DETAILED ACTION
This action is responsive to the application filed on 12/19/2023. Claims 1-20 are pending and have been examined.
This action is Non-final.
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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C.
120, 121, 365(c), or 386(c) is acknowledged.
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.
Claims 1,5-9, 10-11, 15-19 are rejected under 35 U.S.C. 103 as being unpatentable over NPL reference “PACE: Predictive and Contrastive Embedding for Unsupervised Action Segmentation”, by Wang et. al. (referred herein as Wang) in view of NPL reference “ANOMALY TRANSFORMER: TIME SERIES ANOMALY DETECTION WITH ASSOCIATION DISCREPANCY”, by Xu et. al. (referred herein as Xu) in view of NPL reference “Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data”, by Hsu et. al. (referred herein as Hsu) further in view of US12499217B2, by Cocea et. al. (referred herein as Cocea.)
Regarding claim 1,Wang teaches:
A computing system configured to perform activity detection, the computing system comprising: ([Wang, Abstract] “Action segmentation, inferring temporal positions of human actions in an untrimmed video, is an important prerequisite for various video understanding tasks. Recently, unsupervised action segmentation (UAS) has emerged as a more challenging task due to the unavailability of frame-level annotations.”, wherein the examiner interprets “inferring temporal positions of human actions” to be the same as “activity detection” because they are both directed to a computational process that determines when and where specific activities occur within a sequential temporal data stream, as inferring when human actions occur in video is the same operation as detecting activities from time-series data.)
… process the plurality of input vectors to obtain a sequence of time ordered segments that maintain a time order of the plurality of input vectors; ([Wang, page 1424, sec 3.1] “Given extracted frame-level features F ∈ ℝ^{n×d}, we first project them into hidden representations R ∈ ℝ^{n×h} with a single fully connected layer, where n is the number of video frames, and d, h are feature dimensionalities. We then employ an encoder network for temporal modeling, which is of an auto-regressive transformer architecture”, AND [Wang, page 1424, sec 1] “we divide the whole video into short clips and form clip-level representations by merging predictive embeddings within each clip.”, [Wang, page 1426, sec 4] “The encoder takes as input video sequences of 100 frames (i.e. n = 100) and the sequence is then divided into clips with length s = 5.”, AND [Wang, page 1424, sec 3.1] “Since the predictive embedding prohibits any leftward information flow, M is used here to make our encoder auto-regressive.”, wherein the examiner interprets “extracted frame-level features F ∈ ℝ^{n×d}” to be the same as “the plurality of input vectors” because they are both directed to a set of multi-dimensional feature vectors drawn sequentially from a temporal data source that serve as the input to the system's transformer encoder. The examiner further interprets “divide the whole video into short clips and form clip-level representations by merging predictive embeddings within each clip” to be the same as “obtain a sequence of time ordered segments” because they are both directed to partitioning a temporal sequence of input feature vectors into contiguous ordered groups, where each group represents a distinct temporal portion of the input data stream. Finally, the examiner interprets “Since the predictive embedding prohibits any leftward information flow” to be the same as “maintain a time order of the plurality of input vectors” because they are both directed to a mechanism that enforces strict left-to-right temporal ordering during processing, ensuring that each segment representation is derived only from temporally prior observations and not from future ones.)
Wang does not teach a memory configured to store a plurality of input vectors representative of time-series data; processing circuitry coupled to the memory, and configured to implement an unsupervised machine learning transformer, wherein the unsupervised machine learning transformer is configured to: … encode the sequence of time ordered segments to obtain a single semantic embedding vector that identifies an activity occurring over at least a portion of the time-series data represented by the plurality of input vectors; and output an indication of the activity detected based on the semantic embedding vector.
Xu teaches a memory configured to store a plurality of input vectors representative of time-series data; ([Xu, page 1, sec 1], “we focus on time series anomaly detection under the unsupervised setting”, and [Xu, p. 3] “The observed time series X is denoted by a set of time points {x₁, x₂, · · · , x_N}, where x_t ∈ ℝ^d represents the observation of time t.”, wherein the examiner interprets “a set of time points {x₁, x₂, · · · , x_N}, where x_t ∈ ℝ^d represents the observation of time t” to be the same as “a plurality of input vectors representative of time-series data” because they are both directed to a collection of multi-dimensional numerical vectors that are each indexed by a distinct point in time, together constituting a time-ordered sequence of observations held within the computing system for processing.)
processing circuitry coupled to the memory, and configured to implement an unsupervised machine learning transformer, wherein the unsupervised machine learning transformer is configured to: ([Xu, page 2] “we adapt Transformers (Vaswani et al., 2017) to time series anomaly detection in the unsupervised regime.” AND [Xu, page 18] “Our paper focuses on unsupervised time series anomaly detection. Experimentally, each dataset includes training, validation and testing subsets. Anomalies are only labeled in the testing subset. Thus, we select the hyper-parameters following the Gap Statistic method (Tibshirani et al., 2001) in K-Means. Here is the selection procedure: • After the training phase, we apply the model to the validation subset (without label) and obtain the anomaly scores (Equation 6) of all time points.”, wherein the examiner interprets “Transformers” operating “in the unsupervised regime” and “unsupervised time series anomaly detection. Experimentally, each dataset includes training, validation and testing subsets. Anomalies are only labeled in the testing subset” to be the same as “unsupervised machine learning transformer” because they are both directed to a transformer neural network architecture that is trained to process sequential data without labeled supervision, and which is further configured to perform the operations described below.)
Wang and Xu does not teach encode the sequence of time ordered segments to obtain a single semantic embedding vector that identifies an activity occurring over at least a portion of the time-series data represented by the plurality of input vectors; and output an indication of the activity detected based on the semantic embedding vector.
Hsu teaches encode the sequence of time ordered segments to obtain a single semantic embedding vector that identifies an activity occurring over at least a portion of the time-series data represented by the plurality of input vectors; ([Hsu, page 5, sec 3], “We use recurrent network architectures for encoders that capture the temporal relationship among time steps, and generate a summarized fixed-dimension vector after consuming an entire sub-sequence.”, wherein the examiner interprets “encoders that capture the temporal relationship among time steps” to be the same as “encode the sequence of time ordered segments” because they are both directed to an encoding process that operates over a temporally ordered set of inputs by learning and representing the sequential relationships among those inputs. The examiner further interprets “generate a summarized fixed-dimension vector after consuming an entire sub-sequence” to be the same as “obtain a single semantic embedding vector that identifies an activity occurring over at least a portion of the time-series data represented by the plurality of input vectors” because they are both directed to producing one fixed-dimensional vector that compactly encodes the semantic content of the consumed temporal sub-sequence, which corresponds to the characteristic activity occurring over that portion of the time-series data.)
Wang, Xu, and Hsu do not teach output an indication of the activity detected based on the semantic embedding vector.
Cocea teaches output an indication of the activity detected based on the semantic embedding vector. ([Cocea, [0041]], “the cache scorer 118 can output the corresponding anomaly score indicated in the anomaly score cache 120.” AND [Cocea, [0042]] “The command line embeddings 116 can be numerical values that indicate semantic and/or contextual meaning of the corresponding command line tokens 114. As discussed above, the transformer model 110 can use NLP principles to determine semantic and/or contextual meaning of textual components of command line entries”, wherein the examiner interprets “output the corresponding anomaly score” to be the same as output an indication of the activity detected because they are both directed to outputting a detection result indicating whether the input data corresponds to a detected condition. The examiner further interprets “command line embeddings 116 can be numerical values that indicate semantic and/or contextual meaning” to be the same as semantic embedding vector because they are both directed to learned numerical embedding representations that encode semantic or contextual meaning of input data. The examiner further interprets outputting the anomaly score based on the command line embeddings to be the same as outputting an indication of the activity detected based on the semantic embedding vector because they are both directed to generating a detection output using semantic/contextual embeddings produced from the input data.)
Wang, Xu, Hsu, Cocea, and the instant application are analogous art because they are all directed to computer-implemented machine-learning processing of sequential or time-series data to generate learned representations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify temporal ordering disclosed by Wang to include the anomaly detection process disclosed by Xu. One would be motivated to do so to effectively apply unsupervised transformer-based temporal modeling to time-series data in an unsupervised detection environment, as suggested by Xu ([Xu, page 2] “we adapt Transformers to time series anomaly detection in the unsupervised regime”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to include the fixed dimension vector process disclosed by Hsu. One would be motivated to do so to efficiently represent a consumed temporal sequence or sub-sequence using a compact learned vector for subsequent activity or sequence-level detection, as suggested by Hsu ([Hsu, page 5, sec 3] “We use recurrent network architectures for encoders that capture the temporal relationship among time steps, and generate a summarized fixed-dimension vector after consuming an entire sub-sequence.”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to include the anomaly score disclosed by Cocea. One would be motivated to do so to effectively generate a detection output from transformer-generated embeddings so that the system can indicate whether the processed sequential data includes anomalous or otherwise notable activity, as suggested by Cocea (([Cocea, [0041]], “the cache scorer 118 can output the corresponding anomaly score indicated in the anomaly score cache 120.” AND [Cocea, [0042]] “The command line embeddings 116 can be numerical values that indicate semantic and/or contextual meaning of the corresponding command line tokens 114. As discussed above, the transformer model 110 can use NLP principles to determine semantic and/or contextual meaning of textual components of command line entries”) Claim 11 is analogous to claim 1, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 5, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1).
Hsu further teaches wherein the unsupervised machine learning transformer implements a variational auto encoder. ([Hsu, page 1, Abstract] “We present a factorized hierarchical variational autoencoder, which learns disentangled and interpretable representations from sequential data without supervision. ... VAEs jointly learn an inference model and a generative model, allowing them to infer latent variables from observed data.”, wherein the examiner interprets “factorized hierarchical variational autoencoder” to be the same as variational auto encoder because they are both directed to a variational autoencoder model. The examiner further interprets “learns disentangled and interpretable representations from sequential data without supervision” to be the same as unsupervised machine learning because they are both directed to learning representations from sequential data without supervised labels. The examiner further interprets “VAEs jointly learn an inference model and a generative model” to be the same as implements a variational auto encoder because they are both directed to using a variational autoencoder architecture that learns latent-variable representations from observed data.)
Wang, Xu, Hsu, Cocea, and the instant application are analogous art because they are all directed to unsupervised machine-learning processing of sequential or time-series data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the autoencoder disclosed by Hsu. One would be motivated to do so to effectively learn disentangled and interpretable latent representations from sequential data without supervised labels, as suggested by Hsu ([Hsu, page 1, Abstract] “learns disentangled and interpretable representations from sequential data without supervision”). Claim 15 is analogous to claim 5, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 6, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1).
Hsu further teaches:
wherein the processing circuitry is further configured to perform preprocessing of the time-series data ([Hsu, page 6, sec 4] “All speech is represented as a sequence of 80 dimensional Mel-scale filter bank (FBank) features or 200 dimensional log-magnitude spectrum (only for audio reconstruction), computed every 10ms.”, wherein the examiner interprets “All speech is represented as a sequence of 80 dimensional Mel-scale filter bank (FBank) features or 200 dimensional log-magnitude spectrum” to be the same as perform preprocessing of the time-series data because they are both directed to transforming raw time-dependent data (speech) into numerical feature representations. The examiner further interprets “computed every 10ms” to be the same as conditioning the time-series data because both involve extracting structured features from raw time-based data at regular intervals).
to condition the time-series data during generation of the plurality of input vectors. ([Hsu, page 6, sec 4] “We consider a sample x to be a 200ms sub-sequence, which is on the order of the length of a syllable, and implies T = 20 for each x.”, wherein the examiner interprets “a 200ms sub-sequence” to be the same as time-series data because both refer to sequential portions of temporal data. The examiner further interprets “implies T = 20 for each x” to be the same as generation of the plurality of input vectors because both are directed to dividing time-series data into structured segments or vectorized representations for further model input).
Wang, Xu, Hsu, Cocea, and the instant application are analogous art because they are all directed to processing sequential or time-series data using learned representations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the filter bank disclosed by Hsu. One would be motivated to do so to effectively condition raw time-series data into structured numerical input vectors suitable for neural-network processing, as suggested by Hsu ([Hsu, page 6] “All speech is represented as a sequence of 80 dimensional Mel-scale filter bank (FBank) features or 200 dimensional log-magnitude spectrum (only for audio reconstruction), computed every 10ms.”). Claim 16 is analogous to claim 6, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 7, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1).
Hsu further teaches:
wherein the processing circuitry is further configured to perform semantic enrichment with respect to the time-series data ([Hsu, page 6, sec 4] “All speech is represented as a sequence of 80 dimensional Mel-scale filter bank (FBank) features or 200 dimensional log-magnitude spectrum (only for audio reconstruction), computed every 10ms.” AND [Hsu, page 2, sec 1] “In addition, linguistic prior also encourages gradually merging nearby tokens (words) into larger semantic units (phrases), which naturally leads to a shorter sequence of representations.”, wherein the examiner interprets “speech represented as a sequence of ... features” and “merging nearby tokens (words) into larger semantic units (phrases), which naturally leads to a shorter sequence of representations” to be the same as perform preprocessing of the time-series data because both involve transforming raw time-based input (speech) into structured features. The examiner further interprets “computed every 10ms” to be the same as processing time-series data because both refer to processing sequential time-based data at regular intervals).
to condition the time-series data during generation of the plurality of input vectors. ([Hsu, page 6, sec 4] “We consider a sample x to be a 200ms sub-sequence, which is on the order of the length of a syllable, and implies T = 20 for each x.”, wherein the examiner interprets “a 200ms sub-sequence” to be the same as conditioning the time-series data because both involve segmenting the time-series into meaningful units. The examiner further interprets “T = 20 for each x” to be the same as generation of the plurality of input vectors because both are directed to dividing time.)
Wang, Xu, Hsu, Cocea, and the instant application are analogous art because they are all directed to processing sequential or time-series data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the filter bank disclosed by Hsu. One would be motivated to do so to effectively generate meaningful feature representations from time-series data for machine-learning processing, as suggested by Hsu ([Hsu, page 6, sec 4] “All speech is represented as a sequence of 80 dimensional Mel-scale filter bank (FBank) features or 200 dimensional log-magnitude spectrum (only for audio reconstruction), computed every 10ms.” AND [Hsu, page 2, sec 1] “In addition, linguistic prior also encourages gradually merging nearby tokens (words) into larger semantic units (phrases), which naturally leads to a shorter sequence of representations.”). Claim 17 is analogous to claim 7, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 8, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1).
Wang further teaches:
wherein the processing circuitry is further configured to: perform preprocessing of the time-series data to condition the time-series data to obtain a plurality of preprocessed embedded vectors; ([Wang, page 1425, sec 3.2] “Given extracted frame-level features F ∈ Rn×d, we first project them into hidden representations R ∈ Rn×h with a single fully connected layer”, wherein the examiner interprets “extracted frame-level features F ∈ Rn×d” to be the same as time-series data because they are both directed to temporally ordered feature data. The examiner further interprets “we first project them into hidden representations R ∈ Rn×h with a single fully connected layer” to be the same as perform preprocessing of the time-series data to condition the time-series data to obtain a plurality of preprocessed embedded vectors because they are both directed to transforming temporal feature data into learned hidden/vector representations before further processing).
perform semantic enrichment with respect to the plurality of preprocessed embedded vectors to generate the plurality of input vectors. ([Wang, page 1425, sec 3.2] “we employ a transformer encoder for temporal modeling. In the training stage, predictive embeddings are learned by predicting future representations from the past ones, while contrastive embeddings are generated with contrastive learning on clip-level representations ... contrastive embeddings are based on clip-level representations, which capture more contextual semantics than frame-level representations ... Next, we employ a contrastive projection head Pctrst to transform C′ into contrastive embeddings C ∈ Rt×m with m being the new dimensionality … Fortunately, clustering-based UAS methods have demonstrated that similarity information among visual representations can also be utilized to separate actions with different semantics. Hence, we propose to further exploit similarity information via contrastive learning to circumvent the limitations in predictive embeddings.”, wherein the examiner interprets “predictive embeddings are learned by predicting future representations from the past ones” to be the same as the plurality of preprocessed embedded vectors because they are both directed to learned embedded vector representations derived from temporal input data. The examiner further interprets “contrastive embeddings are based on clip-level representations, which capture more contextual semantics than frame-level representations” to be the same as perform semantic enrichment with respect to the plurality of preprocessed embedded vectors because they are both directed to generating embeddings that add contextual semantic information beyond the earlier frame-level or predictive embedded representations. The examiner further interprets “transform C′ into contrastive embeddings C ∈ Rt×m” and “similarity information via contrastive learning to circumvent the limitations in predictive embeddings” to be the same as generate the plurality of input vectors because they are both directed to producing a plurality of vector embeddings from contextually processed temporal representations.) Claim 18 is analogous to claim 8, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 9, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1).
Wang further teaches:
Wherein the processing circuitry to: perform activity detection with respect to the single semantic embedding vector to identify the activity; ([Wang, page 1424, sec 3] “In the test stage, we fuse prediction errors and embedding similarities for action boundary detection.” ... [Wang, page 3] “Embedding Similarity. To infer the embedding similarities for the entire video, we utilize a sliding window of 2s frames.” ... [Wang, page 1426, sec 3.3] “Representations within each window can form two adjacent contrastive embeddings…After the fusion, Vi indicates the potential of frame i being an action boundary.”, wherein the examiner interprets “embedding similarities” and “contrastive embeddings” to be the same as the single semantic embedding vector because they are both directed to learned embedding representations used to capture semantic information from temporal video/activity data. The examiner further interprets “action boundary detection” to be the same as activity detection because they are both directed to detecting actions or activities within temporal data. The examiner further interprets “Vi indicates the potential of frame i being an action boundary” to be the same as to identify the activity because they are both directed to identifying portions of temporal data corresponding to action/activity boundaries or segments).
output an indication of the activity. ([Wang, page 1425, sec 3.2] “Specifically, let Ro denote the output representations from the encoder.” AND [Wang, page 1426, sec 3.3] “Action Boundary Detection. After the fusion, Vi indicates the potential of frame i being an action boundary. We then employ an action boundary detection method to generate the final action segmentation results.”, wherein the examiner interprets “generate the final action segmentation results” to be the same as output an indication of the activity because they are both directed to producing an output result indicating detected action/activity segments in the temporal data). Claim 19 is analogous to claim 9, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 10, Wang, Xu, Hsu, and Cocea teach The computing system of claim 9, (see the rejection of claim 9).
Cocea further teaches:
wherein the activity detection includes anomaly detection with respect to the single semantic embedding vector. ([Cocea, [0049]] “The anomaly detection model 112 can be configured to use command line embeddings 116 associated with a command line entry to generate a corresponding anomaly score for the command line entry.”, wherein the examiner interprets “anomaly detection model 112” that is configured to “generate a corresponding anomaly score” to be the same as activity detection includes anomaly detection because they are both directed to detecting an anomalous condition and generating an anomaly-related detection output. Furthermore, the examiner interprets “use command line embeddings 116 associated with a command line entry to generate a corresponding anomaly score” to be the same as anomaly detection with respect to the single semantic embedding vector because they are both directed to performing anomaly detection using an embedding vector generated from the input data).
Wang, Xu, Hsu, Cocea, and the instant application are analogous art because they are all directed to machine-learning-based detection using learned embeddings generated from sequential or time-based input data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 9 disclosed by Wang, Xu, Hsu, and Cocea to include the “anomaly detection model 112” disclosed by Cocea. One would be motivated to do so to effectively identify anomalous activity or behavior using embeddings generated from input data, as suggested by Cocea ([Cocea, [0049]] “The anomaly detection model 112 can be configured to use command line embeddings 116 associated with a command line entry to generate a corresponding anomaly score for the command line entry.”).
Claims 2-4, and 12-14 are rejected under 35 U.S.C. 103 as being unpatentable over Wang in view Xu in view of Hsu in view of Cocea further in view of NPL reference “Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing”, by Dai et. al. (referred herein as Dai.).
Regarding claim 2, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1)
Wang further teaches wherein the hierarchical framing transformer includes a feed forward neural network having multiple attention heads, ([Wang, page 1424, sec 3.1] “Let SA, FFN, LN denote multi-head self-attention, position-wise feed-forward network and layer normalization, respectively. The basic encoder layer can be formulated as: R_l' = LN(SA(R_{l-1}, M) + R_{l-1}), R_l = LN(FFN(R_l') + R_l')”, wherein the examiner interprets “multi-head self-attention” to be the same as “multiple attention heads” because they are both directed to an attention mechanism that employs multiple parallel attention computations over the same input sequence; and further interprets “position-wise feed-forward network” to be the same as “feed forward neural network” because they are both directed to a feedforward layer composed of learned linear transformations and nonlinearities applied to each position in the sequence to produce updated hidden representations.)
wherein the feed forward neural network is trained via unsupervised learning. ([Wang, page 1424, sec 2] “Recent works start to focus on training action segmentation models without any supervision signals. It is a much more challenging problem since no information can be directly utilized to optimize frame-wise classification models.”, wherein the examiner interprets “training action segmentation models without any supervision signals” to be the same as “trained via unsupervised learning” because they are both directed to a model training process in which no labeled ground truth annotations are provided, and the model instead learns directly from the structure and patterns inherent in the unlabeled input data.)
Wang, Xu, Hsu, and Cocea do not teach wherein the unsupervised machine learning transformer includes a hierarchical framing transformer.
Dai teaches wherein the unsupervised machine learning transformer includes a hierarchical framing transformer, ([Dai, page 3, sec 2.2] “our model employs an encoder that gradually reduces the sequence length of the hidden states as the layer gets deeper.” AND ([Dai, page 2, sec 1] “linguistic prior also encourages gradually merging nearby tokens (words) into larger semantic units (phrases), which naturally leads to a shorter sequence of representations.” AND [Dai, page 1, Abstract] “Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost.”)., wherein the examiner interprets “an encoder that gradually reduces the sequence length of the hidden states as the layer gets deeper” to be the same as “hierarchical framing transformer” because they are both directed to a transformer encoder that operates across multiple successive levels, with each deeper level progressively compressing the input sequence into a shorter, higher-abstraction representation, which constitutes the hierarchical framing of the input data stream; and further interprets “gradually merging nearby tokens (words) into larger semantic units (phrases)” to be the same as “hierarchical framing” because they are both directed to aggregating sequences of lower-level inputs into larger meaningful units at successive levels of the hierarchy, which is the defining characteristic of hierarchical framing.)
Wang, Xu, Hsu, Cocea, Dai, and the instant application are analogous art because they are all directed to transformer-based processing of ordered sequence data using attention and feed-forward neural-networks.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the Transformer disclosed by Dai. One would be motivated to do so to efficiently reduce sequential redundancy and form higher-level semantic representations of the input sequence, as suggested by Dai ([Dai, page 1, Abstract] “Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost.”). Claim 12 is analogous to claim 2, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 3, Wang, Xu, Hsu, Cocea, and Dai teach The computing system of claim 2, (see the rejection of claim 1)
Wang further teaches wherein the feed forward neural network includes multiple layers, ([Wang, p. 1426, sec 4] “The encoder has 3 basic layers in total. There are 4 attention heads in SA, and the inner-layer of FFN is 2,048-d.”, wherein the examiner interprets “3 basic layers in total” to be the same as “multiple layers” because they are both directed to a neural network encoder composed of more than one successive processing stage, with each layer building upon the output representations of the previous one to produce increasingly refined features.)
Hsu further teaches each of the multiple layers generating a separate sub-semantic embedding vector ([Hsu, page 2, sec 2] “We refer to the first type of attributes as sequence-level attributes, and the other as segment-level attributes. In this work, we achieve disentanglement and interpretability by encoding the two types of attributes into latent sequence variables and latent segment variables respectively, where the former is regularized by an sequence-dependent prior and the latter by a sequence-independent prior.”, wherein the examiner interprets encoding “the two types of attributes into latent sequence variables and latent segment variables respectively” at two distinct levels of the hierarchy to be the same as “each of the multiple layers generating a separate sub-semantic embedding vector” because they are both directed to a hierarchical model where each level of the architecture produces its own distinct latent representation, which is that the segment-level encoder produces latent segment variables and the sequence-level encoder produces latent sequence variables, each being a separate semantic embedding vector generated by a different layer of the hierarchy.)
for each of the sequence of time ordered segments ([Hsu, page 3 sec 2] “N i.i.d. latent sequence variables {z_2^{(n)}}{n=1}^N and latent segment variables {z_1^{(n)}}{n=1}^N are drawn from a sequence-dependent prior distribution p_θ(z_2|μ_2) and a sequence-independent prior distribution p_θ(z_1) respectively”, wherein the examiner interprets “latent segment variables {z_1^{(n)}}_{n=1}^N”, N latent vectors individually indexed by segment position n, to be the same as “for each of the sequence of time ordered segments” because they are both directed to producing one distinct embedding vector per temporal segment, where n = 1 through N indexes each consecutive segment across the full time-ordered sequence.)
