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 .
Claim Objections
Claim 1 is objected to because of the following informalities: Claim 1 line 6 has the term “based model” and it should be “base model”. Appropriate correction is required.
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.
Claims 1-5 are rejected under 35 U.S.C. 103 as being unpatentable over Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna) in view of Tan et al. (US 2020/0251100 A1 – hereinafter Tan) and further in view of Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer).
In regards to claim 1, Bapna disclose a method comprising:
obtaining a base model associated with one or more machine learning models; (Bapna abstract cites “Our proposed approach consists of injecting tiny task specific adapter layers into a pre-trained model.” And section 3 first paragraph teaches “Our approach consists of two phases: (i) Training a generic base model, and (ii) adapting it to new tasks with added small network modules.”)
determining a first domain specific part associated with the one or more machine learning models, the first domain specific part corresponding to the first domain; (Bapna section 4.2 second bullet teaches “Adaptation: Inject a set of domain-specific adapter layers for every target domain. These adapters are then fine-tuned to maximize performance on the corresponding domains. This step can be applied any time a new domain is added to the model.” Also see figure 1 wherein the input is the Chinese input, and the selected adapter is Chinese to English in the encoder.)
causing the first domain specific part to be coupled to the based model (Bapna section 3 first paragraph and section 4.2 second bullet teaches causing the first domain specific part to couple to based model.)
determining, based at least on the base model and the first domain specific part processing input data, an output associated with the input data. (Bapna figure 1 shows a transformer encoder/decoder enhanced with adapters, receiving a Chinese input and producing English output. The caption for figure 1 says the same formulation can be used for domain adaptation. Further section 4.2 states domain-specific adapters are fine-tuned to maximize performance on the corresponding domains.)
However, Bapna does not explicitly disclose determining a first domain based on input data; and causing the first domain specific part to be coupled to the base model and a second domain specific part be decoupled from the base model, the second domain specific part corresponding to second domain.
Tan disclose determining a first domain based on input data. (Tan para. [0036] teaches determine a domain an input data wherein it cites “The processor 102 is further configured to provide the input text 106 to the domain identifier 122. The domain identifier 122 is configured to generate, based on the input text 106, a weight vector 124. The weight vector 124 includes a classification weight for each domain of the plurality of domains, and the classification weight for a particular domain is associated with a probability that the input text 106 is associated with the particular domain.” This teaches determining a first domain of input data.)
It would have been obvious to one of ordinary skill in the art before earliest effective filing date of the claimed invention to modify the teachings of the Bapna with that of Tan in order to allow for determining the domain of input data as both references deal with multi-domain classification. The benefit of doing so it makes Bapna more efficient by determining if the data also matches domains in the system or whether a new one needs to be added.
However, Bapna in view of Tan does not explicitly causing the first domain specific part to be coupled to the base model and a second domain specific part be decoupled from the base model, the second domain specific part corresponding to second domain.
Shazeer discloses causing the first domain specific part to be coupled to the based model and a second domain specific part to be decoupled from the base model, the second domain specific part corresponding to a second domain. (Shazeer figure 1 teaches a base neural network comprising the first neural network layer (104) and the second neural network layer (108). Shazeer abstract, para. [0004] and figure 1 element 124, 110 and 112, and fig. 2 element 202 and 204 teaches based on the input to the gating function, selecting appropriate expert(s). Examiner interprets each expert to be a different domain. Shazeer para. [0027-0033] teaches the gating system selects a small number of experts based on the given input, combines the output the selected experts, and provides the output of the mixture of experts (MoE) to the next layer. Shazeer para. [0036] teaches conditional computation, wherein parts of the MoE subnetwork are active on a per-input basis, meaning only experts selected based on the input are active. This also means that the remaining experts are inactive which the examiner interprets to be decoupled.)
It would have been obvious to one or ordinary skill in the art before earliest effective filing date of the claimed invention to modify the teachings of the Bapna in view of Tan with that of Shazeer to make the second domain specific parts (layers, modules, subnetwork or experts) inactive (decoupled) as both Bapna and Shazeer references deal with machine learning models being adapted based on data. The benefit of doing so it creates a more efficient system by increasing the computation capacity as cited in para. [0036] of the Shazeer by only using experts relevant to the input.
In regards to claim 2, Bapna in view of Tan in view of Shazeer disclose the method of claim 1, further comprising: determining that the second domain specific part corresponds to the second domain, wherein the causing the second domain specific part to be decoupled from the base model is based at least on the second domain specific part corresponding to the second domain. (Shazeer para. [0032] a sparsity output of a weight vector, wherein any expert with 0 does not need to be computed, which means it is inactive and not relevant to input the data. Thus, it is not part of the first domain and is a second domain.)
In regards to claim 3, Bapna in view Tan in view of the Shazeer discloses the method of claim 1,further comprising: wherein the determining the output associated with the input data occurs while the first domain specific part is coupled to the base model and the second domain specific part is decoupled from the base model. (See Shazeer figure 1 wherein the output is based on the experts 116 being active/coupled and experts 114, 118, and 122 being inactive/decoupled.)
In regards to claim 4, Bapna in view of Tan in view of Shazeer discloses the method of claim 3, wherein the determining that the input data is associated with the specific domain comprises 1, further comprising at least one of: receiving, from a user device, an indication that the input data is associated with the specific first domain; or analyzing the input data to determine that the input data is associated with the specific first domain. (Tan para. [0036] teaches analyzing input data to determine a domain of the input data wherein it cites “The processor 102 is further configured to provide the input text 106 to the domain identifier 122. The domain identifier 122 is configured to generate, based on the input text 106, a weight vector 124. The weight vector 124 includes a classification weight for each domain of the plurality of domains, and the classification weight for a particular domain is associated with a probability that the input text 106 is associated with the particular domain.” This teaches determining a first domain of input data.)
In regards to claim 5, Bapna in view of Tan in view of Shazeer discloses the method of claim 1, wherein the causing the first domain specific part to be coupled to the base model and the second domain specific part to be decoupled from the base model comprises: causing one or more first layers associated with the first domain specific part to be coupled to one or more second layers associated with the base model; and causing one or more third layers associated with the second domain specific part to be decoupled from the one or more second layers associated with the base model. (See Shazeer figure 1 wherein element 116 and element 120 are first domain and coupled to the second layer of the base model and elements 114, 118, and 122 are second domain and third layers decoupled from the second layer of the neural network (base model).)
Claims 6-7 are rejected under 35 U.S.C. 103 as being unpatentable over Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna) in view of Tan et al. (US 2020/0251100 A1 – hereinafter Tan) in view of Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer) and further in view of Houlsby et al. (“Parameter-Efficient Transfer Learning for NLP” – hereinafter Houlsby).
In regards to claim 6, Bapna in view of Tan in view of Shazeer discloses the method of claim 1, but does not explicitly disclose updating, using first training data associated with one or more general domains, one or more first parameters of one or more first layers associated with the base model; and updating, using second training data associated with the first domain, one or more second parameters of one or more second layers associated with the first domain specific part.
Houlsby discloses updating, using first training data associated with one or more general domains, one or more first parameters of one or more first layers associated with the base model; (Houlsby abstract, section 1 first paragraph and section 3.1 teaches using a pre-trained BERT network as the based model and the Bert network is pre- trained on a large text corpora (general domain). This teaches updating the parameters of the base model as training is what determines those parameters.) and updating, using second training data associated with a specific domain, one or more second parameters of one or more second layers associated with the domain specific part. (Houlsby abstract cites "Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones." And page 2 section 2 teaches in adapter-tuning, the parameters of the original network are frozen...". Both of these teaches when the adapter layer is being trained that based model layers are frozen. Also fig. 2 text teaches adapters are trained on downstream data.)
It would have been obvious to one of ordinary skill in the art before the earliest effective filing date of the claimed invention to modify the teachings of Bapna in view of Tan in view of Shazeer with the teachings of Houlsby in order to allow update base models using general domain information as Bapna, Shazeer and Houlsby deal with adapting machine learning models to domains and the benefit of doing so it creates more accurate models by updating the models when needed.
In regards to claim 7, Bapna in view of Tan in view of Shazeer in view of Houlsby discloses the method of claim 6, further comprising: during the updating using the first training data, refraining from updating the one or more second parameters of the one or more second layers associated with the domain specific part; and during the updating using the second training data, refraining from updating the one or more first parameters of the one or more first layers associated with the base model. (Houlsby abstract, section 1 first paragraph and section 3.1 teaches using a pre-trained BERT network as the based model and the Bert network is pre-trained on a large text corpora (general domain), so the parameters of the base model are tuned without updating he second parameters for the second layers (adapter layer) as it hasn't been added yet. Houlsby abstract cites "Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones." and page 2 section 2 teaches in adapter-tuning, the parameters of the original network are frozen...". Both of these teaches when the adapter layer is being trained that based model layers are frozen. Also fig. 2 text teaches adapters are trained on downstream data.)
