DETAILED ACTION This action is responsive to the Application f iled on 12/01/2023. Claims 1-24 are pending in the case. Claims 1 and 1 3 are independent claims. Domestic Benefit Domestic Benefit of provisional dated 12/02/2022 is acknowledged 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. 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. Claim Rejections - 35 USC § 102 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)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale , or otherwise available to the public before the effective filing date of the claimed invention. (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 , 3-5 , 12-13 , 15-17 and 24 is/are rejected under 35 U.S.C. 102 (a) as being anticipated by Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) . Regarding claim 1 : Houlsby discloses a computer-implemented method for training a modular neural network model, the computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations ( Houlsby , Page 3, Col. 2, Paragraph 4,“ All runs are trained on 4 Google Cloud TPUs with a batch size of 32 ” where running the model on a TPU corresponds to using a computer-implemented method for training a modular neural network mode and being executed on data processing hardware that causes the data processing hardware to perform operations ) Houlsby discloses during an initial training stage, training only a backbone model to provide a first model configuration of the modular neural network model, the first model configuration comprising only the trained backbone model and adding an intrinsic sub-model to the trained backbone model ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Adapters are new modules added between layers of a pre-trained network. ” where pre-trained network corresponds to a first model configuration where the backbone is trained without an adapter module is not added to the model yet and training only a backbone model and the adaptor module corresponds to an intrinsic sub-model being added to a backbone model ) Houlsby discloses during a fine-tuning training stage: freezing parameters of the trained backbone model ( Houlsby , Page 2 , Col. 2, Paragraph 2 , “The weights of the original network are untouched…the parameters of the original network are frozen and therefore may be shared by many tasks.” ) and fine-tuning parameters of the intrinsic sub-model added to the trained backbone model while the parameters of the trained backbone model are frozen to provide a second model configuration, the second model configuration comprising the backbone model initially trained during the initial training stage and the intrinsic sub- model having the parameters fine-tuned during the fine-tuning stag e ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Only the new, task specific, parameters, v, are then trained…During training, only v are tuned” where only training the adapters task specific parameters v corresponds creating a second model configuration where the fine-tuning parameters of the intrinsic sub model while the pre-trained model parameters are frozen ) Regarding claim 3 : The rejection of claim 1 with prior art Houlsby is incorporated and further: Houlsby discloses wherein the operations further comprise, after fine-tuning parameters of the intrinsic sub-model: removing the intrinsic sub-model ( Houlsby , Page 2 , Col. 2, Paragraph 3, “The adapter modules may also be ignored if not required” where each adapter for each task may be ignored if not required corresponds to removing the intrinsic sub-model as it’s no longer being considered in the network ( See also Houlsby , Page 6 , Col. 2, Paragraph 2, “ For this, we remove some trained adapters and re-evaluate the model (without re-training) ” where the removal of the adapters after training can also be considered removing an intrinsic sub-model after fine-tuning parameters ) Houlsby discloses adding another intrinsic sub-model ( Houlsby , Page 1 , Col. 1, Abstract, “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” where adding adapter modules for new tasks corresponds with disclosing adding another intrinsic sub-model as for each new task there is a new adapter ) to the trained backbone model ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Adapters are new modules added between layers of a pre-trained network. ” ) Houlsby discloses during another fine-tuning training stage : freezing parameters of the trained backbone model ( Houlsby , Page 2 , Col. 2, Paragraph 2 , “The weights of the original network are untouched…the parameters of the original network are frozen and therefore may be shared by many tasks.” ) and fine-tuning parameters of the other intrinsic sub-model added to the trained backbone model while the parameters of the trained backbone model are frozen to provide a third model configuration, the third model configuration comprising the backbone model initially trained during the initial training stage and the other intrinsic sub-model having the parameters fine-tuned during the other fine-tuning stage. ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Only the new, task specific, parameters, v, are then trained…During training, only v are tuned” where only training the adapters task specific parameters v corresponds creating a third model configuration where the fine-tuning parameters of the intrinsic sub model while the pre-trained model parameters are frozen ( See also Houlsby , Page 2 , Col. 1, Paragraph 3 , “…Adapters differ in that the tasks do not interact and the shared parameters are frozen” that further shows that each task has a corresponding adapter ) Regarding claim 4 : The rejection of claim 3 with prior art Houlsby is incorporated and further: Houlsby discloses during the fine-tuning training stage, the parameters of the intrinsic sub-model are trained on a first domain and/or first application ( Houlsby , Page 1 , Col. 1, Abstract , “ 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 ” where each adapter is tied to a separate task which is considered a domain and having a task trained on a task corresponds to the intrinsic-sub-model being trained on a first domain (See also, Houlsby , Page 2 , Col. 1, Paragraph 2 , “… Only the new, task specific, parameters, v, are then trained. ”)) Houlsby discloses during the other fine-tuning training stage, the parameters of the other intrinsic sub-model are trained on a second domain different than the first domain and/or a second application different than the first application ( Houlsby , Page 1 , Col. 1, Abstract , “ 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 … To demonstrate adapter’s effectiveness, we transfer the recently proposed BERT Transformer model to 26 diverse text classification tasks” where each task using an adapter and 26 tasks being defined as diverse with Table 1 and Table 2 showing the task s being different corresponds to having other intrinsic sub-model are trained on a second domain different than the first domain ) Regarding claim 5 : The rejection of claim 4 with prior art Houlsby is incorporated and further: Houlsby discloses wherein the trained backbone model is domain-independent ( Houlsby , Page 3 , Col. 2, Paragraph 3 , “ We use the public, pre-trained BERT Transformer network as our base model. ” where the pre-trained network used with adapters being public pre-trained network corresponds to being domain-independent ( See also Houlsby , Page 1 , Col. 1, Paragraph 1 , “ BERT, a Transformer network trained on large text corpora with an unsupervised loss ” )) Regarding claim 13: Houlsby discloses data processing hardware; and memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations ( Houlsby , Page 3, Col. 2, Paragraph 4,“ All runs are trained on 4 Google Cloud TPUs with a batch size of 32 ” where running the model on a TPU corresponds to using data processing hardware and memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations ) Houlsby discloses during an initial training stage, training only a backbone model to provide a first model configuration of the modular neural network model, the first model configuration comprising only the trained backbone model and adding an intrinsic sub-model to the trained backbone model ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Adapters are new modules added between layers of a pre-trained network. ” where pre-trained network corresponds to a first model configuration where the backbone is trained without an adapter module is not added to the model yet and training only a backbone model and the adaptor module corresponds to an intrinsic sub-model being added to a backbone model ) Houlsby discloses during a fine-tuning training stage: freezing parameters of the trained backbone model ( Houlsby , Page 2 , Col. 2, Paragraph 2 , “The weights of the original network are untouched…the parameters of the original network are frozen and therefore may be shared by many tasks.” ) and fine-tuning parameters of the intrinsic sub-model added to the trained backbone model while the parameters of the trained backbone model are frozen to provide a second model configuration, the second model configuration comprising the backbone model initially trained during the initial training stage and the intrinsic sub- model having the parameters fine-tuned during the fine-tuning stage ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “ Only the new, task specific, parameters, v, are then trained…During training, only v are tuned” where only training the adapters task specific parameters v corresponds creating a second model configuration where the fine-tuning parameters of the intrinsic sub model while the pre-trained model parameters are frozen ) Regarding claim 12 : The rejection of claim 1 with prior art Houlsby is incorporated and further: Houlsby discloses wherein during inference, the trained modular neural network model is configured to operate in any one of: the second model configuration comprising the backbone model initially trained during the initial training stage and the intrinsic sub-model having the parameters fine- tuned during the fine-tuning stage ( Houlsby , Page 2 , Col. 1, Paragraph 2 , “For adapter tuning, a new function, ψ w,v (x), is defined, where parameters w are copied over from pre-training. …During training, only v are tuned” where ψ w,v (x) represents the output of running input x through the model that includes the pretrained