloud connection, the co-processor accesses the stored weights to locally classify and differentiate the received signal stream, including distinguishing recognized keywords, identifying specific speakers, and categorizing image classes. One would be motivated to make this combination for the purpose of improving performance and efficiency, such as by reducing the high overheads of analog-to-digital conversion in neural network accelerators (Shafiee, Abstract).
Regarding Claim 2, van der Made/Shafiee teaches the signal stream is comprised of signals received by way of any of sensors comprised of any of infrared sensors, pressure sensors, temperature sensors, proximity sensors, motion sensors, fingerprint scanners, photo eye sensors, wireless signal antennae (van der Made, ¶ [0020] and [0025]—the sensor may be an image sensor or camera).
Regarding Claim 3, van der Made/Shafiee teaches the signal stream is comprised of any of speech or non-verbal acoustic signals received by way of a microphone, and images types or classes received by a smart camera (van der Made, ¶ [0025] – [0027]—the signal stream may come from a microphone).
Regarding Claim 4, van der Made/Shafiee teaches the one or more target signals are comprised of any of spoken keywords, specific sounds, desired image types or classes, and signal patterns among sensor data (van der Made, ¶ [0023]).
Regarding Claim 7, van der Made/Shafiee teaches the offline state is comprised of an absence of connectivity between the integrated circuit and an external communications network (Shafiee, section III—the weights are loaded into the memristor array after training. The training is not performed in-the-field, but it is clear that inference is performed in-the-field by performing in-situ multiply-accumulate operations using only the weights stored in the memristor, i.e. it is performed in an offline state without an active cloud connection).
Regarding Claim 21, van der Made teaches a system for signal detection in an offline state (fig. 2; ¶ [0024]), comprising a host processor on a first integrated circuit configured to receive a signal stream (fig. 2, Microprocessor 204; ¶ [0026]);
a neuromorphic co-processor on a second integrated circuit including an artificial neural network that is configured to identify one or more target signals among one or more signals received from the host processor (fig. 2, spiking neuron adaptive processor 208; ¶ [0027] – [0028]—the spiking neural networks identify target signals in the received audio stream); and
a communications interface between the host processor and the co-processor configured to transmit information therebetween (¶ [0028]—commands sent between the neuromorphic co-processor and the host processor indicate a communication interface);
whereby any one of recognized keywords, specific speakers, and classes of images are differentiated using signals received by the integrated ciruit in the absence of the cloud connection (¶ [0023] – [0024]—the co-processor differentiates the received signal stream to distinguish recognized keywords. van der Made is directed towards a system that operates in small, low-power, portable devices, as described in ¶ [0010] and [0023] – [0024]; there is no mention of a cloud connection or other type of server because the system functions in a standalone manner).
van der Made does not specifically teach:
wherein a set of weights are produced by way of training a neural network and translated into a weight file suitable for being stored in a memory storage that is accessible to the integrated circuit;
wherein the set of weights are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and
wherein the weights are stored and downloaded in a non-volatile memory accessible to the neuromorphic co- processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection.
However, Shafiee teaches:
wherein a set of weights are produced by way of training a neural network and translated into a weight file suitable for being stored in a memory storage that is accessible to the integrated circuit (section II. B. and fig. 1—training generates weights that are stored in the crossbar array {i.e. in a weight file});
wherein the set of weights are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network (section II. D. and fig. 1—the memristor array is a core that performs multiplication of input current values by weights to provide output current values); and
wherein the weights are stored and downloaded in a non-volatile memory accessible to the neuromorphic co- processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection (sections II. D and IV, and fig. 2—the ISAAC system stores weights in a memristor array, which is a non-volatile memory accessible to the neuromorphic co-processor. Section III explains that the weights are loaded into the memristor array after training. The training is not performed in-the-field, but it is clear that inference is performed in-the-field by performing in-situ multiply-accumulate operations using only the weights stored in the memristor, i.e. it is performed in an offline state without an active cloud connection);
whereby any one of recognized keywords, specific speakers, and classes of images are differentiated using signals received by the integrated ciruit in the absence of the cloud connection (p. 17—weights are stored in the memristor cells for performing inference locally, without a cloud connection. The inference may categorize image classes, such as the face detection problem described on p. 16, second paragraph and in the experiments using the DeepFace dataset in section VIII).
All of the claimed elements were known in van der Made and Shafiee and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the multiplier array and input and output currents of Shafiee with the co-processor and neural network of van der Made to yield the predictable result of wherein a set of weights are produced by way of training a neural network and translated into a weight file suitable for being stored in a memory storage that is accessible to the integrated circuit; wherein the set of weights are multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and wherein the weights are stored in a non-volatile memory accessible to the neuromorphic co- processor, such that the integrated circuit is configured to recognize target signals in an offline state without requiring an active cloud connection. One would be motivated to make this combination for the purpose of improving performance and efficiency, such as by reducing the high overheads of analog-to-digital conversion in neural network accelerators (Shafiee, Abstract).
Claims 6-7 are rejected under 35 U.S.C. 103 as being unpatentable over van der Made in view of Shafiee, as applied to claims 1 and 4, above, in view of Khellah et al. (U.S. 2019/0115011, hereinafter “Khellah”).
Regarding Claim 6, van der Made/Shafiee does not specifically teach wherein the set of weights comprises a programmed file that is formed by way of training an external software model of the artificial neural network to recognize the one or more target signals. However, Khellah teaches wherein the set of weights comprises a programmed file that is formed by way of training an external software model of the artificial neural network to recognize the one or more target signals (Khellah, ¶ [0061] and [0073]—training may be performed by 33an external computer and transmitted to the integrated circuit over a network).
All of the claimed elements were known in van der Made/Shafiee and Khellah and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the programmed file of Khellah with the weights of van der Made/Shafiee to yield the predictable result of wherein the set of weights comprises a programmed file that is formed by way of training an external software model of the artificial neural network to recognize the one or more target signals. One would be motivated to make this combination for the purpose of facilitating training of the neural network to detect keywords, as described by Khellah, ¶ [0016].
Regarding Claim 7, van der Made/Shafiee/Khellah teaches the offline state is comprised of an absence of connectivity between the integrated circuit and an external communications network (Khellah, ¶ [0061]—training may be performed offline and weights are transmitted to the integrated circuit, indicating that the integrated circuit processes input using the weights in an absence of connectivity to an external network by using the stored weights).
Claims 8-12 and 16-19 are rejected under 35 U.S.C. 103 as being unpatentable over Hofer et al. (U.S. 2016/0379629, hereinafter “Hofer”) in view of Parada San Martin et al. (U.S. 2015/0127594, hereinafter “Parada”), and further in view of Shafiee.
Regarding Claim 8, Hofer teaches a method for generating a weight file that causes an integrated circuit to detect desired user-specified signals (¶ [0020], [0024], and [0026]), comprising:
listing desired target signals that may be detected by a signal detector (¶ [0020]—word lists are desired target signals);
retrieving one or more signal databases that are comprised of standard target signals that may be detected by the signal detector (¶ [0024]—the static or pre-built vocabulary is a database of standard target signals);
combining the desired target signals and the one or more signal databases to build a modified database (¶ [0024] and [0037]); and
using the modified database to recognize the target signals and the standard signals (¶ [0027]); and
enabling differentiation between signals received by the integrated ciruit in the absence of the cloud connection, so as to classify any one of recognized keywords, specific speakers, and classes of images (fig. 1; ¶ [0016], [0020], and [0026] – [0031]—the integrated circuit operates as a standalone device to classify recognized keywords. There is no mention of or need for a cloud connection. ¶ [0002] and [0023] state that the invention is directed towards a device with limited resources, in contrast to a system that connects to a remote server.).
Hofer teaches training the neural network implementation (¶ [0001]), but does not explicitly teach:
using the modified database to train a neural network implementation to recognize the target signals and the standard signals;
producing a set of weights by way of training the neural network implementation; and
translating the set of weights into the weight file suitable for being stored in a memory storage that is accessible to the integrated circuit;
wherein the weights are:
i) multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and
ii) stored in a non-volatile memory accessible to the co-processor, such that the integrated circuit is configured to recognize target signals in an offline state.
However, Parada teaches:
using a modified database to train a neural network implementation to recognize target signals and standard signals (¶ [0017] – [0018] and [0036] – [0037]);
producing a set of weights by way of training the neural network implementation (¶ [0003] – [0004] and [0034]); and
translating the set of weights into the weight file suitable for being stored in a memory storage that is accessible to the integrated circuit (¶ [0003] – [0004] and [0034]—the weights are clearly stored for use by the neural network; ¶ [0109] describes storing data in files).
These claimed elements were known in Hofer and Parada and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the training a neural network of Parada with the neural network and weights of Hofer to yield the predictable result of using the modified database to train a neural network implementation to recognize the target signals and the standard signals; producing a set of weights by way of training the neural network implementation; and translating the set of weights into the weight file suitable for being stored in a memory storage that is accessible to the integrated circuit. One would be motivated to make this combination for the purpose of enabling the neural network to recognize new desired target signals through the process of training.
Hofer/Parada does not specifically teach wherein the weights are:
i) multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and
ii) stored in a non-volatile memory accessible to the co-processor, such that the integrated circuit is configured to recognize target signals in an offline state.
However, Shafiee teaches wherein the weights are:
i) multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network (section II. D. and fig. 1—the memristor array is a core that performs multiplication of input current values by weights to provide output current values); and
ii) stored in a non-volatile memory accessible to the co-processor, such that the integrated circuit is configured to recognize target signals in an offline state (sections II. D and IV, and fig. 2—the ISAAC system stores weights in a memristor array, which is a non-volatile memory accessible to the neuromorphic co-processor. Section III explains that the weights are loaded into the memristor array after training. The training is not performed in-the-field, but it is clear that inference is performed in-the-field by performing in-situ multiply-accumulate operations using only the weights stored in the memristor, i.e. it is performed in an offline state); and
enabling differentiation between signals received by the integrated ciruit in the absence of the cloud connection, so as to classify any one of recognized keywords, specific speakers, and classes of images (p. 17—weights are stored in the memristor cells for performing inference locally, without a cloud connection. The inference may categorize image classes, such as the face detection problem described on p. 16, second paragraph and in the experiments using the DeepFace dataset in section VIII).
All of the claimed elements were known in Hofer/Parada and Shafiee and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the core and input and output currents of Shafiee with the weight file of Hofer/Parada to yield the predictable result of wherein the weights are: i) multiplied in a core by input current values to provide output current values that are combined to arrive at a decision of the neural network; and ii) stored in a non-volatile memory accessible to the neuromorphic co-processor, such that the integrated circuit is configured to recognize target signals in an offline state. One would be motivated to make this combination for the purpose of improving performance and efficiency, such as by reducing the high overheads of analog-to-digital conversion in neural network accelerators (Shafiee, Abstract).
Regarding Claim 9, Hofer/Parada/Shafiee teaches listing comprises entering the target signals into a web-based application that is configured to generate the weight file (Hofer, ¶ [0080], [0094], [0098], and [0101]—many types of graphical and wireless interfaces may be used, including internet devices, which use web-based applications).
Regarding Claim 10, Hofer/Parada/Shafiee teaches listing comprises entering the target signals into a cloud-based application that is configured to generate the weight file (Hofer, ¶ [0080], [0094], [0098], and [0101]—many types of graphical and wireless interfaces may be used, including internet devices, which use cloud-based applications).
Regarding Claim 11, Hofer/Parada/Shafiee teaches listing comprises entering the target signals into a stand-alone software that is configured to generate the weight file (Hofer, ¶ [0072]—programs may be provided on any form of media, including stand-alone software).
Regarding Claim 12, Hofer/Parada/Shafiee teaches the target signals are comprised of signal patterns within input signals received by way of one or more sensors comprised of any of infrared sensors, pressure sensors, temperature sensors, proximity sensors, motion sensors, fingerprint scanners, photo eye sensors, wireless signal antennae (Hofer, ¶ [0077]).
Regarding Claim 16, Hofer/Parada/Shafiee teaches the neural network implementation is a software model of a neural network that is implemented in the integrated circuit comprising the signal detector (Parada, fig. 1; ¶ [0003], [0021], [0034], and [0107]).
Regarding Claim 17, Hofer/Parada/Shafiee teaches the weight file may be provided to an end-user upon purchasing a mobile device (Hofer, ¶ [0016], [0020], and [0023]—the mobile device employs a user-specific vocabulary for recognition, indicating that the weights are determined and provided to an end-user upon purchasing the mobile device).
Regarding Claim 18, Hofer/Parada/Shafiee teaches the weight file may be programmed into one or more chips that may be purchased by an end-user for use in a mobile device (Hofer, ¶ [0016], [0020], and [0023]—the mobile device employs a user-specific vocabulary for recognition, indicating that the weights are determined and provided to an end-user upon purchasing the mobile device, which inherently includes chips in the device that are purchased by the user).
Regarding Claim 19, Hofer/Parada/Shafiee teaches upon an end-user installing the weight file the mobile device, the signal detector may detect the target signals by way of the set of weights (Hofer, ¶ [0026] – [0027]).
Claims 13-14 are is rejected under 35 U.S.C. 103 as being unpatentable over Hofer in view of Parada in view of Shafiee, as applied to claim 8, above, and further in view of van der Made.
Regarding Claim 13, Hofer/Parada/Shafiee does not specifically teach the target signals may be any type of signal that an end-user wants to detect. However, van der Made teaches the target signals may be any type of signal that an end-user wants to detect¶ [0020] and [0025]—the signals may be audio, images, video, or other types from a variety of sensors).
All of the claimed elements were known in Hofer/Parada/Shafiee and van der Made and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the types of signals and sensors of van der Made with target signals of Hofer/Parada/Shafiee to yield the predictable result of the target signals may be any type of signal that an end-user wants to detect. One would be motivated to make this combination for the purpose of expanding the functionality of the device by enabling the use of other sensors on the device to detect target signals.
Regarding Claim 14, Hofer/Parada/Shafiee/van der Made teaches the target signals may be spoken keywords, non- verbal acoustic signals such as specific sounds, image types or classes to be detected by a smart camera (Hofer, ¶ [0029] and [0052]).
Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Hofer in view of Parada in view of Shafiee, as applied to claim 8, above, and further in view of Ramos et al. (U.S. 2019/0244113, hereinafter “Ramos”).
Regarding Claim 15, Hofer/Parada/Shafiee does not specifically teach wherein combining comprises labeling the target signals with corresponding labels and labeling all other signals by way of a generic label. However, Ramos teaches labeling target signals with corresponding labels and labeling all other signals by way of a generic label (¶ [0028] – [0029]—target signals are labelled with a classification for the target and a positive indication, while all other signals that do not comprise a target signal are labelled with the generic “negative” label).
All of the claimed elements were known in Hofer/Parada/Shafiee and Ramos and could have been combined by known methods with no change in their respective functions. It therefore would have been obvious to a person of ordinary skill in the art at the time of filing of the applicant’s invention to combine the labels of Ramos with target signals of Hofer/Parada to yield the predictable result of wherein combining comprises labeling the target signals with corresponding labels and labeling all other signals by way of a generic label. One would be motivated to make this combination for the purpose of improving the efficiency of labeling and classification of target signals (Ramos, ¶ [0002]).
Response to Arguments
The amendments to the claims filed 23 February 2026 are accepted as overcoming the previous claim objections and rejections under 35 U.S.C. 112(b). Note, however, the new objections necessitated by the amendments, as detailed above.
Applicant’s arguments filed 23 February 2026 with respect to the rejections under 35 U.S.C. 103 have been fully considered but they are not persuasive. The examiner finds than the new limitations added to independent claims 1, 8, and 21 are taught by the prior art previously recited. As detailed above, van der Made, Shafiee, and Hofer are all directed towards small devices that perform their operations in a standalone manner. None of them utilize a cloud connection to perform classification or other forms of inference. In fact, these references make little or no reference to a cloud connection, server, or other remote computing system. These prior art refences clearly demonstrate that storing weights in a co-processor and using the weights by a neural network to perform inference such as recognizing keywords and classifying images was well-known in the art at the time of the applicant’s invention.
Conclusion
Applicant’s amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HAL W SCHNEE whose telephone number is (571) 270-1918. The examiner can normally be reached M-F 7:30 a.m. - 6:00 p.m.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Michael Huntley can be reached at 303-297-4307. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/HAL SCHNEE/Primary Examiner, Art Unit 2129