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 . This action is in response to the application filed on February 21, 2024. Claims 1-20 are pending in the application and have been examined.
Specification
The disclosure is objected to because of the following informalities: The disclosure is objected to because of the following informalities: The output h “204” and output h “206” in para. [0026] refers to the same component. The figure component numbers refer to the same component in Fig. 2 where the “206” needs to be corrected to “204” as “206” is not present in the drawings. Appropriate correction is required.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The following is a quotation of pre-AIA 35 U.S.C. 112, sixth paragraph:
An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph:
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 USC 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action.
The limitations in Claim 20 recite a means for receiving an input for processing through a layer of a neural network; and means for generating an output invoking the 35 USC 112(f) interpretation, the “means” being interpreted using para. [0058] of the specification.
Claim Objections
Claims 6 and 16 are objected to because of the following informalities: the limitation calculate the first product by multiplying the first diagonal matrix the first frozen matrix, which appears to be a clear typographical error. For the purpose of examination, the claim will be interpreted as if it had read: “calculate the first product by multiplying the first diagonal matrix, and the first frozen matrix”. Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
In reference to step 1 of the subject matter eligibility guidance, claim 1 recites a processing system, comprising: at least one memory having executable instructions stored thereon; and one or more processors configured to execute the executable instructions, thus an article of manufacture, one of the four statutory categories of patentable subject matter. However, in reference to step 2A prong 1, the claim further recites the step generate an output … based on a first product, pertains to a mathematical calculation, thus the claim recites an abstract idea.
In reference to step 2A prong 2, the claim does not include any additional elements which could integrate the abstract idea into a practical application, because the additional elements consist of:
generic computer components on which to perform the abstract idea (e.g. one or more processors, at least one memory) which, by MPEP 2106.05(f)(2), consists of no more than “using a computer of other machinery as a tool” to perform the mental process, i.e. “apply it”, which cannot integrate the abstract idea into a practical application; and
using a machine learning model to perform the mathematical concept step of generating a result (e.g. for processing through a layer of a neural network) which by MPEP 2106.05(h), consists of “restricting the abstract idea to a particular technological environment” to perform the mental process, cannot integrate the abstract idea into a practical application.
mere data gathering and outputting to perform the abstract idea (e.g. receive an input, generate an output) which can consist of “gathering statistics generated … ; the statistics are then used to calculate”, by MPEP 2106.05(g), which cannot integrate the abstract idea into a practical application.
Thus, the claim is directed to the abstract idea of generating and evaluating a result based on an input.
In reference to step 2B, the additional elements cannot, either alone or in combination, provide an inventive concept nor significantly more than the abstract idea itself, because (a) by MPEP 2106.05(f)(2), merely implementing the abstract idea on generic computer components or other machinery cannot provide an inventive concept, because (a,c) receiving an input and generating an output can be “storing and retrieving information in memory” and “receiving or transmitting data over a network”, which by MPEP 2105.06(g) is well-understood, routine, conventional activity, because (b) the additional element of through a neural network is “merely indicating a field of use or technological environment in which to apply a judicial exception do not amount to significantly more than the exception itself”, by MPEP 2106.05(h), and because there is no nexus between the additional elements which, in combination, can provide significantly more than the abstract idea itself.
Claims 2-5, dependent on Claim 1, merely recites the details of the mathematical concepts being computed in the system (e.g. the first product corresponds to a low-rank decomposition of an accumulated weight update, the first product corresponds to, comprise random matrices) which cannot integrate the abstract idea into a practical application nor provide an inventive concept, by MPEP 2106.05(h).
Claims 6,7, dependent upon Claim 1, merely recites additional steps of the mathematical concept performed by the system, in claim 1, to perform the generating of a result, thus there are no new additional elements introduced, thus no new elements which integrate a practical application or provide significantly more than the abstract idea.
Claim 8, dependent upon Claim 1, merely recites comprises a transformer neural network which is merely the field of use of the system, and thus by MPEP 2106.05(h), limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not amount to significantly more than the abstract idea itself, and cannot integrate a judicial exception into a practical application.
Claim 9, dependent upon Claim 1, merely recites one or more actions based on the generated output, where to take one or more actions, in light of the specification, is as an abstract idea step of inference [0041] as well as forward propagating [0042] demonstrating “to take one or more actions” as an abstract idea being a mental process and mathematical concept respectively. Thus, claim 9 recites no new additional elements, and thus no additional elements which could integrate the abstract idea into a practical application, by MPEP 2106.05(h).
Claims 10, dependent upon Claim 9, merely recites the details of a mental process step of the abstract idea in claim 9 (e.g. to generate an inference based on the generated output), and are thus rejected for reasons set forth in the rejection of Claim 9.
Claims 11, dependent upon Claim 9, merely recites the details of the mathematical concepts in claim 9 (e.g. forward propagate the output of the layer), and are thus rejected for reasons set forth in the rejection of Claim 9.
Claims 12-17, recites the method performed by the system of Claims 1-7, respectively, and are thus rejected for reasons set forth in the rejections of Claims 1-7.
Claims 18 and 19 recite, respectively, the method performed by the system of Claim 10 and Claim 11 and are thus rejected for reasons set forth in the rejection of Claims 10,11.
Claim 20 recites the system of Claim 1 and is thus rejected for reasons set forth in the rejections of Claim 1.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1-5,7-15,17-20 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Zhang et. al, “LORA-FA: MEMORY-EFFICIENT LOW-RANK ADAPTATION FOR LARGE LANGUAGE MODELS FINE-TUNING”.
Regarding Claim 1, Zhang teaches a processing system, comprising: at least one memory having executable instructions stored thereon; and one or more processors configured to execute the executable instructions (Zhang et. al, pg. 6, sec. 4.1, “Due to the limited budget, we conduct our experiments on different devices: we use NVIDIA Turing RTX2080Ti for small-sized RoBERTa models, …, and NVIDIA Ampere A100 for large-sized LLaMA models.” Experimentation on the disclosed method with GPU hardware inherently shows the application of the method requires one or more processors and memory.) to cause the processing system to: receive an input for processing through a layer of a neural network (Zhang et. al, pg. 2, sec. 1, “This is because the large input activation of X needs to be stored during the feed forward pass”); and generate an output of the layer of the neural network (Zhang et. al, pg. 3, sec. 2.1, “linear layers transform the input into query, key, and value, they are fed into the self-attention for interaction, and the attention output is sent to another linear layer.”) based on a first product, the first product being based on a first trainable scaling vector, a first frozen matrix (Zhang et. al, pg. 2, sec. 1, “Specifically, LoRA-FA chooses to freeze both pretrained weight of W and projection-down weight of A, and only update projection-up weight of B, as shown in Figure 1(c). By doing so, we only need to compute the gradient of B, which requires storing a much smaller input of AX during the feed-forward pass.” Where the weight of the trainable structure B is made of vectors which can correspond to the trainable scaling vector, the frozen projection-down weight of A corresponds to the frozen matrix), a second trainable scaling vector, a second frozen matrix (Zhang et. al, pg. 1, Abstract, “the number of LoRA layers”, where the mention of multiple LoRA layers teach multiple iterations of the frozen and trainable vectors are present in the method of the art that can corresponding to the first and second data structures (e.g. vector, matrix) in the claim. Given the structure taught by Zhang for the frozen weight and trainable scaling vector, Zhang anticipates a second training scaling vector and second frozen matrix), and the received input (Zhang et. al, pg. 2, sec. 1, “the projection-down weight of A has mapped a d-dimensional input of X into an r-dimensional input of XA”, where X is the received input).
Regarding Claim 2, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated). Zhang further teaches to generate the output further based on addition, to the first product, of a second product based on a trained weight matrix and the received input (Zhang et. al, pg. 2, sec. 1, “During the inference, similar to LoRA, it can merge low-rank weights by adding AB into W”) where AB can correspond to the first product and W the second product. Zhang discloses equation (1) that demonstrates the output generation based on addition of the products.
Regarding Claim 3, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the first product corresponds to a low-rank decomposition of an accumulated weight update during adaptation of the neural network (Zhang et. al, pg. 5, sec. 3.3, “This implies that LoRA-FA is equivalent to a low-rank gradient compression method for full parameter fine-tuning, in which the calculated weight gradient is compressed … and then it is uncompressed”).
Regarding Claim 4, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the first frozen matrix and the second frozen matrix comprise matrices, the second frozen matrix, as identified in the rejection of Claim 1, comes from a second layer, shared across the layer of the neural network (Zhang et. al, pg. 2, Fig. 1, Where LoRA-FA will fix the projection-down weight A across its entirety) and one or more other layers of the neural network (Zhang et. al, pg. 2, sec. 1, “LoRA-FA does not change feed-forward and back-propagation computations of LoRA (except skipping the gradient calculation of A)”, where through multiple iterations of the model, the projection-down weight A does not have its weights updated, keeping its same weight throughout the one or more other layers).
Regarding Claim 5, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the first frozen matrix and the second frozen matrix comprise random matrices (Zhang et. al, pg. 2, sec. 1, “We initialize A randomly from a normal distribution (which is typically a rank-r matrix) … During the model adaptation, LoRA-FA will fix A and update B to improve the model performance”).
Regarding Claim 7, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the one or more processors are further configured to cause the processing system to generate at least one of the first trainable scaling vector or the second trainable scaling vector based on gradient descent and backpropagation of values from one or more other layers of the neural network (Zhang et. al, pg. 2, sec. 1, para. 5-6, “By doing so, we only need to compute the gradient of B, … LoRA-FA does not change feed-forward and back-propagation computations of LoRA (except skipping the gradient calculation of A)”).
Regarding Claim 8, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the neural network comprises a transformer neural network (Zhang et. al, pg. 5, sec. 3.2, “The activation memory overhead exists in other components of a transformer model, such as attention, layernorm, GeLU, and dropout (Korthikanti et al., 2023). To address it, we can use full activation recomputation to store the input of each transformer block.”).
Regarding Claim 9, Zhang teaches the processing system of Claim 1 (and thus the rejection of Claim 1 is incorporated), wherein the one or more processors are further configured to cause the processing system to take one or more actions based on the generated output of the layer of the neural network (Zhang et. al, pg. 2, sec. 1, “During the inference, similar to LoRA, it can merge low-rank weights by adding AB into W, introducing no extra inference latency compared to a fully fine-tuned model.” where in light of the specifications [0041], one or more actions can be interpreted as an inference based on the generated output of the layer).
Regarding Claim 10, Zhang teaches the processing system of Claim 9 (and thus the rejection of Claim 9 is incorporated), to generate an inference based on the generated output of the layer of the neural network (Zhang et. al, pg. 2, sec. 1, “During the model adaptation, … the change of model weight i.e. ∆W = AB is residing in a low-rank space, defined by the column space of the initialized A. Empirically, our experimental results show that LoRA-FA is sufficient to fine-tune LLMs. … During the inference, similar to LoRA, it can merge low-rank weights by adding AB into W.”, where the fine-tuning stage is where the generated output of the layer of the neural network for the change of model weight is produced, and the inference stage is where the generation of an inference occurs based on the fine-tuning output in the merge of the change of weights with the pre-trained weight).
Regarding Claim 11, Zhang teaches the processing system of Claim 9 (and thus the rejection of Claim 9 is incorporated), to forward propagate the output of the layer of the neural network to one or more further layers of the neural network; and trigger generation of a first scaling vector and a second scaling vector for the one or more further layers of the neural network (Zhang et. al, pg. 2, sec. 1, “compute the gradient of B, which requires storing a much smaller input of AX during the feed-forward pass. … LoRA-FA does not change feed-forward and back-propagation computations of LoRA (except skipping the gradient calculation of A), thus, it will not increase the computational overhead during the fine-tuning phase.” Where the feed-forward and back-propagation encompasses the functionality of forward propagation that is used to trigger updating the trainable weight structure for further layer iterations. Gradient B corresponds to the trainable scaling weights, where B is a multitude of scaling vectors).
Claims 12-15,17-19 recite a processor-implemented method to perform precisely the system of claims 1,2,4,5,7,10,11 respectively. As Zhang et. al performs the method with their system on a memory and processor inherently due to GPUs used in experimentation, e.g. “we use NVIDIA Turing RTX2080Ti for small-sized RoBERTa models” (pg. 6, sec. 4.1), claims 12-15,17-19 are rejected for reasons set forth in the rejections of claims 1,2,4,5,7,10,11 respectively.
Claim 20 recites a means for receiving an input for processing through a layer of a neural network; and means for generating to perform precisely the system of claim 1. As the rejection of claim 1 demonstrates a system (e.g. means) for performing these functions, claim 20 is thus rejected for reasons set forth in the rejection of Claim 1.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 6 is/are rejected under 35 U.S.C. 103 as being unpatentable over Zhang et. al, “LORA-FA: Memory-Efficient Low-Rank Adaption for Large Language Models Fine-Tuning”, in view of Wang et. al, “Transformed Low-Rank Parameterization Can Help Robust Generalization for Tensor Neural Networks”.
Regarding Claim 6, Zhang recites in claim 1 the teachings to calculate the first product (and thus the rejection of claim 1 is incorporated).
Zhang does not teach, but Wang teaches to transform the first trainable scaling vector into a first diagonal matrix; transform the second trainable scaling vector into a second diagonal matrix (Wang et. al, pg. 3, sec. 2.1, “As the M-block-diagonal matrix TM is defined after transforming tensor T from the original domain to the transformed domain, the relationship … indicates that the tubal rank can be chosen as a measure of transformed low-rankness” where the original tensor, which can be defined as either a matrix or vector, can map to the vector that is transformed to the diagonal matrix TM). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Zhang, to offer scaling capabilities from the diagonal matrix for performing the computations of the neural network. The motivation to do so being the “additional expressivity to neural networks, while the controllable transformed low-rank structure in t-NNs enables a flexible balance between model accuracy and robustness”. (Wang et. al, pg. 1, sec. 1)
Regarding Claim 16, the Zhang/Wang combination of Claim 6 teaches generating the output of the layer of the neural network (and thus the rejection of claim 6 is incorporated). Wang does not teach, but Zhang further teaches, a processor-implemented method. Claim 16 recites precisely the method performed by the system of Claim 6. As Zhang/Wang teaches the system of Claim 6, which performs the method of Claim 16, Claim 16 is thus rejected for reasons set forth in the rejection of Claim 6.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HAYDEN Y ARDALE whose telephone number is (571)270-5128. The examiner can normally be reached Monday-Thursday (7:30-5).
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, Kakali Chaki can be reached at (571) 272-3719. 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.
/H.Y.A./Examiner, Art Unit 2122
/KAKALI CHAKI/Supervisory Patent Examiner, Art Unit 2122
