DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . This action is in response to the application filed 08/21/2023. Claims 1, 13, 19, 21, 22, and 23 have been amended. Claims 1-23 are being examined.
Claim Objections
Claim 22 is objected to because of the following informalities: a spelling/typing error in the following limitation: and performing a non-linear function using en the multidimensional prediction to generate a multidimensional output comprising the next element. 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.
Claim 1-23 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1:
Step 1: Claim 1 recites a method which falls into the statutory category of process.
Step 2A Prong 1: Claim 1 recites multiple mathematical concepts: diagonally initializing a state matrix to produce a 1x1 discretized state convolutional kernel for performing pointwise state convolution; computing a multidimensional state at a current timestep by applying the 1x1 discretized state convolutional kernel to the multidimensional state at a previous timestep2 wherein computing the multidimensional state comprises using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type; computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep; and performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence.
Step 2A Prong 2: Claim 1 does not integrate the abstract idea into a practical application since the additional elements of:
A computer-implemented method, is merely an instruction to apply the abstract ideas using a generic computer.
Step 2B: Claim 1 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above, the additional element of implementing the method using a computer is nothing more than mere instructions to apply the abstract idea using a generic computer. Claim 1 is not patent eligible.
Regarding claim 2, the rejection of claim 1 is incorporated, further the claim recites further comprising computing a subsequent multidimensional prediction using the extended spatiotemporal sequence as the input spatiotemporal sequence; and performing the non-linear function using the subsequent multidimensional prediction to generate a subsequent multidimensional output. which amounts to applying the method of claim 1 using a prediction as the input. This merely recites the same abstract ideas as in claim 1 but with a different input and nothing more. Claim 2 is not patent eligible.
Regarding claim 3, the rejection of claim 1 is incorporated, further the claim recites wherein the multidimensional prediction is computed using an encoded version of the input spatiotemporal sequence. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 3 is not patent eligible.
Regarding claim 4, the rejection of claim 1 is incorporated, further the claim recites wherein the non-linear function generates an intermediate output and the intermediate output is decoded to generate the multidimensional output. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 4 is not patent eligible.
Regarding claim 5, the rejection of claim 1 is incorporated, further the claim recites wherein the at least one dimension is time. This limitation amounts to more specifics of the output produced by the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 5 is not patent eligible.
Regarding claim 6, the rejection of claim 1 is incorporated, further the claim recites wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 6 is not patent eligible.
Regarding claim 7, the rejection of claim 1 is incorporated, further the claim recites wherein the input spatiotemporal sequence comprises data for at least one of weather forecasting, traffic modeling, video prediction, video generation, and physics simulation. This limitation amounts to linking the use of the judicial exception to a particular technological environment or field of us (MPEP 2109.05(h)). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 7 is not patent eligible.
Regarding claim 8, the rejection of claim 1 is incorporated, further the claim recites wherein the input spatiotemporal sequence comprises biomedical or robotics data. This limitation amounts to linking the use of the judicial exception to a particular technological environment or field of us (MPEP 2109.05(h)). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 8 is not patent eligible.
Regarding claim 9, the rejection of claim 1 is incorporated, further the claim recites wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed on a server or in a data center and the multidimensional output is streamed to a user device. This limitation amounts to merely the use of a generic computer (server) to apply the judicial exception and insignificant extra-solution activity (streaming to a user) given it’s merely the transmitting of data and is therefore considered well-known, routine, conventional as evidenced by MPEP §2106.05(d)(II)(I). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 9 is not patent eligible.
Regarding claim 10, the rejection of claim 1 is incorporated, further the claim recites wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed within a cloud computing environment. This limitation amounts to applying the judicial exception with or by a generic machine (cloud computing). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 10 is not patent eligible.
Regarding claim 11, the rejection of claim 1 is incorporated, further the claim recites wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed for training, testing, or certifying a neural network employed in a machine, robot, or autonomous vehicle. This limitation amounts to applying the judicial exception with or by a generic machine. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 11 is not patent eligible.
Regarding claim 12, the rejection of claim 1 is incorporated, further the claim recites wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed on a virtual machine comprising a portion of a graphics processing unit. This limitation amounts to applying the judicial exception with or by a generic machine (virtual machine). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 12 is not patent eligible.
Regarding claim 13:
Step 1: Claim 13 recites a system which falls into the statutory category of machine.
Step 2A Prong 1: Claim 13 recites multiple mathematical concepts: diagonally initializing a state matrix to produce a 1x1 discretized state convolutional kernel for performing pointwise state convolution; computing a multidimensional state at a current timestep by applying the discretized state convolutional kernel to the multidimensional state at a previous timestep wherein computing the multidimensional state comprises using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type; computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep; and performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence.
Step 2A Prong 2: Claim 13 does not integrate the abstract idea into a practical application since the additional elements of:
a memory that stores the input spatiotemporal sequence, is insignificant extra-solution activity.
a processor that is connected to the memory, wherein the processor is configured to extend the input spatiotemporal sequence, is merely an instruction to apply the abstract ideas using a generic computer.
Step 2B: Claim 13 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above, the additional element of implementing the method using a processor is nothing more than mere instructions to apply the abstract idea using a generic computer and the additional element of storing data in memory Is considered insignificant extra-solution activity because it is well-understood, routine, conventional activity as evidenced by MPEP §2106.05(d)(II)(Iv). Claim 13 is not patent eligible.
Regarding claim 14, the rejection of claim 13 is incorporated, further the claim recites wherein the multidimensional prediction is computed using an encoded version of the input spatiotemporal sequence. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 14 is not patent eligible.
Regarding claim 15, the rejection of claim 13 is incorporated, further the claim recites wherein the non-linear function generates an intermediate output and the intermediate output is decoded to generate the multidimensional output. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 15 is not patent eligible.
Regarding claim 16, the rejection of claim 13 is incorporated, further the claim recites wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 16 is not patent eligible.
Regarding claim 17, the rejection of claim 13 is incorporated, further the claim recites wherein the input spatiotemporal sequence comprises data for at least one of weather forecasting, traffic modeling, video prediction, video generation, and physics simulation. This limitation amounts to linking the use of the judicial exception to a particular technological environment or field of us (MPEP 2109.05(h)). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 17 is not patent eligible.
Regarding claim 18, the rejection of claim 13 is incorporated, further the claim recites wherein the input spatiotemporal sequence comprises biomedical or robotics data. This limitation amounts to linking the use of the judicial exception to a particular technological environment or field of us (MPEP 2109.05(h)). As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 18 is not patent eligible.
Regarding claim 19:
Step 1: Claim 19 recites a non-transitory computer-readable media which falls into the statutory category of manufacture.
Step 2A Prong 1: Claim 19 recites multiple mathematical concepts: diagonally initializing a state matrix to produce a 1x1 discretized state convolutional kernel for performing pointwise state convolution; computing a multidimensional state at a current timestep by applying the discretized state convolutional kernel to the multidimensional state at a previous timestep, wherein computing the multidimensional state comprises using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type; computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep; and performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence.
Step 2A Prong 2: Claim 19 does not integrate the abstract idea into a practical application since the additional elements of:
when executed by one or more processors, cause the one or more processors to perform the steps of, is merely an instruction to apply the abstract ideas using a generic computer.
Step 2B: Claim 19 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above, the additional element of implementing the method using a processor is nothing more than mere instructions to apply the abstract idea using a generic. Claim 19 is not patent eligible.
Regarding claim 20, the rejection of claim 19 is incorporated, further the claim recites wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence. This limitation amounts to more specifics of the judicial exception. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical application, nor to significantly more than the judicial exception. Claim 20 is not patent eligible.
Regarding claim 21:
Step 1: Claim 21 recites a method which falls into the statutory category of process.
Step 2A Prong 1: Claim 21 recites multiple abstract ideas: reshaping a diagonal state matrix to produce a 1 x 1 discretized state convolutional kernel for performing pointwise state convolution; forming initial scan tuples for a recurrence for computing a multidimensional state; performing parallel scan computations on the initial scan tuples using an associative operator that takes at least two tuples and produces a new tuple of a same type; applying the 1x1 discretized state convolutional kernel to the multidimensional state at a previous timestep to produce the multidimensional state at the current timestep; and computing the next element in the input spatiotemporal sequence using the input spatiotemporal sequence and the multidimensional state at the current timestep. These limitations amount to mental processes (reshaping and forming tuples) or mathematical concepts (parallel computations using an associative operator, computing a multidimensional state, and computing the next element).
Step 2A Prong 2: Claim 21 does not integrate the abstract idea into a practical application since the additional elements of:
A computer-implemented method, is merely an instruction to apply the abstract ideas using a generic computer.
Receiving an input spatiotemporal sequence of data for weather forecasting, traffic modeling, video generation, or physics simulation, is insignificant extra-solution activity.
Step 2B: Claim 21 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above, the additional element of implementing the method using a computer is nothing more than mere instructions to apply the abstract idea using a generic computer and the additional element of receiving data Is considered insignificant extra-solution activity because it is well-understood, routine, conventional activity as evidenced by MPEP §2106.05(d)(II)(I). Claim 21 is not patent eligible.
Regarding claim 22, the rejection of claim 21 is incorporated, further the claim recites wherein computing the next element comprises: computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep; and performing a non-linear function using en the multidimensional prediction to generate a multidimensional output comprising the next element. This limitation amounts to more specifics of the abstract idea of computing the next element and merely describes mathematical concepts. As such the claim does not have any additional elements that amount to an integration of the judicial exception into a practical exception, nor to significantly more than the judicial exception. Claim 22 is not patent eligible.
Regarding claim 23, the rejection of claim 21 is incorporated, further the claim recites further comprising computing a subsequent element using the input spatiotemporal sequence and the next element which amounts to applying the method of claim 21 using a prediction alongside the sequence as the input. This merely recites the same abstract ideas as in claim 21 but with a different input and nothing more. Claim 23 is not patent eligible.
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) 21, 23 are rejected under 35 U.S.C. 103 as being unpatentable over Gu (Efficiently Modeling Long Sequences with Structured State Spaces) in view of Shi (Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting) and Martin (PARALLELIZING LINEAR RECURRENT NEURAL NETS OVER SEQUENCE LENGTH).
Regarding claim 21, Gu teaches receiving an input spatiotemporal sequence of data for weather forecasting, traffic modeling, video generation, or physics simulation (Table 9, Notably S4 is better on the longest setting in each task, e.g. reducing MSE by 37% when forecasting 30 days of weather data); reshaping a diagonal state matrix to produce a (Section 2.3, Notationally, throughout this paper we use Ā ,... to denote discretized SSM matrices defined by (3)); applying the discretized state convolutional kernel to the multidimensional state at a previous timestep to produce the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state; Pg. 10, S4 requires constant memory and computation per time step); and computing the next element in the input spatiotemporal sequence using the input spatiotemporal sequence and the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state, y is the multidimensional prediction, and u is the input sequence).
Gu fails to teach forming initial scan tuples for a recurrence for performing parallel scan computations on the initial scan tuples using an associative operator that takes at least two tuples and produces a new tuple of a same type. Martin teaches teach forming initial scan tuples for a recurrence for performing parallel scan computations on the initial scan tuples using an associative operator that takes at least two tuples and produces a new tuple of a same type (Section 1, Throughout this work, the scan of ⊕ with initial value b is defined as SCAN(⊕,[a1,a2,...,an],b) = [(a1 ⊕ b),(a2 ⊕ a1 ⊕ b),...,(an ⊕ an−1... ⊕ a1 ⊕ b)]. The reduction of ⊕ over array A and initial value b is denoted REDUCE(⊕,A,b) and is the final element of SCAN(⊕,A,b). Despite their dependent computation graph, algorithms exist to parallelize scans and reductions when ⊕ is associative (Ladner and Fischer, 1980), Algorithm 1 Parallel linear recurrences on p processors).
Martin is analogous to the claimed invention because it is in the field of recurrent neural networks and linear recurrences. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the method used in Martin to the recurrence relations in Gu to “speed up training and inference” (Martin Abs).
The combination of Gu and Martin fails to teach a 1 x 1 convolutional kernel for performing pointwise state convolution. Shi teaches a 1 x 1 convolutional kernel for performing pointwise state convolution (Table 1, ‘-5x5’ and ‘-1x1’ represent the corresponding state-to-state kernel size, which is either 5×5 or 1×1, Section 3.2, As our prediction target has the same dimensionality as the input, we concatenate all the states in the forecasting network and feed them into a 1 × 1 convolutional layer to generate the final prediction.).
Shi is analogous to the claimed invention because it Is in the field of sequence prediction using convolution and states. Therefore, it would have been obvious to one of ordinary skill in the art before the claimed invention to have added the 1x1 convolutional kernel to state-to-state transitions (such as those found in Gu and Shi) to better capture spatiotemporal correlations (Abs Shi) while limiting the size of the matrix to limit the cost of the parallel scanning method found in Martin when used on state-to-state recurrence relations (Martin Section 2.1, The O(n3) cost of the matrix multiplication in the parallel algorithm can counter-act any parallel speedups for sufficiently large hidden states and lead to a slower algorithm overall.)
Regarding claim 23, Gu teaches further comprising computing a subsequent element using the input spatiotemporal sequence and the next element (Title, Efficiently Modeling Long Sequences with Structured State Spaces; Table 9, Univariate long sequence time-series forecasting results, SSMs are used to generate long sequence therefore a sequence of predictions each depending on the previous prediction are generated).
Claim(s) 1-7, 9, 13-17, 19, 20, 22 are rejected under 35 U.S.C. 103 as being unpatentable over Gu (Efficiently Modeling Long Sequences with Structured State Spaces) in view of Shi (Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting) and Martin (PARALLELIZING LINEAR RECURRENT NEURAL NETS OVER SEQUENCE LENGTH) and Nazir (US10996374B1).
Regarding claim 1, Gu teaches method of extending an input spatiotemporal sequence in at least one dimension (Title, Efficiently Modeling Long Sequences, Section 4.3, The tasks we focus on (generative modeling, image classification, time-series forecasting) are considered as LRD tasks in the literature, and serve as additional validation that S4 handles LRDs efficiently) comprising: diagonally initializing a state matrix (Section 1, special state matrices A, Abs, Our technique involves conditioning A with a low-rank correction, allowing it to be diagonalized stably) to produce a discretized state convolutional kernel (Section 2.3, Notationally, throughout this paper we use Ā ,... to denote discretized SSM matrices defined by (3)); computing a multidimensional state at a current timestep by applying the discretized state convolutional kernel to the multidimensional state at a previous timestep2 (Fig 1, Fast Discrete Representations, where x is the multidimensional state); computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state, y is the multidimensional prediction, and u is the input sequence; Pg. 10, S4 requires constant memory and computation per time step).
Gu fails to teach using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type. Martin teaches using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type (Section 1, Throughout this work, the scan of ⊕ with initial value b is defined as SCAN(⊕,[a1,a2,...,an],b) = [(a1 ⊕ b),(a2 ⊕ a1 ⊕ b),...,(an ⊕ an−1... ⊕ a1 ⊕ b)]. The reduction of ⊕ over array A and initial value b is denoted REDUCE(⊕,A,b) and is the final element of SCAN(⊕,A,b). Despite their dependent computation graph, algorithms exist to parallelize scans and reductions when ⊕ is associative (Ladner and Fischer, 1980), Algorithm 1 Parallel linear recurrences on p processors).
Martin is analogous to the claimed invention because it is in the field of recurrent neural networks and linear recurrences. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the method used in Martin to the recurrence relations in Gu to “speed up training and inference” (Martin Abs).
The combination of Gu and Martin fails to teach a 1 x 1 convolutional kernel for performing pointwise state convolution. Shi teaches a 1 x 1 convolutional kernel for performing pointwise state convolution (Table 1, ‘-5x5’ and ‘-1x1’ represent the corresponding state-to-state kernel size, which is either 5×5 or 1×1, Section 3.2, As our prediction target has the same dimensionality as the input, we concatenate all the states in the forecasting network and feed them into a 1 × 1 convolutional layer to generate the final prediction.).
Shi is analogous to the claimed invention because it Is in the field of sequence prediction using convolution and states. Therefore, it would have been obvious to one of ordinary skill in the art before the claimed invention to have added the 1x1 convolutional kernel to state-to-state transitions (such as those found in Gu and Shi) to better capture spatiotemporal correlations (Abs Shi) while limiting the size of the matrix to limit the cost of the parallel scanning method found in Martin when used on state-to-state recurrence relations (Martin Section 2.1, The O(n3) cost of the matrix multiplication in the parallel algorithm can counter-act any parallel speedups for sufficiently large hidden states and lead to a slower algorithm overall.)
The combination of Gu, Martin, and Shi fails to teach performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence. Nazir teaches performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence (Fig 7, As illustrated at 704, a forecast of local weather indicators can also be received and processed. As illustrated 706, a difference between the forecast and the measured local weather indicators can be determined. Such difference is determined particularly for historical data that can be used in generating a correction module, as illustrated at 708. In an example, the correction module can take the form of a neural network model, such as a LSTM module, the correction module, a neural network containing non-linear layers, is the non-linear function).
Nazir is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data such as weather. Furthermore, because Nazir states that the weather forecast engine can be a state space model (Col. 5 lines 20-25, Weather forecast engine 210 can utilize heuristics or algorithms to form regression models, Markov chains, time series models, state space models), it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in the prediction produced by the state space model and runs it through a correction module in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Regarding claim 2, Gu teaches further comprising computing a subsequent multidimensional prediction using the extended spatiotemporal sequence as the input spatiotemporal sequence (Title, Efficiently Modeling Long Sequences with Structured State Spaces; Table 9, Univariate long sequence time-series forecasting results, SSMs are used to generate long sequence therefore a sequence of predictions each depending on the previous prediction are generated) and Nazir teaches performing the non-linear function using the subsequent multidimensional prediction to generate a subsequent multidimensional output (Fig 7, As illustrated at 704, a forecast of local weather indicators can also be received and processed. As illustrated 706, a difference between the forecast and the measured local weather indicators can be determined. Such difference is determined particularly for historical data that can be used in generating a correction module, as illustrated at 708. In an example, the correction module can take the form of a neural network model, such as a LSTM module, the correction module, a neural network containing non-linear layers, is the non-linear function).
Regarding claim 3, Gu teaches the multidimensional prediction is computed using an encoded version of the input spatiotemporal sequence (Fig 1, Fast Discrete Representations, y is produced using a version of U encoded through the use of B).