that identifies a sub-activity performed during the overarching activity. ([Hsu, page 1, sec 1] “The information encoded in sequential data, such as speech, video, and text, is naturally multi-scaled; in speech for example, information about the channel, speaker, and linguistic content is encoded in the statistics at the session, utterance, and segment levels, respectively… learns disentangled and interpretable representations from sequential data without supervision.”, wherein the examiner interprets the encoding of linguistic content at the segment level - the locally-varying attribute of each individual temporal sub-sequence within the broader context of speaker information captured at the utterance level - to be the same as “sub-activity performed during the overarching activity” because they are both directed to capturing the locally-occurring characteristic of each individual temporal segment (the sub-activity, corresponding here to the linguistic content specific to that segment) within the broader temporal activity captured at the sequence level (the overarching activity, corresponding here to the speaker's identity sustained across the full utterance)).
Wang, Xu, Hsu, Cocea, Dai, and the instant application are analogous art because they are all directed to hierarchical representation learning for sequential or temporal data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 2 disclosed by Wang, Xu, Hsu, Cocea, and Dai to include the encoding technique disclosed by Hsu. One would be motivated to do so to effectively learn disentangled sequence-level and segment-level representations from sequential data, thereby allowing the model to separately represent information associated with the broader sequence and information associated with individual segments of the sequence, as suggested by Hsu ([Hsu, page 1, sec 1] “learns disentangled and interpretable representations from sequential data without supervision.”) Claim 13 is analogous to claim 3, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 4, Wang, Xu, Hsu, and Cocea teach The computing system of claim 1, (see the rejection of claim 1)
Xu further teaches based on the plurality of reconstructed input vectors and the plurality of input vectors, ([Xu, page 5], “The loss function for input series X ∈ RN×d is formalized as: LTotal(X̂, P, S, λ; X) = ||X - X̂||2F - λ × ||AssDis(P, S; X)||1”, wherein the examiner interprets “X” to be the same as the plurality of input vectors and “X̂” to be the same as the plurality of reconstructed input vectors because they are both directed to an original input series and a corresponding reconstructed version of the input series used in a reconstruction loss. The examiner further interprets “||X - X̂||2F” to be the same as based on the plurality of reconstructed input vectors and the plurality of input vectors because they are both directed to comparing reconstructed input data with original input data during training.)
Hsu further teaches to adjust one or more weights applied to a plurality of subsequent input vectors when obtaining a subsequent single semantic embedding vector. ([Hsu, page 5] “During testing, we may want to use the s-vector μ2 of an unseen sequence X˜ = {x˜(n)}N˜n=1 as the sequence-level attribute representation” ... “where LSTM refers to a long short-term memory recurrent neural network [14], and MLP refers to a multi-layer perceptron, φ∗ are the related weight matrices.”, wherein the examiner interprets “φ∗ are the related weight matrices” to be the same as one or more weights because they are both directed to trainable neural-network weight parameters. The examiner further interprets “unseen sequence X˜ = {x˜(n)}N˜n=1” to be the same as a plurality of subsequent input vectors because they are both directed to later-processed sequence data input after model training. The examiner further interprets “s-vector μ2 ... as the sequence-level attribute representation” to be the same as a subsequent single semantic embedding vector because they are both directed to a single learned vector representation obtained for a later-processed sequence. The examiner further interprets “φ∗ are the related weight matrices” to be the same as weights applied to a plurality of subsequent input vectors when obtaining a subsequent single semantic embedding vector because they are both directed to trainable neural-network weight parameters applied during processing of later-received sequence input data to obtain a sequence-level vector representation.)
Wang, Xu, Hsu, and Cocea do not teach wherein the unsupervised machine learning transformer includes an unsupervised machine learning model that acts as a decoder and is configured to decode the semantic embedding vector to reconstruct the plurality of input vectors and obtain a plurality of reconstructed input vectors, and wherein the unsupervised machine learning transformer performs unsupervised learning.
Dai teaches:
wherein the unsupervised machine learning transformer includes an unsupervised machine learning model that acts as a decoder and is configured to decode the semantic embedding vector to reconstruct the plurality of input vectors and obtain a plurality of reconstructed input vectors, ([Dai, page 3, sec 2.2] “In addition, for tasks involving per-token predictions like pretraining, a simple decoder is used to reconstruct a full sequence of token-level representations from the compressed encoder output.” AND [Dai, page 2, sec 1] “In addition, linguistic prior also encourages gradually merging nearby tokens (words) into larger semantic units (phrases), which naturally leads to a shorter sequence of representations.”, wherein the examiner interprets “a simple decoder” to be the same as a machine learning model that acts as a decoder because they are both directed to a decoder component used in a transformer-based model to generate reconstructed sequence representations from an encoded representation, and Dai’s decoder as included in the unsupervised machine learning transformer because Dai teaches that the decoder is part of the proposed Funnel-Transformer architecture used during pretraining.)
and wherein the unsupervised machine learning transformer performs unsupervised learning, ([Dai, page 2, sec 2] “For a length-T natural language sequence x sample from a large unlabeled set D, the MLM objective first constructs a corrupted sequence xˆ by randomly replacing 15% of the tokens of x with a special token [mask] and then trains a Transformer model [2] to reconstruct the original x based on xˆ”, wherein the examiner interprets “a large unlabeled set D” to be the same as unsupervised learning because they are both directed to training using unlabeled data, and wherein the examiner interprets “trains a Transformer model [2] to reconstruct the original x based on xˆ” to be the same as the unsupervised machine learning transformer performs unsupervised learning because they are both directed to training a transformer model through reconstruction of original input data).
Wang, Xu, Hsu, Cocea, Dai, and the instant application are analogous art because they are all directed to unsupervised neural-network processing of sequential or time-series data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to further modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the use of trainable neural-network weight matrices as disclosed by Hsu. One would be motivated to do so to effectively obtain a sequence-level representation for subsequently received or unseen sequential input data using trained neural-network parameters, as suggested by Hsu ([Hsu, page 5, sec 2.2] “During testing, we may want to use the s-vector μ2 of an unseen sequence X˜ = {x˜(n)}N˜n=1 as the sequence-level attribute representation” ... “where LSTM refers to a long short-term memory recurrent neural network [14], and MLP refers to a multi-layer perceptron, φ∗ are the related weight matrices.”)
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computing system of claim 1 disclosed by Wang, Xu, Hsu, and Cocea to include the “simple decoder” disclosed by Dai. One would be motivated to do so to efficiently reconstruct full sequence representations from compressed transformer representations during unsupervised pretraining, as suggested by Dai ([Dai, page 3, sec2.2] “for tasks involving per-token predictions like pretraining, a simple decoder is used to reconstruct a full sequence of token-level representations from the compressed encoder output.”). Claim 14 is analogous to claim 4, aside from claim type and minute differences, hence the same rejection can apply.
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Cocea in view of Xu in view of Wang further in view of Dai.
Regarding claim 20, Cocea teaches:
A non-transitory computer-readable storage medium having instructions stored thereon that, when executed, cause one or more processors to: ([Cocea, [0083]] “The memory 504 can further include non-transitory computer-readable media, such as volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data... [Cocea, [0085]] “the processors 510 can access data and computer-executable instructions stored in the memory 504, and execute such computer-executable instructions... The machine readable medium 520 can store one or more sets of instructions, such as software or firmware, that embodies any one or more of the methodologies or functions described herein”, wherein the examiner interprets the non-transitory computer-readable media storing computer-readable instructions, and processors executing computer-executable instructions stored in memory, to be the same as a non-transitory computer-readable storage medium having instructions stored thereon that, when executed, cause one or more processors to because they are both directed to stored instructions on computer-readable media that are executed by processors to perform the disclosed operations).
invoke an unsupervised machine learning transformer that: ([Cocea, Abstract] “the first machine learning model of the transformer model is trained, via unsupervised machine learning”, wherein the examiner interprets execution of the transformer model to be the same as invoking a machine learning transformer because they are both directed to causing a transformer model to run on a computing system)., wherein the examiner interprets the transformer model trained via unsupervised machine learning to be the same as an unsupervised machine learning transformer because they are both directed to a transformer machine-learning model trained using unsupervised machine learning).
Cocea does not teach processes a plurality of input vectors representative of time-series data to obtain a sequence of time ordered segments that maintain time order of the plurality of input vectors; encodes the sequence of time ordered segments to obtain a single semantic embedding vector that identifies an overarching activity occurring over at least a portion of the time-series data represented by the plurality of input vectors; and outputs the semantic embedding vector.
Xu teaches processes a plurality of input vectors representative of time-series data ([Xu, page 3, sec 3] “The observed time series X is denoted by a set of time points {x1, x2, · · · , xN}, where xt ∈ Rd represents the observation of time t”, wherein the examiner interprets “a set of time points {x1, x2, · · · , xN}, where xt ∈ Rd represents the observation of time t” to be the same as a plurality of input vectors representative of time-series data because they are both directed to multiple vector-valued observations arranged over time).
Cocea and Xu does not teach to obtain a sequence of time ordered segments that maintain time order of the plurality of input vectors; … encodes the sequence of time ordered segments to obtain a single semantic embedding vector that identifies an overarching activity occurring over at least a portion of the time-series data represented by the plurality of input vectors;
Wang teaches:
to obtain a sequence of time ordered segments that maintain time order of the plurality of input vectors; ([Wang, page 1424, sec 1] “we divide the whole video into short clips and form clip-level representations by merging predictive embeddings within each clip... by pulling temporally adjacent clips closer and pushing nonadjacent ones farther”, AND [Wang, page 1425, sec 3.2] “we merge features within each consecutive clip of s frames, resulting in clip-level representations C′ ∈ Rt×d, where t = n/s indicates the total number of clips”, wherein the examiner interprets “short clips,” “clip-level representations,” and “each consecutive clip of s frames” to be the same as a sequence of time ordered segments because they are both directed to dividing temporal input data into ordered clip-level or segment-level portions. The examiner further interprets “temporally adjacent clips,” “nonadjacent ones,” and “each consecutive clip of s frames” to be the same as maintaining time order of the plurality of input vectors because they are both directed to preserving or using the temporal ordering and adjacency relationships among portions of the input sequence.)
that identifies an overarching activity occurring over at least a portion of the time-series data represented by the plurality of input vectors; ([Wang, page 1423] “Action segmentation, inferring temporal positions of human actions in an untrimmed video, is an important prerequisite for various video understanding tasks... Aiming to classify video frames into predefined action categories, action segmentation is an important step towards fine-grained human action understanding”, wherein the examiner interprets “inferring temporal positions of human actions in an untrimmed video” and “classify video frames into predefined action categories” to be the same as identifies an overarching activity occurring over at least a portion of the time-series data because they are both directed to identifying human actions or activity categories occurring within temporal portions of input video/frame data).
Cocea, Xu, and Wang does not teach encodes the sequence of time ordered segments to obtain a single semantic embedding vector … outputs the semantic embedding vector.
Dai teaches encodes the sequence of time ordered segments to obtain a single semantic embedding vector ([Dai, page 3, sec 2.2] “our model employs an encoder that gradually reduces the sequence length of the hidden states as the layer gets deeper”, AND [Dai, page 4, sec 2.2] “the linguistic prior that nearby tokens could be gradually merged (or compressed) into a larger semantic component.”, AND [Dai, page 3, sec 2.1] “many sequence-level downstream tasks like classification or ranking only need a single-vector summary of the entire sequence”, wherein the examiner interprets “our model employs an encoder that gradually reduces the sequence length of the hidden states as the layer gets deeper” to be the same as encodes the sequence of time ordered segments because they are both directed to using an encoder to process an ordered sequence into progressively reduced sequence representations. The examiner further interprets “nearby tokens could be gradually merged (or compressed) into a larger semantic component” to be the same as semantic because they are both directed to forming higher-level semantic representations from portions of an ordered sequence. The examiner further interprets “a single-vector summary of the entire sequence” to be the same as a single semantic embedding vector because they are both directed to representing the sequence using one learned vector that summarizes the semantic content of the sequence.)
outputs the semantic embedding vector. ([Dai, page 3, sec 2.2] “In addition, for tasks involving per-token predictions like pretraining, a simple decoder is used to reconstruct a full sequence of token-level representations from the compressed encoder output”, AND [Dai, page 4, sec 2.2] “Specifically, given the output sequence hM of length TM = T/2M-1 from an M-block encoder, we directly up-sample it to a full-length sequence hup = hup 1 ,··· ,hup T by repeating each hidden vector 2M-1 times:”, wherein the examiner interprets simple decoder is used to reconstruct a full sequence of token-level representations from the compressed encoder output to be the same as outputs the semantic embedding vector because they are both directed to producing a learned sequence-level vector representation from the transformer encoder output for downstream use).
Cocea, Xu, Wang, Dai, and the instant application are analogous art because they are all directed to computer-implemented machine-learning processing of sequential or time-series data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the non-transitory computer-readable storage medium and invoking of unsupervised learning disclosed by Cocea to include the time series processing disclosed by Xu. One would be motivated to do so to effectively process vector-valued time-series observations as suggested by Xu ([Xu, page 3, sec 3] “The observed time series X is denoted by a set of time points {x1, x2, · · · , xN}, where xt ∈ Rd represents the observation of time t”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to further include the representations disclosed by Wang. One would be motivated to do so to effectively preserve temporal adjacency and generate temporally consistent segment-level representations of actions in time-series/video data, as suggested by Wang ([Wang, page 1424, sec 1] “we divide the whole video into short clips and form clip-level representations by merging predictive embeddings within each clip... by pulling temporally adjacent clips closer and pushing nonadjacent ones farther”, AND [Wang, page 1425, sec 3.2] “we merge features within each consecutive clip of s frames, resulting in clip-level representations C′ ∈ Rt×d, where t = n/s indicates the total number of clips”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to further include the sequence summary disclosed by Dai. One would be motivated to do so to efficiently produce a compact sequence-leve representation for downstream use while reducing sequential redundancy, as suggested by Dai ([Dai, page 3, sec 2.2] “In addition, for tasks involving per-token predictions like pretraining, a simple decoder is used to reconstruct a full sequence of token-level representations from the compressed encoder output”, AND [Dai, page 4, sec 2.2] “Specifically, given the output sequence hM of length TM = T/2M-1 from an M-block encoder, we directly up-sample it to a full-length sequence hup = hup 1 ,··· ,hup T by repeating each hidden vector 2M-1 times:”)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DEVAN KAPOOR whose telephone number is (703)756-1434. The examiner can normally be reached Monday - Friday: 9:00AM - 5:00 PM EST (times may vary).
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, David Yi can be reached at (571) 270-7519. 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.
/DEVAN KAPOOR/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights