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 .
Specification
The disclosure is objected to because of the following informalities: paragraph [019] of the specification as filed states “the system 100 includes…communication interface device(s) (106) or Input/Output (I/O) interface(s) (106) or user interface (106)”, within which reference character 106 is used to refer to several different components. See MPEP 608.01(g), which states in part “The reference characters must be properly applied, no single reference character being used for two different parts”.
Appropriate correction is required.
Claim Objections
Claims 1, 6, and 11 are objected to because of the following informality: feeding the feature vector and previous group of seasonal lags should read “feeding the feature vector and a previous group of seasonal lags”. Appropriate correction is required.
Claim 6 is additionally objected to because of the following informality: the one or more communication interfaces , should read “the one or more communication interfaces,” (remove space between “interfaces” and “,”). Appropriate correction is required.
Prior Art
The following references are used for prior art claim rejections:
Achar and Pachal “Seasonal Encoder-Decoder Architecture for Forecasting”, hereinafter Achar
Qin et al. “A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction”, hereinafter Qin
Flunkert et al. (U.S. Patent No. 10,936,947), hereinafter Flunkert
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.
Claims 1-15 are rejected under 35 U.S.C. 103 as being unpatentable over Achar, in view of Qin, further in view of Flunkert.
Regarding claim 1,
Achar teaches A…method, comprising:
obtaining,…a seasonal time series data for a plurality of sequential time steps, (Achar Pg. 6, Algorithm 1 states that a “Set of sequences T” is used as input, and “each sequence (time-series) T”, therefore each sequence T is a time series)
PNG
media_image1.png
207
437
media_image1.png
Greyscale
wherein the seasonal time series data is divided into a plurality of groups of seasonal lags; ((Achar Pg. 5) “Note the lags shaded in blue and grouped as ‘standard auto-regressive’ lags in Fig. 1. These inputs of category(a) are fed as input to Encoder 0 (GRU units shaded in blue). This is followed by P groups of seasonal lags”, Achar Pg. 4, Fig. 1 shows a sequence, which is a time series, divided into P groups of seasonal lags)
PNG
media_image2.png
508
868
media_image2.png
Greyscale
and training,…a deep neural network architecture using the seasonal time series data, (Achar Pg. 7, Algorithm 2 shows training examples being formed from sequences, which are time series, and being used to build an RNN model M, which is a deep neural network architecture)
PNG
media_image3.png
445
438
media_image3.png
Greyscale
wherein the deep neural network comprises of a) a plurality of encoders,… ((Achar Pg. 5) “we propose to have multiple encoders depending on the order P of the model”) …and c) a decoder comprising a plurality of sequential decoder units, (Achar Pg. 6, Fig. 2 shows a decoder in green consisting of several sequential GRU units)
PNG
media_image4.png
762
888
media_image4.png
Greyscale
generating a plurality of summary vectors ((Achar Pg. 5) “Context vectors obtained at the last time-step of each of these P encoders are appended before feeding further as initial state at the decoder’s first time-step”, each of the P encoders have their own context vector, which each correspond to a summary vector) pertaining to the plurality of groups of seasonal lags using the plurality of encoders, ((Achar Pg. 2) “in our approach, we avoid skip connections and feed the seasonal lags from each cycle into a separate encoder”, therefore each encoder corresponds to a different group of seasonal lags, with each encoder having its own corresponding summary vector)
wherein each of the plurality of summary vectors encodes one or more seasonal correlations ((Achar Abstract) “We propose a novel RNN architecture capturing (stochastic) seasonal correlations intelligently”, (Achar Pg. 2) “in our method, we consider multiple encoders where the kth encoder captures the correlations from exactly k periods behind the current instant…This is also unlike [21], where a convex combination of all states (of the single encoder) is the context vector”, context vectors correspond to summary vectors) among each of the plurality of groups of seasonal lags; ((Achar Pg. 2) “in our approach, we avoid skip connections and feed the seasonal lags from each cycle into a separate encoder”, Achar Pg. 6, Fig. 2 shows that each encoder, each of which is fed its own group of seasonal lags, produces its own context vector, which each correspond to summary vectors)
generating a feature vector by processing the plurality of summary vectors… (Achar Pg. 6, Fig. 2 shows that the context vectors from all the encoders are appended as input g0 to the decoder, an appending of multiple context vectors corresponds to a feature vector)
and feeding the feature vector (Achar Pg. 6, Fig. 2 shows that the appended context vectors are given as inputs to each decoder unit, an appending of multiple context vectors corresponds to a feature vector) and previous group of seasonal lags into a current decoder unit among the plurality of sequential decoder units ((Achar Pg. 5) “That’s why this block of lags is fed as a separate encoder. Its immediately succeeding lags are fed at the decoder end as they can be exactly matched (OR synchronized) with one of the multi-step targets exactly one period ahead. That’s why we see data from time points y(t−S) to y(t−S+k) fed as decoder inputs from the first to the (K+1)th time-step respectively”, Achar Pg. 6, Fig. 2 shows y(t - S) to y(t + K - S) each given as inputs to the sequential GRU units of the decoder) to predict a seasonal time series data at a subsequent time step ((Achar Pg. 5) “Let us start with P = 1 case. To predict for y(t), recall that the one-step endogenous model eqn. (8) needs as input y(t−S) from exactly one cycle behind and Q1 lags preceding it. A (K+1)-step ahead predictive model (eqn. 12 without exogenous inputs) from y(t) to y(t+K) would depend on data from y(t−S) to y(t−S+k) (exactly one cycle behind the prediction instants) and Q1 lags preceding y(t−S)”, predicting from y(t) to y(t + K) is predicting time series data at subsequent time steps)
Qin teaches the following further limitations that Achar does not teach:
wherein the deep neural network comprises of a) a [plurality of] encoder[s], b) a neural network layer and c) a decoder comprising a plurality of sequential decoder units, (Qin Pg. 2, Fig. 1 shows at (b) an attention layer between the encoder and decoder, Achar but not Qin explicitly teaches a plurality of encoders)
PNG
media_image5.png
362
855
media_image5.png
Greyscale
generating a feature vector by processing the plurality of summary vectors ((Qin Pg. 2, Fig. 1 Caption) “(b) The temporal attention system…represents the input information as a weighted sum of the encoder hidden states across all the time steps. The generated context vector ct is then used as an input to the decoder LSTM unit”, the hidden states correspond to a plurality of summary vectors, the context vector of Qin corresponds to a feature vector) and a state of a previous decoder unit ((Qin Pg. 2, Fig. 1 Caption) “(b) The temporal attention system computes the attention weights βtt based on the previous decoder hidden state dt−1”) among the plurality of sequential decoder units (Qin Pg. 2, Fig. 1 shows that the previous decoder hidden state dt−1 used by the temporal attention layer is from one of the plurality of sequential decoder LSTM units) using the neural network layer; (Qin Pg. 2, Fig. 1 shows that the context vector is derived from the temporal attention layer)
At the time of filing, one of ordinary skill in the art would have motivation to combine Achar and Qin by taking the method for training an encoder-decoder architecture to predict seasonal time series data, taught by Achar, and placing a temporal attention neural network layer, which takes a plurality of vector inputs and produces one vector output, between the encoder and decoder, taught by Qin, as Qin teaches (Qin Pg. 6) “The temporal attention mechanism can naturally capture the long-range temporal information of the encoded inputs”, solving the issue as Qin describes as (Qin Pgs. 1-2) “One problem with encoder-decoder networks is that their performance will deteriorate rapidly as the length of input sequence increases…To resolve this issue, the attention-based encoder decoder network employs an attention mechanism to select parts of hidden states across all the timesteps”. Such a combination would be obvious.
Flunkert teaches the following further limitations that neither Achar nor Qin teach explicitly:
A processor implemented method, comprising: ((Flunkert Col. 23, lines 8-12) “In at least some embodiments, a server that implements a portion or all of one or more of the technologies described herein, including the RNN-based forecasting techniques, as well as various components of a machine learning service may include a general-purpose computer system”, (Flunkert Col. 23, line 16) “computing device 9000 includes one or more processors”)
obtaining, via one or more hardware processors, ((Flunkert Col. 23, line 16) “computing device 9000 includes one or more processors”) a seasonal time series data for a plurality of sequential time steps ((Flunkert Col. 3, lines 16-22) “In various embodiments, a large number (e.g., millions) of time series representing observed demands for respective items of a large inventory for relatively long time periods may be used to train the RNN model. In some cases, several years of weekly or daily demand may be used for training the RNN model, and as a result the model may be able to incorporate seasonal variations in demand automatically”)
and training, via the one or more hardware processors, ((Flunkert Col. 23, line 16) “computing device 9000 includes one or more processors”) a deep neural network architecture using the seasonal time series data, ((Flunkert Col. 3, lines 16-22) “In various embodiments, a large number (e.g., millions) of time series representing observed demands for respective items of a large inventory for relatively long time periods may be used to train the RNN model. In some cases, several years of weekly or daily demand may be used for training the RNN model, and as a result the model may be able to incorporate seasonal variations in demand automatically”)
and wherein training the deep neural network architecture ((Flunkert Col. 3, lines 16-19) “a large number (e.g., millions) of time series representing observed demands for respective items of a large inventory for relatively long time periods may be used to train the RNN model”, an RNN is a deep neural network architecture) comprises of iteratively performing ((Flunkert Col. 19, lines 37-38) “Multiple iterations of training and testing may be performed in various embodiments”) for each of the plurality of sequential decoder units: ((Flunkert Col. 12, lines 1-5) “feeding in the previous sample versus the observed demand during training may be important in order to generate correlated samples and in order for the decoder to perform the same task during training as during prediction”, Flunkert Fig. 3 shows that the decoder consists of sequentially arranged units)
PNG
media_image6.png
698
1115
media_image6.png
Greyscale
At the time of filing, one of ordinary skill in the art would have motivation to combine Achar, Qin, and Flunkert by taking the method for training an encoder-decoder architecture with an attention layer to predict seasonal time series data, jointly taught by Achar and Qin, and implementing the method using at least one processor and iteratively, taught by Flunkert, as both are well-known in the art as providing the predictable benefit of practical implementation of methods for training machine learning models. Such a combination would be obvious.
Regarding claim 2,
Achar, Qin, and Flunkert jointly teach The method of claim 1,
Achar further teaches:
wherein each of the plurality of encoders comprises of a plurality of sequential encoder units, (Achar Pg. 6, Fig. 2 shows that each of the P encoders comprise a sequence of GRU units)
and wherein number of sequential encoder units is determined by an order of seasonality of the seasonal time series data ((Achar Pg. 5) “The first group in the rearrangement of eqn. (12) comes from standard immediate consecutive lags of order p”, Achar Pg. 6, Fig. 2 shows that the first GRU unit in the encoder receives inputs y(t - p) and x(t - p), the second GRU unit in the encoder receives inputs y(t - p - 1) and y(t - p - 1), all the way until the final GRU unit in the encoder receives inputs y(t - 1) and x(t - 1). This means that the number of GRU units in the encoder goes from 1 (last) to p (first), and p is the order of seasonality)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Achar, Qin, Flunkert for the parent claim of claim 2, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 3,
Achar, Qin, and Flunkert jointly teach The method of claim 1,
Qin further teaches:
wherein the neural network layer is one of: (i) an attention layer and (ii) a Recurrent Neural Network (RNN) layer (Qin Pg. 2, Fig. 1 shows at (b) an attention layer between the encoder and decoder)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Achar, Qin, Flunkert for the parent claim of claim 3, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 4,
Achar, Qin, and Flunkert jointly teach The method of claim 1,
Achar further teaches:
wherein the state of the previous decoder unit among the plurality of sequential decoder units ((Qin Pg. 2, Fig. 1 Caption) “(b) The temporal attention system computes the attention weights βtt based on the previous decoder hidden state dt−1”) facilitates encoding contextual information in the feature vector ((Qin Pg. 3) “The attention weight βit represents the importance of the i-th encoder hidden state for the prediction. Since each encoder hidden state hi is mapped to a temporal component of the input, the attention mechanism computes the context vector ct as a weighted sum of all the encoder hidden states {h1, h2, …, hT}”, the context vector of Qin corresponds to a feature vector)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Achar, Qin, Flunkert for the parent claim of claim 4, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 5,
Achar, Qin, and Flunkert jointly teach The method of claim 1,
Qin further teaches:
wherein processing the plurality of summary vectors and a state of the previous decoder unit among the plurality of sequential decoder units using the neural network layer eliminates dependency of the plurality of summary vectors on the order of seasonality (The specification clarifies at [004] “The outputs from the parallel encoders are appended to form a single vector which is then fed into a decoder. However, this results in a large dimensional combined vector dependent on the order of seasonality”. Qin Pg. 2, Fig. 1 shows that the vector outputs, corresponding to summary vectors, of each encoder unit are fed as input to the neural network temporal attention layer, which takes a previous decoder unit state as additional input, then ultimately summed to produce Qin’s context vector. The combination of Qin’s attention layer with Achar’s order of seasonality-dependent summary vectors would eliminate dependency on order of seasonality by summing instead of appending the vectors)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Achar, Qin, Flunkert for the parent claim of claim 5, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claims 6-10,
Claims 6-10 recite a system comprising memory storing instructions, one or more communication interfaces, and one or more processors for performing the function of the method of claims 1-5, respectively. Specifically, claim 6 recites A system, comprising: a memory storing instructions; one or more communication interfaces; and one or more hardware processors coupled to the memory via the one or more communication interfaces , wherein the one or more hardware processors are configured by the instructions to: [perform the method of claim 1]. Flunkert Fig. 12 shows a system comprising a memory, multiple communication interfaces, and multiple processors, coupled together.
PNG
media_image7.png
941
611
media_image7.png
Greyscale
All other limitations in claims 6-10 are substantially the same as those in claim 1-5, respectively, therefore the same rationale for rejection applies.
Regarding claims 11-15,
Claims 11-15 recite at least one non-transitory machine-readable storage medium with instructions for performing the function of the method of claims 1-5, respectively. Specifically, claim 11 recites One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause: [the method of claim 1 to be performed]. Flunkert teaches: (Flunkert Col. 24, lines 28-30) “In some embodiments, system memory 9020 may be one embodiment of a computer-accessible medium configured to store program instructions and data as described above”.
All other limitations in claims 11-15 are substantially the same as those in claim 1-5, respectively, therefore the same rationale for rejection applies.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Achar et al. (U.S. Patent Application Publication No. 2022/0327560) teaches a system for time series data prediction based on seasonal lags with a plurality of encoders and a decoder.
Cinar et al. “Position-based Content Attention for Time Series Forecasting with Sequence-to-sequence RNNs” teaches an architecture for forecasting time series based on an encoder, a decoder, attention, and pseudo-periods.
Du et al. “Multivariate time series forecasting via attention-based encoder–decoder framework” teaches a method for time series forecasting with a Bi-LSTM based encoder-decoder with an attention mechanism.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VICTOR A NAULT whose telephone number is (703) 756-5745. The examiner can normally be reached M - F, 12 - 8.
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, Miranda Huang can be reached at (571) 270-7092. 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.
/V.A.N./Examiner, Art Unit 2124
/MIRANDA M HUANG/ Supervisory Patent Examiner, Art Unit 2124
Prosecution Insights