Regarding claim 4, Nazir teaches wherein the non-linear function generates an intermediate output and the intermediate output is decoded to generate the multidimensional output (Col. 7, lines 52-58, In an example, the correction module can take the form of a neural network model, such as a LSTM module… Utilizing the generated module, future forecasts of local weather indicators can then be used as an input to forecast corrected weather indicators, here the non-linear function is the correction module, a neural network such as a LSTM, which can encode inputs creating some latent or intermediate output, which is later decoded into the output, the corrected forecast). Gu and Nazir are analogous to the claimed invention as described above and it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in a prediction and outputs another prediction in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Regarding claim 5, Gu teaches wherein the at least one dimension is time (Table 9, Univariate long sequence time-series forecasting results).
Regarding claim 6, Gu teaches wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence (Fig 1, Fast Discrete Representations, where x is the multidimensional state, u is the input, and Ā is the kernel).
Regarding claim 7, Gu teaches wherein the input spatiotemporal sequence comprises data for at least one of weather forecasting, traffic modeling, video prediction, video generation, and physics simulation (Table 9, Notably S4 is better on the longest setting in each task, e.g. reducing MSE by 37% when forecasting 30 days of weather data).
Regarding claim 9, Nazir teaches wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed on a server or in a data center and the multidimensional output is streamed to a user device (Abs, A system includes at least one server implementing a weather forecast engine, Col. 2 lines 50-52, In addition, a user device 110 can access the one or more servers 102 to interact with the weather forecasting engine 104).
Gu and Nazir are analogous to the claimed invention as described above and it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a server to perform the forecasting and stream it to a user device because using servers allows the forecast system to draw in more complex data (Col. 2 lines 28-37) and streaming to users improves the forecasting system by allowing users to display and interact with forecasts.
Regarding claim 13, Gu teaches diagonally initializing a state matrix (Section 1, special state matrices A, Abs, Our technique involves conditioning A with a low-rank correction, allowing it to be diagonalized stably) to produce a discretized state convolutional kernel (Section 2.3, Notationally, throughout this paper we use Ā ,... to denote discretized SSM matrices defined by (3)); computing a multidimensional state at a current timestep by applying the discretized state convolutional kernel to the multidimensional state at a previous timestep2 (Fig 1, Fast Discrete Representations, where x is the multidimensional state); computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state, y is the multidimensional prediction, and u is the input sequence; Pg. 10, S4 requires constant memory and computation per time step).
Gu fails to teach using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type. Martin teaches using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type (Section 1, Throughout this work, the scan of ⊕ with initial value b is defined as SCAN(⊕,[a1,a2,...,an],b) = [(a1 ⊕ b),(a2 ⊕ a1 ⊕ b),...,(an ⊕ an−1... ⊕ a1 ⊕ b)]. The reduction of ⊕ over array A and initial value b is denoted REDUCE(⊕,A,b) and is the final element of SCAN(⊕,A,b). Despite their dependent computation graph, algorithms exist to parallelize scans and reductions when ⊕ is associative (Ladner and Fischer, 1980), Algorithm 1 Parallel linear recurrences on p processors).
Martin is analogous to the claimed invention because it is in the field of recurrent neural networks and linear recurrences. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the method used in Martin to the recurrence relations in Gu to “speed up training and inference” (Martin Abs).
The combination of Gu and Martin fails to teach a 1 x 1 convolutional kernel for performing pointwise state convolution. Shi teaches a 1 x 1 convolutional kernel for performing pointwise state convolution (Table 1, ‘-5x5’ and ‘-1x1’ represent the corresponding state-to-state kernel size, which is either 5×5 or 1×1, Section 3.2, As our prediction target has the same dimensionality as the input, we concatenate all the states in the forecasting network and feed them into a 1 × 1 convolutional layer to generate the final prediction.).
Shi is analogous to the claimed invention because it Is in the field of sequence prediction using convolution and states. Therefore, it would have been obvious to one of ordinary skill in the art before the claimed invention to have added the 1x1 convolutional kernel to state-to-state transitions (such as those found in Gu and Shi) to better capture spatiotemporal correlations (Abs Shi) while limiting the size of the matrix to limit the cost of the parallel scanning method found in Martin when used on state-to-state recurrence relations (Martin Section 2.1, The O(n3) cost of the matrix multiplication in the parallel algorithm can counter-act any parallel speedups for sufficiently large hidden states and lead to a slower algorithm overall.)
The combination of Gu, Martin, and Shi fails to teach performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence. Nazir teaches performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence (Fig 7, As illustrated at 704, a forecast of local weather indicators can also be received and processed. As illustrated 706, a difference between the forecast and the measured local weather indicators can be determined. Such difference is determined particularly for historical data that can be used in generating a correction module, as illustrated at 708. In an example, the correction module can take the form of a neural network model, such as a LSTM module, the correction module, a neural network containing non-linear layers, is the non-linear function).
Nazir is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data such as weather. Furthermore, because Nazir states that the weather forecast engine can be a state space model (Col. 5 lines 20-25, Weather forecast engine 210 can utilize heuristics or algorithms to form regression models, Markov chains, time series models, state space models), it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in the prediction produced by the state space model and runs it through a correction module in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Regarding claim 14, Gu teaches the multidimensional prediction is computed using an encoded version of the input spatiotemporal sequence (Fig 1, Fast Discrete Representations, y is produced using a version of U encoded through the use of B).
Regarding claim 15, Nazir teaches wherein the non-linear function generates an intermediate output and the intermediate output is decoded to generate the multidimensional output (Col. 7, lines 52-58, In an example, the correction module can take the form of a neural network model, such as a LSTM module… Utilizing the generated module, future forecasts of local weather indicators can then be used as an input to forecast corrected weather indicators, here the non-linear function is the correction module, a neural network such as a LSTM, which can encode inputs creating some latent or intermediate output, which is later decoded into the output, the corrected forecast). Gu and Nazir are analogous to the claimed invention as described above and it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in a prediction and outputs another prediction in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Regarding claim 16, Gu teaches wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence (Fig 1, Fast Discrete Representations, where x is the multidimensional state, u is the input, and Ā is the kernel).
Regarding claim 17, Gu teaches wherein the input spatiotemporal sequence comprises data for at least one of weather forecasting, traffic modeling, video prediction, video generation, and physics simulation (Table 9, Notably S4 is better on the longest setting in each task, e.g. reducing MSE by 37% when forecasting 30 days of weather data).
Regarding claim 19, diagonally initializing a state matrix (Section 1, special state matrices A, Abs, Our technique involves conditioning A with a low-rank correction, allowing it to be diagonalized stably) to produce a discretized state convolutional kernel (Section 2.3, Notationally, throughout this paper we use Ā ,... to denote discretized SSM matrices defined by (3)); computing a multidimensional state at a current timestep by applying the discretized state convolutional kernel to the multidimensional state at a previous timestep2 (Fig 1, Fast Discrete Representations, where x is the multidimensional state); computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state, y is the multidimensional prediction, and u is the input sequence; Pg. 10, S4 requires constant memory and computation per time step).
Gu fails to teach using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type. Martin teaches using parallel scans to compute a recurrence by forming initial scan tuples and utilizing an associative operator that takes at least two tuples and produces a new tuple of a same type (Section 1, Throughout this work, the scan of ⊕ with initial value b is defined as SCAN(⊕,[a1,a2,...,an],b) = [(a1 ⊕ b),(a2 ⊕ a1 ⊕ b),...,(an ⊕ an−1... ⊕ a1 ⊕ b)]. The reduction of ⊕ over array A and initial value b is denoted REDUCE(⊕,A,b) and is the final element of SCAN(⊕,A,b). Despite their dependent computation graph, algorithms exist to parallelize scans and reductions when ⊕ is associative (Ladner and Fischer, 1980), Algorithm 1 Parallel linear recurrences on p processors).
Martin is analogous to the claimed invention because it is in the field of recurrent neural networks and linear recurrences. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the method used in Martin to the recurrence relations in Gu to “speed up training and inference” (Martin Abs).
The combination of Gu and Martin fails to teach a 1 x 1 convolutional kernel for performing pointwise state convolution. Shi teaches a 1 x 1 convolutional kernel for performing pointwise state convolution (Table 1, ‘-5x5’ and ‘-1x1’ represent the corresponding state-to-state kernel size, which is either 5×5 or 1×1, Section 3.2, As our prediction target has the same dimensionality as the input, we concatenate all the states in the forecasting network and feed them into a 1 × 1 convolutional layer to generate the final prediction.).
Shi is analogous to the claimed invention because it Is in the field of sequence prediction using convolution and states. Therefore, it would have been obvious to one of ordinary skill in the art before the claimed invention to have added the 1x1 convolutional kernel to state-to-state transitions (such as those found in Gu and Shi) to better capture spatiotemporal correlations (Abs Shi) while limiting the size of the matrix to limit the cost of the parallel scanning method found in Martin when used on state-to-state recurrence relations (Martin Section 2.1, The O(n3) cost of the matrix multiplication in the parallel algorithm can counter-act any parallel speedups for sufficiently large hidden states and lead to a slower algorithm overall.)
The combination of Gu, Martin, and Shi fails to teach performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence. Nazir teaches performing a non-linear function using the multidimensional prediction to generate a multidimensional output that extends the input spatiotemporal sequence in the at least one dimension, producing an extended spatiotemporal sequence (Fig 7, As illustrated at 704, a forecast of local weather indicators can also be received and processed. As illustrated 706, a difference between the forecast and the measured local weather indicators can be determined. Such difference is determined particularly for historical data that can be used in generating a correction module, as illustrated at 708. In an example, the correction module can take the form of a neural network model, such as a LSTM module, the correction module, a neural network containing non-linear layers, is the non-linear function).
Nazir is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data such as weather. Furthermore, because Nazir states that the weather forecast engine can be a state space model (Col. 5 lines 20-25, Weather forecast engine 210 can utilize heuristics or algorithms to form regression models, Markov chains, time series models, state space models), it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in the prediction produced by the state space model and runs it through a correction module in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Regarding claim 20, Gu teaches wherein computing the multidimensional state at the current timestep further comprises applying a discretized input convolutional kernel to the input spatiotemporal sequence (Fig 1, Fast Discrete Representations, where x is the multidimensional state, u is the input, and Ā is the kernel).
Regarding claim 22, Gu in view of Martin and Shi teaches the method of claim 21, further Gu teaches wherein computing the next element comprises: computing a multidimensional prediction using the input spatiotemporal sequence and the multidimensional state at the current timestep (Fig 1, Fast Discrete Representations, where x is the multidimensional state, y is the multidimensional prediction, and u is the input sequence). Gu fails to teach and performing a non-linear function using the multidimensional prediction to generate a multidimensional output comprising the next element.
Nazir teaches and performing a non-linear function using the multidimensional prediction to generate a multidimensional output comprising the next element (Fig 7, As illustrated at 704, a forecast of local weather indicators can also be received and processed. As illustrated 706, a difference between the forecast and the measured local weather indicators can be determined. Such difference is determined particularly for historical data that can be used in generating a correction module, as illustrated at 708. In an example, the correction module can take the form of a neural network model, such as a LSTM module, the correction module, a neural network containing non-linear layers, is the non-linear function).
Nazir is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data such as weather. Furthermore, because Nazir states that the weather forecast engine can be a state space model (Col. 5 lines 20-25, Weather forecast engine 210 can utilize heuristics or algorithms to form regression models, Markov chains, time series models, state space models), it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified Gu to incorporate the use of a correction module that takes in the prediction produced by the state space model and runs it through a correction module in order to reduce error and improve forecasts (Col. 7 line 55-65, Table 1 indicates that the root mean square error is reduced using such a correction module which is further illustrated in the figures. For example, FIG. 20 shows an improvement in the six-hour forecast of windspeed).
Claim(s) 8, 10-12, 18 are rejected under 35 U.S.C. 103 as being unpatentable over Gu (Efficiently Modeling Long Sequences with Structured State Spaces) in view of Shi (Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting) and Martin (PARALLELIZING LINEAR RECURRENT NEURAL NETS OVER SEQUENCE LENGTH) and Nazir (US10996374B1) as applied to claim 1 above, and further in view of Rangapuram (US11281969B1).
Regarding claim 8, Gu in view of Shi, Martin, and Nazir teaches the method of claim 1 but fails to teach wherein the input spatiotemporal sequence comprises biomedical or robotics data. Rangapuram teaches wherein the input spatiotemporal sequence comprises biomedical or robotics data (Col. 3 lines 60-65, Other types of domain specific metadata may be used for time series pertaining to autonomous vehicles, robotics, image/speech processing applications and the like).
Rangapuram is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data using machine learning models such as state space models. It would have been obvious for one of ordinary skill in the art before the filling date of the claimed invention to have used robotics data alongside the invention in Gu because Rangapuram states that state space models can be used to forecast time series data that can be robotics data (Abs, A composite time series forecasting model comprising a neural network sub-model and one or more state space sub-models corresponding to individual time series is trained, Col. 9 lines 13-15, In another embodiment, the forecasts may for example be used to cause a robotic system and/or an autonomous vehicle system…) therefore the weather time series data can be substituted with robotics data.
Regarding claim 10, Gu in view of Shi, Martin, and Nazir teaches the method of claim 1 but fails to teach wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed within a cloud computing environment. Rangapuram teaches wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed within a cloud computing environment (Col. 21 lines 15-20, FIG. 10 illustrates an example provider network environment in which a forecasting service may be implemented…..networks set up by an entity such as a company or a public sector organization to provide one or more network-accessible services (such as various types of cloud-based computing, storage or analytics services)).
Rangapuram is analogous to the claimed invention as stated above, therefore it would have been obvious for one of ordinary skill in the art before the filling date of the claimed invention to have used a network environment including cloud-based computing to implement the forecasting to eliminate the need for heavy on-premises infrastructure.
Regarding claim 11, Gu in view of Shi, Martin, and Nazir teaches the method of claim 1 but fails to teach wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed for training, testing, or certifying a neural network employed in a machine, robot, or autonomous vehicle. Rangapuram teaches wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed for training, testing, or certifying a neural network employed in a machine, robot, or autonomous vehicle. (Col. 9 lines 13-17, In another embodiment, the forecasts may for example be used to cause a robotic system and/or an autonomous vehicle system, e.g., within a factory or warehouse, to relocate items from one location to another, or to reposition the robotic system or autonomous vehicle system in anticipation of one or more events).
Rangapuram is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data using machine learning models such as state space models. It would have been obvious for one of ordinary skill in the art before the filling date of the claimed invention to have used the state space model in Gu in robotics or autonomous vehicles because Rangapuram states that forecast systems using state space models can be used in robotic/autonomous vehicle systems (Col. 23 lines 43-50, Similarly, for other problem domains such as human resources planning, automated data center resource provisioning/planning, traffic modeling/planning, autonomous vehicle or robot trajectory planning and the like, the forecasting techniques described may be able to generate high-quality forecasts using relatively short time series…).
Regarding claim 12, Gu in view of Shi, Martin, and Nazir teaches the method of claim 1 but fails to teach wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed on a virtual machine comprising a portion of a graphics processing unit. Rangapuram teaches wherein at least one of the steps of diagonally initializing, computing the multidimensional state, computing the multidimensional prediction, and performing is performed on a virtual machine comprising a portion of a graphics processing unit (Col. 21 lines 41-45, for example, the forecasting service 1043, and/or the machine learning service 1071 may utilize virtual machines instantiated at the virtual computing), Col. 22 lines 1-4, In some embodiments, special execution servers 1076 (e.g., servers comprising graphics processing units (GPUs) or other processors optimized specifically for machine learning) may be available at the MLS 1071, and may be employed for some of the algorithms/models trained and executed by the forecasting service 1043).
Rangapuram is analogous to the claimed invention as stated above, therefore it would have been obvious for one of ordinary skill in the art before the filling date of the claimed invention to have used a network environment including a virtual machine to implement the forecasting to enable multiple environments to run on shared hardware, furthermore, it would have been obvious before the effective filling date of the claimed invention to use a GPU in this network environment because they can handle large scale computations, such as those needed during machine learning, efficiently.
Regarding claim 18, Gu in view of Shi, Martin, and Nazir teaches the system of claim 13 but fails to teach wherein the input spatiotemporal sequence comprises biomedical or robotics data. Rangapuram teaches wherein the input spatiotemporal sequence comprises biomedical or robotics data (Col. 3 lines 60-65, Other types of domain specific metadata may be used for time series pertaining to autonomous vehicles, robotics, image/speech processing applications and the like).
Rangapuram is analogous to the claimed invention because it is in the field of endeavor of predicting or forecasting sequences of data using machine learning models such as state space models. It would have been obvious for one of ordinary skill in the art before the filling date of the claimed invention to have used robotics data alongside the invention in Gu because Rangapuram states that state space models can be used to forecast time series data that can be robotics data (Abs, A composite time series forecasting model comprising a neural network sub-model and one or more state space sub-models corresponding to individual time series is trained, Col. 9 lines 13-15, In another embodiment, the forecasts may for example be used to cause a robotic system and/or an autonomous vehicle system…) therefore the weather time series data can be substituted with robotics data.
Response to Arguments
Regarding Prior Art rejections (35 U.S.C 102 and 103):
Applicant’s arguments with respect to claim(s) 1-23 have been considered but are moot because they do not apply to the new references Martin and Shi.
Regarding 35 U.S.C 101 rejection:
Applicant's arguments filed 06/23/2026 have been fully considered but they are not persuasive. The applicant asserts the claims are directed to a specific improvement in computer processing of spatiotemporal sequences. The examiner respectfully disagrees. Firstly, the claims still recite mathematical concepts as stated in the 101 rejection above. The 1x1 kernel for pointwise state convolution and the parallel scan recurrence still fall under mathematical concepts given they are simply matrix operations and an algorithm consisting of mathematical operations for the parallel scans. Additionally, the applicant asserts even if a mathematical concept is involved, the amended claims incorporate the specific solution, of using parallel scan computations, by requiring the 1x1 discretized state convolution kernel for performing pointwise state convolution. The applicant asserts the parallel scan recurrence mechanism supported by the 1x1 kernel improves computer-processing and as such should be eligible under Step 2A Prong 2 and Step 2B given the claim recite a non-generic arrangement of state convolution, tuple-based associative parallel scan recurrence, and spatiotemporal sequence prediction. The examiner respectfully disagrees. The claims as recited do not explicitly link the 1x1 kernel and the parallel scans to a clear and concrete technological improvement but instead merely cite a series of mathematical concepts in algorithmic fashion. If the claims were amended to better reflect the specific improvements in computer processing, the applicants’ arguments would work. Furthermore, even if the parallel scans were not considered an abstract idea there would still be some chance that they could be considered well-understood, routine, conventional given parallel scans for recurrences are known in the art. To overcome the 101 rejection, the applicant should strengthen the connection between the claims and the specific technological improvements described in the applicant’s argument.
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.
/NATNAEL A ASEGDEW/ Examiner, Art Unit 2122
/MICHAEL H HOANG/ PRIMARY EXAMINER, Art Unit 2122