parameters (w) and the adapter(v) corresponds to inferring , using the being the backbone model initially trained and the intrinsic sub- model having parameters fine-tuned ( See also Houlsby , Page 2, Col. 1, Paragraph 4, “We demonstrate on a large and diverse set of text classification tasks that adapters yield parameter-efficient tuning for NLP.” ) ) The following limitations were not mapped as the claim language requires only one or more of the first, second or third model configurations be shown in prior art and second model configuration is shown in prior art as cited above: the first model configuration comprising only the trained backbone model and having the intrinsic sub-model removed ; o r a third model configuration comprising only the intrinsic sub-model having the parameters fine-tuned during the fine-tuning stage and the trained backbone model removed Regarding claim 15 : The rejection of claim 13 incorporated in claim 15 . Claim 15 is rejected under the same rationale as set forth in the rejection of claim 3 . Regarding claim 16 : The rejection of claim 15 incorporated in claim 16 . Claim 16 is rejected under the same rationale as set forth in the rejection of claim 4 . Regarding claim 17 : The rejection of claim 16 incorporated in claim 17 . Claim 17 is rejected under the same rationale as set forth in the rejection of claim 5 . Regarding claim 24 : The rejection of claim 13 incorporated in claim 24 . Claim 24 is rejected under the same rationale as set forth in the rejection of claim 12 . 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. 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 and 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) in view of Kannan et al.(“ Large-Scale Multilingual Speech Recognition with a Streaming End-to-End Model”, henceforth known as Kannan) Regarding claim 6 : The rejection of claim 4 with prior art Houlsby is incorporated and further: Houlsby doesn’t teach wherein the first domain is associated with speech recognition in a first language and the second domain is associated with speech recognition in a second language different than the first language . Kanna n discloses wherein the first domain is associated with speech recognition in a first language ( Kannan , Page 2, Col. 2, Paragraphs 4-5,“ Here we extend it to multilingual speech recognition … Adapter modules are effectively domain-specific (language-specific in our case) adjustments to the activations coming out of each layer. ”) and the second domain is associated with speech recognition in a second language different than the first language ( Kannan , Page 3, Col.1 , Table 1 and Paragraph 1 , “Importantly, each adapter module contains separate parameters for each language” where each adapter being for a separate language corresponds to a second domain with speech recognition in a second language different than the first language ( See also Kannan, Page 2, Col. 2, Figure 1, “For a Tamil utterance, only the Tamil adapters are applied to each activation” which emphases each adapter is focused on the domain of a specific language )) References Houlsby and Kannan are analogous art because they are from the same field of endeavor of using deep learning with adapters as a techniques and modular architectures with the goal of parameter efficiency. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby and Kannan before him or her, to modify the domain of the adapters of Houlsby to include the language -specific domains for adapters of Kannan because multilingual models tend to be biased toward languages with more data and the use of adapters with language specific adapters to capture the difference of language . The suggestion/motivation for doing so would have been Kannan, Page 2, Col. 1, Paragraph 5(“Imbalanced data typically leads to having a model perform better on languages with larger data”), Kannan, Page 2, Col. 2, Paragraph 3(“it is typical to have varying amounts of transcribed data available for different languages. As a result, a multilingual model will be more influenced by languages which are over-represented in the training set”), Kannan, Page 2, Col. 2, Paragraph 4( “A second architecture extension that we investigate to handle data imbalance is adapter modules.” ) and Kannan having Houlsby cited as a resource/reference. Regarding claim 18 : The rejection of claim 16 incorporated in claim 18 . Claim 18 is rejected under the same rationale as set forth in the rejection of claim 6 . Claim (s) 7, 9 , 19 and 2 1 is/are rejected under 35 U.S.C. 103 as being unpatentable over Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) in view of Eeckt et al.(“USING ADAPTERS TO OVERCOME CATASTROPHIC FORGETTING IN END-TO-END AUTOMATIC SPEECH RECOGNITION”, henceforth known as Eeckt ) Regarding claim 7 : The rejection of claim 1 with prior art Houlsby is incorporated and further: Houlsby doesn’t teach the modular neural network model comprises an end-to-end speech recognition model comprising an audio encoder and a decoder; training only the backbone model comprises updating parameters of the audio encoder or the decoder; and fine-tuning the parameters of the intrinsic sub-model comprises updating the parameters of the audio encoder or the decoder. Eeckt discloses the modular neural network model comprises an end-to-end speech recognition model ( Eeckt , Page 1, Abstract “The same applies to End-to-End (E2E) Automatic Speech Recognition (ASR) models, even for monolingual tasks. In this paper, we aim to overcome CF for E2E ASR by inserting adapters” and Eeckt , Page 1, Col. 2, Paragraph 3, “ X ∈ R F×f the input utterance consisting of F frames of d imension f, and y the ground truth transcription of w word pieces” where the model being an Automatic Speech Recognition model and handling speech corresponds to an end-to-end speech recognition model ) comprising an audio encoder and a decoder ( Eeckt , Page 1, Col. 2, Paragraph 3, “ The model is a hybrid encoder-decoder E2E model, consisting of a Conformer or Transformer encoder and Transformer decoder.”) Eeckt discloses training only the backbone model comprises updating parameters of the audio encoder or the decoder ( Eeckt , Page 1, Col. 2, Equation 1 and Paragraph 3, “…We denote L( X,y;θ ) the loss of the model with parameters θ ∈ R N (N is the number of parameters))” where the model is defined as a encoder and decoder with parameter vector θ representing all model parameters and the loss is minimized with respect to θ meaning the gradients are computed over the full model corresponds to training a backbone model updating parameters of the audio encoder and decoder ( See also Eeckt , Page 1, Col. 2, Paragraph 2, “…we use adapters…which we insert into the model and make task-specific, meaning that each task uses its own adapters.” ) ) Eeckt discloses fine-tuning the parameters of the intrinsic sub-model comprises updating the parameters of the audio encoder or the decoder ( Eeckt , Page 2, Col. 1, Paragraph 5, “ We, too, consider this method, called A/Freeze. In addition, we consider A/CFT (Cautious Fine-Tuning) which consists of two stages: 1) train the adapters of the new task while freezing the shared parameters (as in A/Freeze); 2) adapt the entire model with a ten times smaller learning rate ” where updating the entire model using A/CFT after adapters are inserted corresponds fine-tuning the intrinsic sub-model comprising updating parameters of the audio encoder or decode ) References Houlsby and Eeckt are analogous art because they are from the same field of endeavor of using transfer learning with adapters as a techniques and modular architectures with the goal of parameter efficiency. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby and Eeckt before him or her, to modify the encoder-only transformer of Houlsby with the e ncoder-decoder transformer model using A/CFT of Eeckt to protect shared parameters while adapting the model for new tasks better with A/CFT and handle sequence generation problems . The suggestion/motivation for doing so would have been Eeckt , Page 2 , Col. 1, Paragraph 5 (“ …thus updating the shared parameters cautiously might improve on the new task while preventing forgetting of the old tasks. ”) and Eeckt having Houlsby cited as a resource/reference. Regarding claim 9 : The rejection of claim 7 with prior art Houlsby - Eeckt is incorporated and further: Eeckt further discloses wherein the operations further comprise training another modular neural network, the other modular neural network comprising the other one of the audio encoder or the decoder of the end-to-end speech recognition mode l ( Eeckt , Page 1, Col. 2, Equation 1 and Paragraph 3, “…We denote L ( X , y ;θ ) the loss of the model with parameters θ ∈ RN (N is the number of parameters))” where the model is defined as a encoder and decoder with parameter vector θ representing all model parameters and the loss is minimized with respect to θ meaning the gradients are computed over the full model corresponds to a training a backbone model updating parameters of the audio encoder and the decoder ) Regarding claim 19 : The rejection of claim 13 incorporated in claim 19 . Claim 19 is rejected under the same rationale as set forth in the rejection of claim 7 . Regarding claim 21 : The rejection of claim 19 incorporated in claim 21 . Claim 21 is rejected under the same rationale as set forth in the rejection of claim 9 . Claim (s) 2 and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) in view of Sathyendra et al.(“CONTEXTUAL ADAPTERS FOR PERSONALIZED SPEECH RECOGNITION IN NEURAL TRANSDUCERS”, henceforth known as Sathyendra ) Regarding claim 2 : The rejection of claim 1 with prior art Houlsby is incorporated and further Houlsby discloses the backbone model comprises a non-attentive neural network comprising existing residual connections ( Houlsby , Page 3 , Col. 1, Figure 2 and Paragraph 2 , “… A skip-connection is applied across each of the sub-layers ” w here the left having feed-forward layers correspond to a non-attentive neural network and the skip layers correspond to having existing residual connection ) Houlsby discloses the intrinsic sub-model is added to the trained backbone model without requiring any residual adaptors or additional residual connection other than the existing residual connections ( Houlsby , Page 3, Col. 2, Figure 2 and Paragraph 2, “A skip-connection is applied across each of the sub-layers. The output of each sub-layer is fed into layer normalization. We insert two serial adapters after each of these sub-layers. The adapter is always applied directly to the output of the sub-layer, after the projection back to the input size, but before adding the skip connection back” where the pre-trained model containing residual connections in each sublayer and after insertion there is the same number of residual connections corresponds to adding the intrinsic sub-model to the trained backbone model without requiring additional residual connection other than the existing residual connections or requiring a residual adaptor ) Houlsby does not t each the intrinsic sub-model comprises an attention-based sub-model . Sathyendra discloses the intrinsic sub-model comprises an attention-based sub-model ( Sathyendra , Page 1, Col. 2, Paragraph 2, “ In this paper, we propose training lightweight contextual adapter networks to augment pretrained neural sequence transducers, such as the RNN-Transducer (RNN-T) … Specifically, the proposed contextual adapter network consists of a catalog encoder and an attention-based biasing adapter ” where the contextual adapter having an attention-based biasing adapter corresponds to an intrinsic sub-model having an attention-based sub-model ) References Houlsby and Sathyendra are analogous art because they are from the same field of endeavor of using deep learning with adapters as a techniques and modular architectures with the goal of parameter efficiency. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby and Sathyendra before him or her, to modify the adapters of Houlsby with attention-based /biased adapters of Sathyendra to help guide contextual knowledge over inputs and adapting to a user’s unique context . The suggestion/motivation for doing so would have been Sathyendra , Page 1 , Col. 2 , Paragraph 2 (“ …The biasing adapter measures the correlation between the pretrained model’s intermediate representations…and the context embeddings, to determine the contextual entities to attend over”), Sathyendra , Page 1, Col. 1, Abstract(“…we demonstrate that contextual adapters can be applied to any general purpose pretrained ASR model to improve personalization”) and Sathyendra having Houlsby cited as a resource/reference. Regarding claim 14 : The rejection of claim 13 incorporated in claim 14 . Claim 14 is rejected under the same rationale as set forth in the rejection of claim 2 . Claim (s) 8 and 2 0 is/are rejected under 35 U.S.C. 103 as being unpatentable over Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) in view of Eeckt et al.(“USING ADAPTERS TO OVERCOME CATASTROPHIC FORGETTING IN END-TO-END AUTOMATIC SPEECH RECOGNITION”, henceforth known as Eeckt ) and further in view of Kannan et al.(“ Large-Scale Multilingual Speech Recognition with a Streaming End-to-End Model”, henceforth known as Kannan) Regarding claim 8 : The rejection of claim 7 with prior art Houlsby - Eeckt is incorporated and further: Houlsby - Eeckt does not teach wherein the end-to-end speech recognition model comprises a recurrent neural network-transducer (RNN-T) architecture Kannan discloses wherein the end-to-end speech recognition model comprises a recurrent neural network-transducer (RNN-T) architecture ( Kannan , Page 1, Col. 2, Paragraph 2, “ First, we present a streaming E2E multilingual system using the Recurrent Neural Network Transducer (RNN-T)”) References Houlsby - Eeckt and Kannan are analogous art because they are from the same field of endeavor of using deep learning with adapters as a techniques and modular architectures with the goal of parameter efficiency. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby - Eeckt and Kannan before him or her, to modify encoder-decoder transformer model of Houlsby - Eeckt with the encoder- prediction-joint transducer model of Kannan to provide a straightforward streaming implementation . The suggestion/motivation for doing so would have been Kannan , Page 1, Col. 2, Paragraph 2(“ prior E2E multilingual work has been limited to … models that do not admit a straightforward streaming implementation ”) and Kannan having Houlsby cited as a resource/reference. Regarding claim 20 : The rejection of claim 19 incorporated in claim 20. Claim 20 is rejected under the same rationale as set forth in the rejection of claim 8 . Claim (s) 10-11 and 22-23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Houlsby et al.(“ Parameter-Efficient Transfer Learning for NLP ”, henceforth known as Houlsby ) in view of Gulati et al.(“ Conformer: Convolution-augmented Transformer for Speech Recognition ”, henceforth known as Gulati) and further in view of Zhao et al.(“ Tiny-Attention Adapter: Contexts Are More Important Than the Number of Parameters ”, henceforth known as Zhao) Regarding claim 10 : The rejection of claim 1 with prior art Houlsby is incorporated and further: Houlsby does not teach the backbone model comprises a first half feedforward layer; a convolution layer; a second half feedforward layer; and a layer norm layer or the intrinsic sub-model comprises a stack of one or more multi-head self-attention layers . Gulati discloses the backbone model comprises a first half feedforward layer; a convolution layer; a second half feedforward layer; and a layer norm layer (Gulati , Page 1, Figure 1, “Conformer comprises of two macaron-like feed-forward layers with half step residual connections sandwiching the multi-headed self attention and convolution modules. This is followed by a post layernorm ” w here the convolution-augmented transformer has a corresponding first half feedforward layer, a convolution layer, a second half feedforward layer and a layer norm layer ) References Houlsby and Gulati are analogous art because they are from the same field of endeavor of using Transformers and neural network architecture for deep learning Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby and Gulati before him or her, to modify transformer model of Houlsby with the convolution-augmented transformer model of Gulati to capture local patterns for better accuracy and efficiency . The suggestion/motivation for doing so would have been Gulati , Page 1, Col. 2, Paragraph 2(“…we propose a novel combination of self-attention and convolution…self-attention learns the global interaction whilst the convolutions efficiently capture…local correlations.”) and Gulati , Page 1, Col. 1 , Abstract (“ In this work, we achieve the best of both worlds…in a parameter-efficient way… Conformer significantly outperforms the previous Transformer and CNN based models achieving state-of-the-art accuracies ”). Neither Houlsby nor Gulati discloses the intrinsic sub-model comprises a stack of one or more multi-head self-attention layers . Zhao discloses the intrinsic sub-model comprises a stack of one or more multi-head self-attention layers (Zhao, Page 2, Col. 2, Paragraph 2-3, “ Our adapter has an attentative structure: as shown in Figure2b, at each position t, it takes as input the intermediate embeddings…from not only the current position…but also all the other positions … As shown in Figure2c, the internal architecture of our tiny-attention adapter resembles an ordinary multi-head attention mechanism” where the adapter has each token attend to all tokens in the same sequence corresponds to having a multi-head self-attention layers ) References Houlsby - Gulati and Zhao are analogous art because they are from the same field of endeavor of using deep learning with adapters as a techniques and modular architectures with the goal of parameter efficiency. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art, having the teachings of Houlsby - Gulati and Zhao before him or her, to modify adapter of Houlsby - Gulati with the tiny-attention adapter of Zhao to provide better context modeling . The suggestion/motivation for doing so would have been Zhao, Page 1, Col. 1, Abstract(“ Our tiny attention adapter learns to modify the hidden states at each position directly conditioned on the hidden states at all the other positions, which is missed by the previously proposed adapters. ”) , Zhao, Page 1, Col. 2, Paragraph s 1 -2 , “Almost all the previously proposed adapter architectures are feed-forward neural networks. Thus, we suspect that their embedding modifications are not as contextually rich as they should . Therefore, we propose to use the attentative structure that allows the embedding modifications of each token to capture more contextual information ” and Zhao Kannan having Houlsby cited as a resource/reference. Regarding claim 11 : The rejection of claim 1 0 with prior art Houlsby - Gulati -Zhao combination is incorporated and further: Gulati further discloses wherein the second model configuration comprises: the first half feedforward layer; the stack of one or more multi-head self-attention layers; the convolution layer; the second half feedforward layer; and the layer norm layer (Gulati , Page 1, Figure 1, “Conformer comprises of two macaron-like feed-forward layers with half step residual connections sandwiching the multi-headed self attention and convolution modules. This is followed by a post layernorm ” where t he convolution-augmented transformer has a corresponding first half feedforward layer, a stack of one or more multi-head self-attention layers, a convolution layer, a second half feedforward layer and a layer norm layer ) Regarding claim 2 2 : The rejection of claim 1 3 incorporated in claim 2 2 . Claim 2 2 is rejected under the same rationale as set forth in the rejection of claim 10 . Regarding claim 23 : The rejection of claim 22 incorporated in claim 2 3 . Claim 2 3 is rejected under the same rationale as set forth in the rejection of claim 11 . Conclusion Any inquiry concerning this communication or earlier communications from the examiner should be directed to FILLIN "Examiner name" \* MERGEFORMAT CHARLES JEFFREY JONES JR whose telephone number is FILLIN "Phone number" \* MERGEFORMAT (703)756-1414 . The examiner can normally be reached FILLIN "Work Schedule?" \* MERGEFORMAT Monday - Friday 8:00 - 5:00 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, FILLIN "SPE Name?" \* MERGEFORMAT Kakali Chaki can be reached at FILLIN "SPE Phone?" \* MERGEFORMAT 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. /C.J.J./ Examiner, Art Unit 2122 /KAKALI CHAKI/ Supervisory Patent Examiner, Art Unit 2122