DETAILED ACTION
1. This office action is in response to the Application No. 18885750 filed on 03/23/2026. Claims 13-18 has been cancelled, claims 1-12 are presented for examination and are currently pending. Applicant’s arguments have been carefully and respectfully considered.
Response to Arguments
2. Applicant’s arguments are moot in view of the new grounds of rejection. The Examine is withdrawing the rejections in the previous office action because the Applicant’s amendments necessitated the new grounds of rejection.
Claim 1 now recites “use the temporal derivatives to modify a plurality of parameters for a plurality of attention mechanisms within the latent transformer and to detect when latent space vectors deviate from temporal patterns determined by the temporal derivatives”.
As a result, a new secondary reference has been added to address the above limitations. Gligorijevic in view of Huang has been applied to the independent claims.
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.
3. Claims 1 and 7 are rejected under 35 U.S.C. 103 as being unpatentable over Gligorijevic et al. (US20240087674 filed 11/17/2023) in view of Huang et al. ("Coupled graph ode for learning interacting system dynamics." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, KDD ’21, August 14–18, 2021, Virtual Event, Singapore).
Regarding claim 1, Gligorijevic teaches a deep learning system (In some variations, the transformer deep learning model may generate the length transformed embedding [0031])
with a latent transformer core (The length transformer 120 [0117], Fig. 1B, 9A) and
a latent dynamics analyzer (The length predictor 124 [0118], Fig. 1B; The length predictor 124 may be implemented in a variety of ways. FIG. 9A depicts a block diagram illustrating one example of the length predictor 124 [0129]; In some variations, the length predictor may further include a neural network configured to determine a categorical distribution of possible length changes between the first sequence of residues and the second sequence of residues [0054]. The Examiner notes length predictor 124 is a latent dynamics analyzer), comprising
one or more computers with executable instructions that, when executed, cause the deep learning system to (In some example embodiments, there is provided a system that includes at least one processor and at least one memory. The at least one memory may include program code that provides operations when executed by the at least one processor [0004]):
receive a plurality of input vectors (The input sequence 108 [0108], Fig. 1B; input protein sequence, x=(x1, x2, . . . , xL), wherein each token xt is an item from a finite vocabulary V [0101]);
generate a plurality of latent space vectors by processing the plurality of input vectors through a variational autoencoder' s encoder (The encoder 116 can determine a mapping from a protein sequence space (e.g., corresponding to the features of the protein data structures, such as the input sequence 108 or modified protein data structures) to a latent space (e.g., a position or set of positions on the manifold) [0115]);
replicate the latent space vectors (hidden embedding 820, Fig. 9A. The Examiner notes hidden embedding 820 replicates the latent space vectors)
to create two latent space vector copies (a first copy is received by length transformer 120 and a second copy is received by length predictor 124, Fig. 9A);
process a first copy of the latent space vectors through a latent transformer to generate predictions (a first copy is received by length transformer 120 and the output of length transformer 120 are predictions, Fig. 9A);
process a second copy of the latent space vectors through the latent dynamics analyzer (a second copy is received by length predictor 124, Fig. 9A; length dL determined by the length predictor 124 [0135])
generate output vectors (output sequence 140 [0121], Fig. 1B)
by passing the plurality of generated predictions through a variational autoencoder' s decoder (output of length transformer 120 are predictions passed through decoder 136, Fig. 1B; the decoder 136 can generate the output sequence 140 by at least decoding the length transformed output Z of the length transformer 120 [0121]); and
Gligorijevic does not explicitly teach process a copy of the latent space vectors through the latent dynamics analyzer that approximates temporal derivatives of the latent space vectors from a sequence of the latent space vectors representing successive time points, wherein the temporal derivatives represent how the latent space vectors change over time; use the temporal derivatives to modify a plurality of parameters for a plurality of attention mechanisms within the latent transformer or to detect when latent space vectors deviate from temporal patterns determined by the temporal derivatives.
Huang teaches process a copy of the latent space vectors through the latent dynamics analyzer that approximates temporal derivatives of the latent space vectors from a sequence of the latent space vectors representing successive time points (The dynamic nature of a multi-agent dynamical system can be captured by a series of first-order ordinary differential equations (ODE), which describes the state evolution for a set of 𝑀 latent de pendent variables over continuous time 𝑡 ∈ R., pg. 706, right col., last para.)
wherein the temporal derivatives represent how the latent space vectors change over time (Therefore, the set of 𝑀 latent state variables 𝑧𝑡 𝑖 are only for nodes, i.e. 𝑀 = 𝑁. In reality, the network structure may change over time, which requires the modeling of latent edge state 𝑧𝑡 𝑖→𝑗 as well. pg. 707, left col., second para.)
the temporal derivatives (In this paper, we propose coupled graph ODE: a novel latent ordinary differential equation (ODE) generative model that learns the coupled dynamics of nodes and edges with a graph neural network (GNN) based ODE in a continuous manner (abstract); To learn a structural representation for each observation over the weighted, directed temporal graph, we propose an attention-based spatial-temporal GNN that attends over the immediate neighbors of a node, pg. 708, right col., second para.) to modify a plurality of parameters for a plurality of attention mechanisms (Model parameters 𝜙 and θ ... Update the parameters 𝜙 and 𝜃 by optimizing ELBO (evidence lower bound) loss in Eq.7, Algorithm 1: Coupled Graph ODE training procedure, pg. 710, left col., first para.) within the latent transformer (The transfer function 𝑓edge2value is also a simple MLP, which transforms latent edge states to a scalar and is then utilized in the node ODE (pg. 714, right col., last para.); The edge ODE function consists of two parts.𝑓𝑒 :R2𝑑→R𝑑 is a mapping function that transforms the concatenation of two nodes to the latent state of their corresponding edge. 𝑓self :R𝑑→R𝑑 accounts for the self-evolution of edges, pg. 709, right col., first para. The Examiner notes Generative Model: Coupled-ODE for nodes and edges in Fig. 3 is the latent transformer) and
to detect when latent space vectors deviate from temporal patterns determined by the temporal derivatives (By introducing temporal edges and stacking multiple layers of GNN, we can capture the influence from historical observations to the current observation, pg. 708, left col., last para. The Examiner notes deviation from historical observations to the current observation is detected)
Since Gligorijevic as primary reference teaches the trained machine learning model samples from the data distribution during inference time [0103] and Huang as secondary reference discloses an inference model (Huang, pg. 712, left col., first para.) then, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Gligorijevic to incorporate the teachings of Huang for the benefit of achieving accurate prediction especially for long-term predictions (Huang, pg. 713, left col., first para.)
Regarding claim 7, claim 7 is similar to claim 1. It is rejected in the same manner and reasoning applying.
4. Claims 2 and 8 are rejected under 35 U.S.C. 103 as being unpatentable over Gligorijevic et al. (US20240087674 filed 11/17/2023) in view of Huang et al. ("Coupled graph ode for learning interacting system dynamics." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, KDD ’21, August 14–18, 2021, Virtual Event, Singapore) and further in view of Luan et al. (Uncovering Closed-form Governing Equations of Nonlinear Dynamics from Videos. arXiv preprint arXiv:2106.04776. 2021 Jun 9.)
Regarding claim 2, Modified Gligorijevic teaches the system of claim 1, Modified Gligorijevic does not explicitly teach wherein the latent dynamics analyzer comprises: a temporal encoding layer; a neural ordinary differential equation (ODE) module; a symbolic regression network; an equation decoder; and a regularization module.
Luan teaches wherein the latent dynamics analyzer (Here, we consider the nonlinear dynamics of moving object whose governing equations can be expressed by d dtxp(t) = f(xp(t)), where xp(t) = (xp(t), yp(t))T denotes the physical states extracted from the autoencoder, pg. 5, last para.) comprises:
a temporal encoding layer (In addition, because the dynamics to be discovered involves temporal evolution, consecutive video snapshots are considered for temporal integration (time marching) of the physical states (pg. 3, last para.); The output layer of the encoder depends on the number of coordinates needed to describe the object location and the size of image. For a 2D physical system with image size of H × H, the latent space has two variables and the activation of output layer has a saturating non-linearity H/2 · tanh (·) + H/2, which leads to the spatial coordinate of the moving object xs with values in [0, H], pg. 5, first para.);
a neural ordinary differential equation (ODE) module (In the designed network, the constraint by the underlying physical law, represented by a set of nonlinear ODEs, is imposed to the physical states xp extracted from autoencoder, pg. 5, last para.);
a symbolic regression network (The closed-form governing ODEs will be distilled by sparse regression as part of the network architecture, pg. 5, last para.);
an equation decoder (Then the forward/backward video frames can be reconstructed via the decoder as ˆIj+q = ψ(T˜(xp(j +q))), which leads to the forward and backward frame reconstructions loss from the temporal integration of the physical states, pg. 6, first para.); and
a regularization module (sparse relaxed regularized regression (SR3) (pg. 6, first para.).
Gligorijevic as primary reference teaches molecular dynamics simulations 127 (e.g., to determine an energy state) [0152] and designing a computational model 115, which may be implemented as one or more machine learning models e.g., autoencoders [0150], similarly, Huang discloses system dynamics in a continuous fashion through two coupled neural ordinary differential equations (conclusion). While Luan as secondary reference teaches simulating the dynamics forward/backward propagation in time with q-steps for the physical states (pg. 6, first para.), then,
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Gligorijevic to incorporate the teachings of Luan for the benefit of a network trained by optimizing the autoencoder reconstruction loss (pg. 17, second para.) and advancing our understanding and prediction of nonlinear dynamics (Luan, abstract)
Regarding claim 8, claim 8 is similar to claim 2. It is rejected in the same manner and reasoning applying.
5. Claims 3 and 9 are rejected under 35 U.S.C. 103 as being unpatentable over Gligorijevic et al. (US20240087674 filed 11/17/2023) in view of Huang et al. ("Coupled graph ode for learning interacting system dynamics." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, KDD ’21, August 14–18, 2021, Virtual Event, Singapore) and further in view of Brdiczka et al. (US20240129601 filed 04/24/2023)
Regarding claim 3, Modified Gligorijevic teaches the system of claim 1, Modified Gligorijevic does not explicitly teach wherein the input vectors may contain a plurality of appended metadata.
Brdiczka teaches wherein the input vectors may contain a plurality of appended metadata (An identifier such as a user number, user name, email address, phone number, etc. may be appended to input 120, associated with input 120 metadata, and the like [0032]; For example, in one implementation, one or more word2vec vectors can provide an initial list of colors related to the input 120 using vector similarity run on all the nouns in the input 120 [0029]).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Gligorijevic to incorporate the teachings of Brdiczka for the benefit of the description of content which may be appended with user identification information such that the user identification information associated with the description of content can be mapped to a user profile (Brdiczka [0091])
Regarding claim 9, claim 9 is similar to claim 3. It is rejected in the same manner and reasoning applying.
6. Claims 4-6 and 10-12 are rejected under 35 U.S.C. 103 as being unpatentable over Gligorijevic et al. (US20240087674 filed 11/17/2023) in view of Huang et al. ("Coupled graph ode for learning interacting system dynamics." Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, KDD ’21, August 14–18, 2021, Virtual Event, Singapore) and further in view of Jiang et al. (Forecasting movements of stock time series based on hidden state guided deep learning approach, Available online 1 March 2023)
Regarding claim 4, Modified Gligorijevic teaches the system of claim 1, Modified Gligorijevic does not explicitly teach wherein the executable instructions further cause the system to generate alerts or signals when substantial changes in the underlying system dynamics are detected.
Jiang teaches wherein the executable instructions further cause the system to generate alerts or signals when substantial changes in the underlying system dynamics are detected (In addition, deep learning models with attention mechanisms such as ALSTM and the proposed HMM-ALSTM achieve significant improvements, which can adaptively mine the important signals in the stock sequence for predicting future price movements (pg. 11, last para.); The multi-task learning method is designed to estimate the stock state modeling and stock prediction in parallel. It is proved that adversarial signals are provided among different tasks to improve the robustness and generalization of the model in multi-task learning, pg. 3, first para.).
Gligorijevic as primary reference teaches encoder stack of a transformer deep learning model that also includes decoder stack coupled to the output of the encoder stack [0131] and training machine learning model samples from the data distribution during inference time [0103], similarly, Huang discloses a variational autoencoder framework with the goal of learning the latent dynamics of the system (pg. 707, right col., section 4 Model). While Jiang as secondary reference teaches an Encoder–Decoder module (section 4.4, pg. 7) and prediction task based on time series data (abstract), then,
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Gligorijevic to incorporate the teachings of Jiang for the benefit of inputting the representation learned by the Encoder and Decoder and utilize the Attention mechanism to capture the timing information (Jiang, pg. 2, last para.)
Regarding claim 5, Modified Gligorijevic teaches the system of claim 1, Modified Gligorijevic does not explicitly teach wherein the input vectors comprise market data, and wherein the system is configured to analyze and predict financial market behavior.
Jiang teaches wherein the input vectors comprise market data (For brevity, the input features for every timestep contain 12 features overall. Different stocks have different opening prices, closing prices, highest prices, and lowest prices, and there is a large gap between the mean and variance of these prices. This is not conducive to the convergence of the trained model. Therefore, it is necessary to normalize these input features before feeding them into our model, pg. 4, second to the last para.), and
wherein the system is configured to analyze and predict financial market behavior (Stock Prediction Module, Fig. 2; The two sequences 𝑃𝑇 and 𝑄𝑇 are concated to get the input features 𝑋𝑇 = {𝑥𝑡 |𝑡 = 1, 2, 3, … , 𝑇 }. We describe the stock prediction problem as Task 1, and Task 2 is able to help the model to be more robust and improve on task 1, pg. 6, last para.).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Gligorijevic to incorporate the teachings of Jiang for the benefit of inputting the representation learned by the Encoder and Decoder and utilize the Attention mechanism to capture the timing information (Jiang, pg. 2, last para.)
Regarding claim 6, Modified Gligorijevic teaches the system of claim 1, Modified Gligorijevic does not explicitly teach wherein the executable instructions further cause the system to perform end-to-end training of the entire system by computing a total loss function.
Jiang teaches wherein the executable instructions further cause the system to perform end-to-end training of the entire system by computing a total loss function (We integrate the loss function 𝐿𝑜𝑠𝑠1 for stock prediction and 𝐿𝑜𝑠𝑠2 for stock state modeling to get the multi-task learning formation, pg. 9, section 4.7. Semi-supervised training procedure)
PNG
media_image1.png
66
242
media_image1.png
Greyscale
.
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Gligorijevic to incorporate the teachings of Jiang for the benefit of inputting the representation learned by the Encoder and Decoder and utilize the Attention mechanism to capture the timing information (Jiang, pg. 2, last para.)
Regarding claim 10, claim 10 is similar to claim 4. It is rejected in the same manner and reasoning applying.
Regarding claim 11, claim 11 is similar to claim 5. It is rejected in the same manner and reasoning applying.
Regarding claim 12, claim 12 is similar to claim 6. It is rejected in the same manner and reasoning applying.
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 extension fee 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 date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MORIAM MOSUNMOLA GODO whose telephone number is (571)272-8670. The examiner can normally be reached Monday-Friday 8am-5pm 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, Michelle T Bechtold can be reached on (571) 431-0762. 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.
/M.G./Examiner, Art Unit 2148
/MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights