DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of the Claims
Claims 1-17 and 22-24 are pending for examination.
Claims 18-21 are withdrawn from consideration.
Claims 1-17 and 22-24 are rejected under 35 U.S.C. §103.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-2, 4-5, 7-10, 12-13, 15-17, 22 and 24 is/are rejected under 35 U.S.C. 103 as being unpatentable over Li et al. (U.S. 10,929,744 hereinafter Li) in view of Savvides et al. (U.S. 2018/0053091 hereinafter Savvides).
As Claim 1, Li teaches a method for processing data, comprising:
obtaining a group of data to be quantized for a machine learning model (Li (col. 12 line 50-52, fig. 9 item 920), training data are quantized using fixed-point forward calculation);
using a plurality of point locations to respectively quantize the group of data to be quantized to determine a plurality of groups of quantized data (Li (col. 12 line 52-57, 65-67, fig. 9 item 920, col. 9 line 6-9), system uses a training data set to conduct a fixed-point forward calculation of the fixed-point neural network. The calculation of fixed-point calculation is repeated until desired accuracy is achieved. Converting single values to fixed-point numbers includes determining the length of integer bits, determining the length of decimal bits and conducting fixed-point conversion), wherein
each of the plurality of point locations specifies a position of a decimal point in the plurality of groups of quantized data (Li (col. 9 line 6-9), converting single values to fixed-point numbers includes determining the length of integer bits, determining the length of decimal bits and conducting fixed-point conversion); and
selecting a point location from the plurality of point locations to quantize the group of data to be quantized (Li (col. 5 line 65-66, col. 6 line 5-7 and 10-14, col. 15 line 16-19), neural network produces fixed-point calculation for a value) based on the differences between each of the plurality of groups of quantized data and the group of data to be quantized (Li (col. 13 line 4-11), calculation error between fixed-point calculation and the actual result is used to adjust the model).
Li may not explicitly disclose:
a first piece of data within the group of data to be quantized is quantized to obtain at least first quantized data and second quantized data, the first quantized data and the second quantized data correspond to the first piece of data, and the first quantized data and the second quantized data have different point locations;
wherein each of the differences indicates a quantization precision loss between a respective one of the plurality of groups of quantized data and the group of data to be quantized
wherein the first piece of data is quantized according to the selected point location.
Savvides teaches:
a first piece of data within the group of data to be quantized (Savvides (¶0035 line 3-4), the trained network is unmodified before proceeding to the compression step 56) is quantized to obtain at least first quantized data and second quantized data (Savvides (¶0035, fig. 7), “The sign preservation (block 92) and recoding (block 94) steps are repeated for each value in the matrix produced via the training step 54. Next, a recoding limit is adjusted (block 96). As described above, recoding may adjust the number of bits to approximately eight or nine. At block 96, this recoding is evaluated to determine whether accuracy is significantly decreased. If so, the recoding is adjusted to include more bits”), the first quantized data and the second quantized data correspond to the first piece of data (Savvides (¶0035, fig. 7), “At block 96, this recoding is evaluated to determine whether accuracy is significantly decreased. If so, the recoding is adjusted to include more bits”), and the first quantized data and the second quantized data (Savvides (¶0035, fig. 7), “At block 96, this recoding is evaluated to determine whether accuracy is significantly decreased. If so, the recoding is adjusted to include more bits”, first quantization is the original one, second quantization is the more bits adjustment) have different point locations (Savvides (¶0034 line 17-19), “more bits means the float can be accurate to more decimal places and fewer bits means the float is accurate to fewer decimal places.” Adding more bits is construed as adjusting the decimal places);
wherein each of the differences indicates a quantization precision loss between a respective one of the plurality of groups of quantized data and the group of data to be quantized (Savvides (¶0035, fig. 7), “At block 96, this recoding is evaluated to determine whether accuracy is significantly decreased. If so, the recoding is adjusted to include more bits”, Accuracy is construed as “a quantization precision loss between a respective one of the plurality of groups of quantized data and the group of data to be quantized”)
wherein the first piece of data is quantized according to the selected point location (Savvides (¶0035, fig. 7), “if not, the compression step 56 proceeds. This modified matrix is then saved in a binary form (block 98)”. Figure 7 is a loop until an acceptable accuracy is achieved).
Li discloses a system/method to use machine learning for a neural network quantization. Savvides discloses a system/method to iterate through a plurality of decimal places in order to decide the optimal decimal place. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify quantization method of Li instead be a quantization method selection module taught by Savvides, with a reasonable expectation of success. The motivation would be to allow the system to allow “using the method of the disclosed embodiments, bits may be removed to reduce the size of the network while simultaneously maintaining sufficient accuracy to run the network” (Savvides (¶0034)) (Teaching Suggestion Motivation).
As Claim 2, besides Claim 1, Li in view of Savvides teaches wherein each of the plurality of point locations is represented by an integer, and the method further includes: obtaining one of the plurality of point locations based on a range associated with the group of data to be quantized; and determining other point locations of the plurality of point locations based on integers adjacent to the obtained point location (Li (col. 9 line 6-9), converting single values to fixed-point numbers includes determining the length of integer bits, determining the length of decimal bits and conducting fixed-point conversion).
As Claim 4, besides Claim 1, Li in view of Savvides teaches wherein the selecting the point location from the plurality of point locations includes: determining a plurality of differences between the plurality of groups of quantized data and the group of data to be quantized respectively; selecting the smallest difference from the plurality of differences (Li (col. 13 line 4-11), calculation error between fixed-point calculation and the actual result is used to adjust the model); and selecting a point location corresponding to the smallest difference from the plurality of point locations (Li (col. 12 line 59-67, fig. 9 item 930, 932), system calculate floating point gradient. Floating point locations are selected when desired accuracy is achieved).
As Claim 5, besides Claim 4, Li in view of Savvides teaches wherein respectively the determining the plurality of differences between the plurality of groups of quantized data and the group of data to be quantized includes: for a given group of quantized data of the plurality of groups of quantized data, determining a group of relative differences between the given group of quantized data and the group of data to be quantized, respectively Li (col. 13 line 4-11), calculation error between fixed-point calculation and the actual result is used to adjust the model); and determining one of the plurality of differences based on the group of relative differences (Li (col. 12 line 59-67, fig. 9 item 930, 932), system calculate floating point gradient. Floating point locations are selected when desired accuracy is achieved).
As Claim 7, besides Claim 1, Li in view of Savvides teaches wherein the group of data to be quantized includes a group of floating-point numbers in a neural network model, and the method further includes: using the selected point location to quantize the group of data to be quantized to obtain a group of quantized data, wherein quantizing the group of data to be quantized includes: mapping the group of data to be quantized to the group of quantized data based on the selected point location, wherein the position of the decimal point in the group of quantized data is determined by the selected point location (Li (col. 12 line 50-52, fig. 9 item 920), training data are quantized using fixed-point forward calculation) and inputting the obtained group of quantized data to the neural network model for processing (Li (col. 12 line 52-57, 65-67, fig. 9 item 920, col. 9 line 6-9), system uses a training data set to conduct a fixed-point forward calculation of the fixed point neural network. The calculation of fixed-point calculation is repeated until desired accuracy is achieved. Converting single values to fixed-point numbers includes determining the length of integer bits, determining the length of decimal bits and conducting fixed-point conversion).
As Claim 8, besides Claim 1, Li in view of Savvides teaches further including: obtaining another group of data to be quantized including a group of floating-point numbers in a neural network model; using the selected point location to quantize the other group of data to be quantized to obtain another group of quantized data, wherein quantizing the another group of data to be quantized includes: mapping the another group of data to be quantized to the other group of quantized data based on the selected point location (Li (col. 12 line 50-52, fig. 9 item 920), training data are quantized using fixed-point forward calculation), wherein the position of the decimal point in the another group of quantized data is determined by the selected point location (Li (col. 12 line 52-57, 65-67, fig. 9 item 920, col. 9 line 6-9), system uses a training data set to conduct a fixed-point forward calculation of the fixed point neural network. The calculation of fixed-point calculation is repeated until desired accuracy is achieved. For each iteration, a different decimal point is calculated); and inputting the obtained another group of quantized data to the neural network model for processing (Li (col. 12 line 52-57, 65-67, fig. 9 item 920, col. 9 line 6-9), system uses a training data set to conduct a fixed-point forward calculation of the fixed point neural network. The calculation of fixed-point calculation is repeated until desired accuracy is achieved. Converting single values to fixed-point numbers includes determining the length of integer bits, determining the length of decimal bits and conducting fixed-point conversion).
As Claim 9-10, 12-13 and 15-6, the Claims are rejected for the same reasons as Claims 1-2, 4-5 and 7-8, respectively.
As Claim 17, 22 and 24, the Claims are rejected for the same reasons as Claim 1, 2 and 4, respectively.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 3, 6, 11, 14 and 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Li in view of Savvides in further view of Yoda et al. (U.S. 2019/0114142 hereinafter Yoda).
As Claim 3, besides Claim 2, Li in view of Savvides does not explicitly disclose:
wherein the determining the other point locations of the plurality of point locations includes at least one of the following: incrementing an integer representing the point location to determine one of the other point locations; or decrementing an integer representing the point location to determine one of the other point locations.
Yoda teaches:
wherein the determining the other point locations of the plurality of point locations includes at least one of the following: incrementing an integer representing the point location to determine one of the other point locations; or decrementing an integer representing the point location to determine one of the other point locations (Yoda (¶0093 line 13-22), system increases or decreases the integer part by 1 bit).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify training system of Li in view of Savvides instead be a training system taught by Yoda, with a reasonable expectation of success. The motivation would be to “to reduce the overhead in a program for acquisition of statistical information” and “an application program executed by the information processing apparatus acquires statistical information from the processor to optimize the decimal point position” (Yoda (¶0101 line 3-4 and 8-9)) (Teaching Suggestion Motivation).
As Claim 6, besides Claim 4, Li in view of Savvides does not explicitly disclose:
wherein respectively the determining the plurality of differences between the plurality of groups of quantized data and the group of data to be quantized includes: for a given group of quantized data of the plurality of groups of quantized data, determining a quantized mean value of the given group of quantized data and an original mean value of the group of data to be quantized, respectively; and
determining one of the plurality of differences based on the quantized mean value and the original mean value.
Yoda teaches:
wherein respectively the determining the plurality of differences between the plurality of groups of quantized data and the group of data to be quantized includes: for a given group of quantized data of the plurality of groups of quantized data, determining a quantized mean value of the given group of quantized data and an original mean value of the group of data to be quantized, respectively (Yoda (¶0191 line 14-22, fig. 31), decimal point position of fixed point data is set according to a distribution ranges of digits); and
determining one of the plurality of differences based on the quantized mean value and the original mean value (Yoda (¶0191 line 14-22, fig. 31), decimal point position of fixed point data is set according to a distribution ranges of digits).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify training system of Li in view of Savvides instead be a training system taught by Yoda, with a reasonable expectation of success. The motivation would be to “to reduce the overhead in a program for acquisition of statistical information” and “an application program executed by the information processing apparatus acquires statistical information from the processor to optimize the decimal point position” (Yoda (¶0101 line 3-4 and 8-9)) (Teaching Suggestion Motivation).
As Claims 11 and 14, the Claims are rejected for the same reasons as Claim 3 and 6, respectively.
As Claim 23, the Claim is rejected for the same reasons as Claim 3.
Response to Arguments
Applicant’s Response to the 35 U.S.C. §103 Rejections:
(A) Li does not taches “selecting a point location based on differences between groups of quantized data and the group of data to be quantized”
Applicants’ arguments are fully considered but are not persuasive. Savvides teaches the limitation(s). See the current rejection(s) for details.
(B) Savvides does not cure the deficiencies of Li
As per Li, Applicant argues that “Savvides teaches changing total bit width, not point location” in paragraphs (¶0034, ¶0035) because “changing the total number of bits used to represent value is not the same as changing a “point location” as claimed (second paragraph of page 4 in the remarks).
PNG
media_image1.png
236
640
media_image1.png
Greyscale
Applicant’s arguments are fully considered but are not persuasive. Savvides (¶0034 line 17-19) teaches “more bits means the float can be accurate to more decimal places and fewer bits means the float is accurate to fewer decimal places.” Adding more bits is construed as adjusting the decimal places.
As per Claim 1, Applicant argues that there is no motivation to combine Li and Savvides (third paragraph of page 5 in the remarks).
PNG
media_image2.png
102
654
media_image2.png
Greyscale
Applicant’s arguments are fully considered but are not persuasive. Li discloses a system/method to use machine learning for a neural network quantization. Savvides discloses a system/method to iterate through a plurality of decimal places in order to decide the optimal decimal place. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify quantization method of Li instead be a quantization method selection module taught by Savvides, with a reasonable expectation of success. The motivation would be to allow the system to allow “using the method of the disclosed embodiments, bits may be removed to reduce the size of the network while simultaneously maintaining sufficient accuracy to run the network” (Savvides (¶0034)) (Teaching Suggestion Motivation).
Other independent/dependent Claims are not allowable for the same reasons above.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Koster et al. (U.S. 2017/0316307) teaches a decimal position that provides highest precision without causing an overflow.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NHAT HUY T NGUYEN whose telephone number is (571)270-7333. The examiner can normally be reached M-F: 12:00-8:00 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, Viker Lamardo can be reached at 571-270-5871. 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.
/NHAT HUY T NGUYEN/Primary Examiner, Art Unit 2147
Prosecution Insights