Claims 8-9 are rejected under 35 U.S.C. 103 as being unpatentable over Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna) in view of Tan et al. (US 2020/0251100 A1 – hereinafter Tan) in view of Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer) and further in view of Rebuffi et al. ("Learning Multiple Visual Domains with Residual Adapters" - hereinafter Rebuffi).
In regards to claim 8, Bapna in view of Tan in view of Shazeer discloses the method of claim 1, but does not disclose further comprising: determining, using one or more second machine learning models, a second output associated with the input data; and determining, based at least on the output and the second output, a third output associated with the input data.
Rebuffi discloses determining, using one or more second machine learning models, a second output associated with the input data; and determining, based at least on the output and the second output, a third output associated with the input data. (Rebuffi page 2 second paragraph states "The layers in the resulting parametric network are either domain-agnostic, hence shared between domains, or domain-specific, hence parametric. The domain-specific layers are changed based on the ground truth domain of the input image, or based on an estimate of the latter obtained from an auxiliary network. In the latter configuration, our architecture is analogous to the learnet of [2]." This teaches a parametric network (1st model) and an auxiliary network (second model). Then section 3 paragraph 2 teaches "classification layer of 30 standard ResNet model. The parametric feature extractors the is then used to construct predictors for each domain d as 4, Our where or are domain-specific parameters and SoftMax( is a domain-specific linear classifier V Vs mapping features to wherein this teaches parametric network (фаd), a classifier (φd) and result. The auxiliary network is disclosing in section 3.1 wherein it cites "As shown later, it often easy to construct an auxiliary network that predict d from x." this teaches the second machine learning model taking input X and getting second output, which is the domain. Then page 2 second paragraph states "The domain-specific layers are changed based on the ground-truth domain of the input image, or based on an estimate of the latter obtained from an auxiliary network. In the latter configuration, our architecture is analogous to the learnet of [2]." This teaches that domain parametric model is changed based on the output (domain) of the auxiliary network, thus the third output is based on the first and second output of the first and second models.)
It would have been obvious to one of ordinary in the art before the effective filing date of the claimed invention to modify the teachings of Bapna in view of Tan in view of Shazeer with that of Rebuffi in order to allow combining the output of two models to determine an output as Shazeer and Rebuffi both deal with fine-tuning a based model using domain specific data and the benefit of doing so it allow for more accurate model selection as the auxiliary model in Rebuffi allows for finding the domain of the data and then tuning the primary or based model using that data.
In regards to claim 9, Houlsby in view of Rebuffi discloses the method of claim 8, wherein the determining the third output associated with the input data comprises:
determining, based at least on the output and the second output, the third output associated with the input data and a fourth output associated with the input data; determining, using one or more third machine learning models, a first score associated with the third output and a second score associated with the fourth output; and determining the third output based at least on the first score being greater than the second score. (See Rebuffi page 6 section "challenge and evaluation" for scoring output and page 8 of Rebuffi.)
Claims 10-12, 14, and 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer) in view of Tan et al. (US 2020/0251100 A1 – hereinafter Tan) and further in view of Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna).
In regards to claim 10, Shazeer discloses a system comprising:
one or more processors to: (Shazeer para. [0057] teaches a processor)
cause, based at least on the first domain, the one or more first layers of the one or more machine learning models to be activated and one or more second layers of the one or more machine learning models to be deactivated, the one or more second layers being associated with a second domain; (Shazeer figure 1 teaches a base neural network comprising the first neural network layer (104) and the second neural network layer (108). Shazeer abstract, para. [0004] and figure 1 element 124, 110 and 112, and fig. 2 element 202 and 204 teaches based on the input to the gating function, selecting appropriate expert(s). Examiner interprets each expert to be a different domain. Shazeer para. [0027-0033] teaches the gating system selects a small number of experts based on the given input (fig. 1 elements 116 and 120 are selected and active, while figure 1 elements 114, 118 and 122 are not selected and deactivated), combines the output the selected experts, and provides the output of the mixture of experts (MoE) to the next layer. Shazeer para. [0036] teaches conditional computation, wherein parts of the MoE subnetwork are active on a per-input basis, meaning only experts selected based on the input are active. This also means that the remaining experts are inactive which the examiner interprets to be decoupled. The activated layers are considered the first layers and deactivated layers are considered the second layers)
processing the input data using one or more third layers associated with a base model of the one or more machine learning models and the one or more first layers of the one or more machine learning models to generate an output; and (Shazeer figure 1 teaches the third layer (second neural network layer 108) associated with a base model and processing the input (first layer output 124) using the first layers (elements 116 and 120) and third layers (second neural network layer 108).)
perform one or more operations using the output.(Shazeer para. [0034] teaches performing further action and processing using the output.)
However, Shazeer does not explicitly disclose receiving input data associated with a first domain, the first domain associated with one or more first layers of one or more machine learning models.
Tan disclose receiving input data associated with a first domain. (Tan para. [0036] teaches determine a domain received input data wherein it cites “The processor 102 is further configured to provide the input text 106 to the domain identifier 122. The domain identifier 122 is configured to generate, based on the input text 106, a weight vector 124. The weight vector 124 includes a classification weight for each domain of the plurality of domains, and the classification weight for a particular domain is associated with a probability that the input text 106 is associated with the particular domain.” This teaches determining a first domain of input data.)
It would have been obvious to one of ordinary skill in the art before earliest effective filing date of the claimed invention to modify the teachings of the Shazeer with that of Tan in order to allow for determining the domain of input data as both references deal with multi-domain classification. The benefit of doing so it makes Shazeer more efficient by determining if the data is associated with any of the experts in the system.
However, Shazeer in view of Tan does not explicitly disclose the first domain associated with one more first layers of the one or more machine learning models.
Bapna discloses the first domain associated with one or more layers of the one or more machine learning models. (Bapna section 4.2 second bullet teaches “Adaptation: Inject a set of domain-specific adapter layers for every target domain. These adapters are then fine-tuned to maximize performance on the corresponding domains. This step can be applied any time a new domain is added to the model.” Also see figure 1 wherein the input is the Chinese input, and the selected adapter is Chinese to English in the encoder. This teaches a domain associated with one or more layers of the machine learning models.)
It would have been obvious to one of ordinary skill in the art before the earliest effective filing date of the claimed invention to modify the teachings of Shazeer in view of Tan with that of Bapna in order allow for associating layers domains as Shazeer using different experts depending on input data and Bapna associated layers with domains which are determined from input data, thus creating a more efficient system that is able to route or select experts based on domain on input data to the correct or more appropriate expert for processing.
In regards to claim 11, Shazeer in view of Tan in view of Bapna disclose the system of claim 10 wherein the one or more first layers are activated by at least coupling the one or more first layers to the one or more third layers; and the one or more second layers are deactivated by at least decoupling the one or more second layers from the one or more third layers. (See Shazeer figure 1 wherein active layer are element 116 and 120 and couples second neural network element 108 (third layer) and the second layers (elements 114, 118, and 122) are decoupled (inactive) from the third layer.)
In regards to claim 12, Shazeer in view of Tan in view of Bapna discloses the system of claim 10, wherein the one or more processors are further configured to determine to active the one or more first layers and deactivate the one or more second layers based at least on one or more of: receiving, from a user device, an indication to at least one of activating the one or more first layers or deactivate the one or more second layers; or analyzing the input data to determine that the input data is associated with the first domain. (Tan para. [0036] teaches analyzing input data to determine a domain of the input data wherein it cites “The processor 102 is further configured to provide the input text 106 to the domain identifier 122. The domain identifier 122 is configured to generate, based on the input text 106, a weight vector 124. The weight vector 124 includes a classification weight for each domain of the plurality of domains, and the classification weight for a particular domain is associated with a probability that the input text 106 is associated with the particular domain.” This teaches determining a first domain of input data.)
In regards to claim 14, Shazeer in view of Tan in view of Bapna discloses the system of claim 10, wherein the input data is input into the one or more machine learning models at a first time, and wherein the one or more processors are further to: receive second input data associated with the second domain; input the second input data into the one or more machine learning models at a second time, the one or more machine learning models including the one or more third layers associated with the base model and the one or more second layers associated with the second domain at the second time; and determine, using the one or more machine learning models and based at least one the second input data, a second output associated with the second input data. (Shazeer in view of Tan in view of Bapna does not explicitly disclose receiving a second input at a second time associated with a second domain, however it would have been obvious to do as discloses in claim 10 Shazeer teaches receiving input and based on in the input it is routed via the gating function the appropriate expert. It would have been obvious to receive a second input and based on it routing it the appropriate expert at the second time. Tan would also be used to determine the domain of the second input.)
In regards to claim 17, Shazeer in view of Tan in view of Bapna discloses the system of claim 10, wherein the system is comprised in at least one of: an infotainment system for an autonomous or semi-autonomous machine; an entertainment system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for hosting real-time streaming applications; a system for generating content for one or more of virtual reality (VR), augmented reality (AR), or mixed reality (MR); a system for performing digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing deep learning operations;(Tan para. [0013-0016] teaches the system using deep learning) a system implemented using an edge device; (Shazeer para. [0066] teaches a client-server setup wherein client devices are edge devices.) a system implemented using a robot; a system for performing conversational AI operations; a system for generating synthetic data; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources. (Tan para. [0008] teaches system implemented in a cloud computing environment)
In regard to claim 18, Shazeer discloses one or more processors comprising processing circuitry to: (Shazeer para. [0057] teaches a processor)
receive, from one or more computing devices, a base model of one or more machine learning models and a first part of the one or more machine learning models that is associated with the first domain while refraining from receiving a second part of the one or more machine learning models that is associated with a second domain; (Shazeer figure 1 teaches a base neural network comprising the first neural network layer (104) and the second neural network layer (108). Shazeer abstract, para. [0004] and figure 1 element 124, 110 and 112, and fig. 2 element 202 and 204 teaches based on the input to the gating function, selecting appropriate expert(s). Examiner interprets each expert to be a different domain as it’s not used for the input, if they all were for the same domain all the experts would be used. Shazeer para. [0027-0033] teaches the gating system selects a small number of experts based on the given input (fig. 1 elements 116 and 120 are selected and active/received, while figure 1 elements 114, 118 and 122 are not selected and deactivated which means they are not received), combines the output the selected experts, and provides the output of the mixture of experts (MoE) to the next layer. Shazeer para. [0036] teaches conditional computation, wherein parts of the MoE subnetwork are active on a per-input basis, meaning only experts selected based on the input are active/received. This also means that the remaining experts are inactive/no received, which meaning they not coupled.)
couple the base model of the one or more machine learning models with the first part of the one or more machine learning models; (Shazeer fig. 1 elements 116 and 120 show the experts coupled to the based model) and generate, based at least on processing the input data using the base model and the first part of the one or more machine learning models, an output associated with the input data. (Shazeer para. [0034] teaches performing further action and processing using the output.)
However, Shazeer does not explicitly disclose receiving input data associated with a first domain, the first domain associated with one or more first parts of one or more machine learning models.
Tan disclose receiving input data associated with a first domain. (Tan para. [0036] teaches determine a domain received input data wherein it cites “The processor 102 is further configured to provide the input text 106 to the domain identifier 122. The domain identifier 122 is configured to generate, based on the input text 106, a weight vector 124. The weight vector 124 includes a classification weight for each domain of the plurality of domains, and the classification weight for a particular domain is associated with a probability that the input text 106 is associated with the particular domain.” This teaches determining a first domain of input data.)
It would have been obvious to one of ordinary skill in the art before earliest effective filing date of the claimed invention to modify the teachings of the Shazeer with that of Tan in order to allow for determining the domain of input data as both references deal with multi-domain classification. The benefit of doing so it makes Shazeer more efficient by determining if the data is associated with any of the experts in the system.
However, Shazeer in view of Tan does not explicitly disclose the first domain associated with one more first part of the one or more machine learning models.
Bapna discloses the first domain associated with one or more parts of the one or more machine learning models and second domain associated with a second part . (Bapna section 4.2 second bullet teaches “Adaptation: Inject a set of domain-specific adapter layers for every target domain. These adapters are then fine-tuned to maximize performance on the corresponding domains. This step can be applied any time a new domain is added to the model.” Also see figure 1 wherein the input is the Chinese input, and the selected adapter is Chinese to English in the encoder. This teaches a domain associated with one or more layers/parts and it has many parts or domains of the machine learning models.)
It would have been obvious to one of ordinary skill in the art before the earliest effective filing date of the claimed invention to modify the teachings of Shazeer in view of Tan with that of Bapna in order allow for associating layers domains as Shazeer using different experts depending on input data and Bapna associated layers with domains which are determined from input data, thus creating a more efficient system that is able to route or select experts based on domain on input data to the correct or more appropriate expert for processing.
In regards to claim 19, Shazeer in view of Tan in view of Bapna discloses the one or more processors of claim 18, wherein the one or more processing units are further to: activate the one or more first layers and deactivate the one or more second layers based at least on one or more of: refraining from the receiving the second part of the one or more machine learning models is based at least on the input data being associated with the first domain. (Shazeer abstract, para. [0004] and figure 1 element 124, 110 and 112, and fig. 2 element 202 and 204 teaches based on the input to the gating function, selecting appropriate expert(s). Examiner interprets each expert to be a different domain as it’s not used for the input, if they all were for the same domain all the experts would be used. Shazeer para. [0027-0033] teaches the gating system selects a small number of experts based on the given input (fig. 1 elements 116 and 120 are selected and active/received, while figure 1 elements 114, 118 and 122 are not selected and deactivated which means they are not received), combines the output the selected experts, and provides the output of the mixture of experts (MoE) to the next layer. Shazeer para. [0036] teaches conditional computation, wherein parts of the MoE subnetwork are active on a per-input basis, meaning only experts selected based on the input are active/received. This also means that the remaining experts are inactive/no received, which meaning they not coupled.)
In regards to claim 20, Shazeer in view of Tan in view of Bapna discloses the one or more processors of claim 18 wherein the system is comprised in at least one of: an infotainment system for an autonomous or semi-autonomous machine; an entertainment system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for hosting real-time streaming applications; a system for generating content for one or more of virtual reality (VR), augmented reality (AR), or mixed reality (MR); a system for performing digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing deep learning operations;(Tan para. [0013-0016] teaches the system using deep learning) a system implemented using an edge device; (Shazeer para. [0066] teaches a client-server setup wherein client devices are edge devices.) a system implemented using a robot; a system for performing conversational AI operations; a system for generating synthetic data; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources. (Tan para. [0008] teaches system implemented in a cloud computing environment)
Claims 13 and 15-16 are rejected under 35 U.S.C. 103 as being unpatentable over Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer) in view of Tan et al. (US 2020/0251100 A1 – hereinafter Tan) in view of Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna) and further in view of Houlsby et al. (“Parameter-Efficient Transfer Learning for NLP” – hereinafter Houlsby).
In regards to claim 13, Shazeer in view of Tan in view of Bapna discloses the system of claim 10, but does not explicitly disclose wherein the one or more first layers are caused to activated and the one or more second layers are caused to be deactivated based at least one: a first memory component associate with the one or more first layers being connected to one or more third layers associated with the based model; and a second memory component associated with one or more third layers associated with the based model.
Houlsby discloses wherein the one or more first layers are caused to activated and the one or more second layers are caused to be deactivated based at least one: a first memory component associate with the one or more first layers being connected to one or more third layers associated with the based model; and a second memory component associated with one or more third layers associated with the based model. (Houlsby abstract cites "...we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task,..."; page 2 left column second paragraph cites "Adapters are new modules added between layers of a pre-trained network."; page 2 section 2 second paragraph cites "in adapter-tuning, the parameters of the original network are frozen and therefore may be shared by many tasks."; and page 2 left column third paragraph teaches adapters differ in that tasks do not interact and the shared parameters are froze, meaning the model has perfect memory of previous task...", this means that each task has its own adapter and when one task is active the others are not, and thus the other adapters and base model is frozen, meaning any memory associated with them are also not active. )
It would have been obvious to one of ordinary skill in the art before earliest effective filing date of the claimed invention to modify the teachings of the Shazeer in view of Tan in view of Bapna with that of Houlsby in order to activate and deactivate memory associated with layers as both Shazeer and Houlsby deal with adapting models base input data and the it provides the benefit of create a more robust system.
In regards to claim 15, Bapna in view of Tan in view of Shazeer discloses the system of claim 10, but does not disclose further comprising: determine, using one or more second models, a second output associated with the input data; and determine, based at least on the output and the second output, a third output associated with the input data.
Rebuffi discloses determining, using one or more second machine learning models, a second output associated with the input data; and determining, based at least on the output and the second output, a third output associated with the input data. (Rebuffi page 2 second paragraph states "The layers in the resulting parametric network are either domain-agnostic, hence shared between domains, or domain-specific, hence parametric. The domain-specific layers are changed based on the ground truth domain of the input image, or based on an estimate of the latter obtained from an auxiliary network. In the latter configuration, our architecture is analogous to the learnet of [2]." This teaches a parametric network (1st model) and an auxiliary network (second model). Then section 3 paragraph 2 teaches "classification layer of 30 standard ResNet model. The parametric feature extractors the is then used to construct predictors for each domain d as 4, Our where or are domain-specific parameters and SoftMax( is a domain-specific linear classifier V Vs mapping features to wherein this teaches parametric network (фаd), a classifier (φd) and result. The auxiliary network is disclosing in section 3.1 wherein it cites "As shown later, it often easy to construct an auxiliary network that predict d from x." this teaches the second machine learning model taking input X and getting second output, which is the domain. Then page 2 second paragraph states "The domain-specific layers are changed based on the ground-truth domain of the input image, or based on an estimate of the latter obtained from an auxiliary network. In the latter configuration, our architecture is analogous to the learnet of [2]." This teaches that domain parametric model is changed based on the output (domain) of the auxiliary network, thus the third output is based on the first and second output of the first and second models.)
It would have been obvious to one of ordinary in the art before the effective filing date of the claimed invention to modify the teachings of Bapna in view of Tan in view of Shazeer with that of Rebuffi in order to allow combining the output of two models to determine an output as Shazeer and Rebuffi both deal with fine-tuning a based model using domain specific data and the benefit of doing so it allow for more accurate model selection as the auxiliary model in Rebuffi allows for finding the domain of the data and then tuning the primary or based model using that data.
In regards to claim 16, Shazeer in view of Tan in view Bapna discloses the system of claim 10, but does not explicitly disclose wherein the one or more processors are further configured to: updating, using first training data associated with one or more general domains, one or more first parameters of one or more third layers without updating one or more second parameters associated with the one or more first layers; and update, using second training data associated with the first domain, the one or more second parameters associated with the one or more first layers without updating the one or more first parameters associated with the one more third layers.
Houlsby discloses updating, using first training data associated with one or more general domains, one or more first parameters of one or more third layers without updating one or more second parameters associated with the one or more first layers; and update, using second training data associated with the first domain, the one or more second parameters associated with the one or more first layers without updating the one or more first parameters associated with the one more third layers. (Houlsby abstract, section 1 first paragraph and section 3.1 teaches using a pre-trained BERT network as the based model and the Bert network is pre- trained on a large text corpora (general domain). This teaches updating the parameters of the base model as training is what determines those parameters. Houlsby abstract cites "Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones.", and Houlsby page 2 section 2 teaches in adapter-tuning, the parameters of the original network are frozen...". Both of these teaches when the adapter layer is being trained that and base model layers are frozen. Also fig. 2 text teaches adapters are trained on downstream data.)
It would have been obvious to one of ordinary skill in the art before the earliest effective filing date of the claimed invention to modify the teachings of Bapna in view of Tan in view of Shazeer with the teachings of Houlsby in order to allow update base models using general data while not updating domain specific layer and updating domain specific layers while freezing generic model layers as Bapna, Shazeer and Houlsby deal with adapting machine learning models to domains and the benefit of doing so it allow models to be tuned to new domains while avoiding catastrophic forgetting by freeing the base layers.
Response to Arguments
Applicant’s arguments, filed 02 April 2026, with respect to the claims rejected under 35 USC 101 have been fully considered and are persuasive, therefore the rejection has been withdrawn. Applicant’s arguments with respect to the rejection of the claims under 35 USC 102 and 103 having being overcome based on the amendments to the claims have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Bapna et al. (“Simple, Scalable Adaptation for Neural Machine Translation” – hereinafter Bapna), Tan et al. (US 2020/0251100 A1 – hereinafter Tan), Shazeer et al. (US 2019/0251423 A1 – hereinafter Shazeer) and Houlsby et al. (“Parameter-Efficient Transfer Learning for NLP” – hereinafter Houlsby), see the rejection above for further details.
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 PAULINHO E SMITH whose telephone number is (571)270-1358. The examiner can normally be reached Mon-Fri. 10AM-6PM CST.
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, Abdullah Kawsar can be reached at 571-270-3169. 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.
/PAULINHO E SMITH/Primary Examiner, Art Unit 2127
Prosecution Insights