DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . This action is responsive to pending claims 1-7, 9-15, 17-19 filed 6/8/2026.
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.
Claim(s) 1-16 are rejected under 35 U.S.C. 103 as being unpatentable over Ferdinand ("Attenuating catastrophic forgetting by joint contrastive and incremental learning", published 8/23/2022) in view of He ("Momentum contrast for unsupervised visual representation learning", published 2020).
For claim 1, Ferdinand discloses: a computer-implemented method for training a classification model (fig.1 gives overview of model training, with §5, §5.1 disclosing various computer vision tasks), the computer-implemented method comprising:
obtaining a labeled data set of images comprising data from a rehearsal memory and new input data (fig.1, “augmented minibatch”; §4.2 ¶1: augmented image minibatches are obtained; §4.1 ¶1, §5.2 ¶2: rehearsal memory and implementation);
augmenting the labeled data set to generate a first data set and a second data set, wherein each image of the first data set corresponds to a corresponding image of the second data set (ibid: contrastive and complementary image sets are obtained);
inputting the first data set into a query encoder and inputting the second data set into a key encoder to obtain encodings output by the query encoder and the key encoder (fig.1: augmented input is passed to identically structured query student and key teacher encoders, each of which generate respective projection (gamma) and feature (F) outputs, see §4.2 ¶1-2);
obtaining a contrastive loss based on the encodings using a sum of a first contrastive loss function and a second contrastive loss function (§4 gives overview of total loss as sum of contrastive losses, see also fig.1 outlining contrastive losses generated from supervised data on the query network (L_con) and contrastive distillation loss (L_Dcon)), wherein the first contrastive loss function uses encodings of different views of a same input image as anchor-positive pairs in a feature space (§4.2 ¶1, eq.6: contrastive loss is calculated in the inner sum, in the numerator, by taking the pairwise dot product of the projections of all elements in the same class, which would include all the augmentation or different views of the input image, hence, the first encodings of the different views for anchor positive pairs), and the second contrastive loss function uses, as anchor-positive pairs in the feature space, encodings of two different samples from a same class (§4.3, eq.7-8: L_Dcon is minimizes distance between B^t-1 and B^t (eq.8), each of which incorporate dot products of projections of base and augmented images z and z-hat, hence, anchor positive pairs are formed via encoding projections from different samples of the same class, the set of images from the past classes, see §4.3 ¶3), the two different samples being identified using the corresponding labels (§4.3: the samples being identified based on associated y labels);
updating parameters of the query encoder based on the obtained contrastive loss (§4: updating parameters based on losses); and
updating parameters of the key encoder based on parameters of the query encoder (§2.1 ¶2: transferring parameters from query to key at each incremental step, with §5.1 ¶2 disclosing 250 epochs for each step).
Ferdinand does not disclose: wherein the key encoder is a momentum encoder;
wherein the updating of the key encoder is as an exponentially weighted moving average of the parameters of the query encoder;
maintaining a first-in-first-out (FIFO) queue storing encodings of data samples from previous mini-batches and corresponding labels of the data samples, the FIFO queue decoupling a number of encodings available for contrastive learning from a mini-batch size;
the samples being stored in the FIFO queue.
He discloses: wherein the key encoder is a momentum encoder (§3.2 “Momentum update” gives overview of a momentum encoder with exponentially weighted moving average per iteration, see eq.(2));
wherein the updating of the key encoder is as an exponentially weighted moving average of the parameters of the query encoder (§3.2 “Momentum update”, hence, combination with teacher model technique of Ferdinand yielding use of a momentum encoder as the teacher);
maintaining a first-in-first-out (FIFO) queue storing encodings of data samples from previous mini-batches and corresponding labels of the data samples, the FIFO queue decoupling a number of encodings available for contrastive learning from a mini-batch size (§3.2 ¶2: dictionary);
the samples being stored in the FIFO queue (combination with dictionary queue of He for storing samples would yield storage in the FIFO queue of §3.2).
It would have been obvious before the effective filing date to a person of ordinary skill in the art to modify the method of Ferdinand by incorporating the momentum encoding of He. Both concern the art of contrastive incremental learning, and the incorporation would have, according to He, improve performance in contrastive loss applications by better sampling the underlying visual space and hence providing more consistent representations (§1 ¶3-4).
For claim 2, Ferdinand modified by He discloses the method of claim 1, as described above. Ferdinand further discloses: updating the rehearsal memory with samples of the new input data (Ferdinand §5.1 ¶2: overview of rehearsal memory implementation with memory bank and using herding sampling to update, He §3.2 ¶2).
For claim 3, Ferdinand modified by He discloses the method of claim 2, as described above. Ferdinand modified by He further discloses: wherein the rehearsal memory is updated based on a balanced-fine tuning such that a number of samples of data existing in the rehearsal memory from previous tasks is equal to a number of samples of the new input data to be stored in the rehearsal memory (He §3.2: “Dictionary as a Queue” discloses storing rehearsal memory as a queue, with the oldest removed and the newest added. Hence, the number from previous tasks will be equal to the number of new input as they are both a minibatch size).
For claim 4, Ferdinand modified by He discloses the method of claim 3, as described above. Ferdinand modified by He further discloses: wherein images of the samples of data existing in the rehearsal memory from previous tasks and images of the samples of the new input data to be stored in the rehearsal memory are both selected randomly (He §3.2 ¶1).
For claim 5, Ferdinand modified by He discloses the method of claim 1, as described above. Ferdinand further discloses: wherein the query encoder and the momentum encoder have a same size and configuration (§2.1 ¶2: as the key encoder is the previous step query encoder, they are identically sized and configured).
For claim 6, Ferdinand modified by He discloses the method of claim 1, as described above. He further discloses: wherein the FIFO queue is maintained at a fixed size by discarding oldest encodings from the FIFO queue as encodings of a current mini- batch are inserted, the fixed size being independent of the mini-batch size (§3.2 ¶2)
For claim 7, Ferdinand modified by He discloses the method of claim 6, as described above. Ferdinand discloses: wherein the second contrastive loss function contrasts an anchor encoding of a current mini-batch against one or more same-class encodings retrieved from the FIFO queue (Ferdinand §4.3, He §3.2 ¶2), the one or more same-class encodings being of data samples from previous mini-batches identified using the corresponding labels (Ferdinand §4.3: using data samples from minibatches, the samples associated with labels y, with He §3.2 ¶2, §1 ¶3 disclosing storing a longer queue including prior minibatches) stored in the FIFO queue (He §3.2 ¶2).
Claim 9-16 disclose computer media corresponding to the above limitations and hence are rejected for the same reasons.
Claim(s) 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over Ferdinand ("Attenuating catastrophic forgetting by joint contrastive and incremental learning", published 8/23/2022) in view of He ("Momentum contrast for unsupervised visual representation learning", published 2020) in view of Dunne (US 20200272899 A1).
For claim 17, Ferdinand discloses: obtain a labeled data set of images comprising data from a rehearsal memory stored in the memory and new input data (fig.1, “augmented minibatch”; §4.2 ¶1: augmented image minibatches are obtained; §4.1 ¶1, §5.2 ¶2: rehearsal memory and implementation);
augment the labeled data set to generate a first data set and a second data set, wherein each image of the first data set corresponds to a corresponding image of the second data set (ibid: contrastive and complementary image sets are obtained);
input the first data set into a query encoder and inputting the second data set into a key encoder to obtain encodings output by the query encoder and the key encoder (fig.1: augmented input is passed to identically structured query student and key teacher encoders, each of which generate respective projection (gamma) and feature (F) outputs, see §4.2 ¶1-2);
obtain a contrastive loss based on the encodings using a sum of a first contrastive loss function and a second contrastive loss function (§4 gives overview of total loss as sum of contrastive losses, see also fig.1 outlining contrastive losses generated from supervised data on the query network (L_con) and contrastive distillation loss (L_DCon)), wherein the first contrastive loss function uses encodings of different views of a same input image as anchor-positive pairs in a feature space (§4.2 ¶1, eq.6: contrastive loss is calculated in the inner sum, in the numerator, by taking the pairwise dot product of the projections of all elements in the same class, which would include all the augmentation or different views of the input image, hence, the first encodings of the different views for anchor positive pairs), and the second contrastive loss function uses, as anchor-positive pairs in the feature space, encodings of two different samples from a same class (§4.3, eq.7-8: L_Dcon is minimizes distance between B^t-1 and B^t (eq.8), each of which incorporate dot products of projections of base and augmented images z and z-hat, hence, anchor positive pairs are formed via encoding projections from different samples of the same class, the set of images from the past classes, see §4.3 ¶3), the two different samples being identified using the corresponding labels (§4.3: the samples being identified based on associated y labels);
update parameters of the query encoder based on the obtained contrastive loss (§4: updating parameters based on losses);
update parameters of the key encoder based on parameters of the query encoder (§2.1 ¶2: transferring parameters from query to key at each incremental step, with §5.1 ¶2 disclosing 250 epochs for each step); and
Ferdinand does not disclose: wherein the key encoder is a momentum encoder;
a computing device for training a classification model to be provided to an edge device, the computing device comprising:
a transceiver;
a memory; and
one or more processors configured to:
provide the classification model including parameters of the query encoder to the edge device via the transceiver;
maintaining a first-in-first-out (FIFO) queue storing encodings of data samples from previous mini-batches and corresponding labels of the data samples, the FIFO queue decoupling a number of encodings available for contrastive learning from a mini-batch size;
the samples being stored in the FIFO queue.
He discloses: wherein the key encoder is a momentum encoder (§3.2 “Momentum update” gives overview of a momentum encoder with exponentially weighted moving average per iteration, see eq.(2));
maintaining a first-in-first-out (FIFO) queue storing encodings of data samples from previous mini-batches and corresponding labels of the data samples, the FIFO queue decoupling a number of encodings available for contrastive learning from a mini-batch size (§3.2 ¶2: dictionary);
the samples being stored in the FIFO queue (combination with dictionary queue of He for storing samples would yield storage in the FIFO queue of §3.2).
It would have been obvious before the effective filing date to a person of ordinary skill in the art to modify the method of Ferdinand by incorporating the momentum encoding of He. Both concern the art of contrastive incremental learning, and the incorporation would have, according to He, improve performance in contrastive loss applications by better sampling the underlying visual space and hence providing more consistent representations (§1 ¶3-4).
Ferdinand modified by He does not disclose the remaining limitations.
Dunne discloses: a computing device for training a classification model to be provided to an edge device (fig.1 gives overview of model training and deployment to edge device), the computing device comprising:
a transceiver (fig.1B gives networking overview, with fig.4 showing sending and receiving between edge devices and a centralized training devices via a network, hence, transceiver);
a memory (fig.14: 1402); and
one or more processors (fig.4:1401) configured to:
provide the classification model including parameters of the query encoder to the edge device via the transceiver (fig.3:310 discloses deployment of classification model to an edge device via transceiver, see fig.4 showing sending and receiving of data, combination with Ferdinand and He yielding the provision of parameters of the trained query encoder to the edge device).
It would have been obvious before the effective filing date to a person of ordinary skill in the art to modify the method of Ferdinand modified by He by incorporating the edge deployment of Dunne. Both concern the art of machine learning, and the incorporation would have, according to Dune, address current limitations in AI technology, such as computer power limitations (0002-3, 0047).
Claim 18-19 disclose devices corresponding to methods 2-3 and are hence rejected for the same reasons.
Response to Arguments
Applicant’s arguments have been fully considered. In the remarks, Applicant argued:
1. The art of record does not disclose the FIFO queue storage encodings and labels used to identify data samples from previous minibatches that is used to identify same-class anchor positive pairs for the second contrastive loss function.
Examiner respectfully disagrees, as He teaches such a FIFO queue, as described above.
2. Ferdinand does not maintain a dictionary of prior mini-batch encodings, and instead I limited to use of a teacher model from the immediately preceding learning step. The second loss function of Ferdinand is a contrastive distillation loss computed across incremental steps, and not a contrastive loss function drawing same class positives from a queue of labeled encodings accumulated across mini-batches within a single training stage.
Examiner submits that He teaches a queue using samples from prior minibatches, as described, combination with Ferdinand would yield calculation of a contrastive loss over that set. Further, A POSITA would be motivated to combine in order to improve performance in contrastive loss applications by better sampling the underlying visual space and hence providing more consistent representations (§1 ¶3-4).
3. Ferdinand 3.4 does not describe a contrastive loss, but rather an alignment model that prevents catastrophic forgetting.
Examiner respectfully disagrees, and submits that additional clarification is needed to clarify difference. A POSITA would understand such a function to be a contrastive loss function, as evidenced by the naming itself (loss for “contrastive distillation”) and also in the function of the loss to contrastively push and pull classes according to positive pairs from the same class.
4. Ferdinand computes loss on samples from rehearsal memory (5.1) rather than same-class anchor positive pairs identified via labels stored in the FIFO queue.
Examine appreciates explanation of differences but submits that further clarifications are required. Furthermore, combination with the FIFO queue of He, for the various advantages listed, would yield the claimed queue.
5. He fails to disclose such features, as He does not allow the use of stored class labels, and hence, cannot identify different samples from a same class as anchor-positive pairs.
Examiner respectfully disagrees, since Ferdinand in combination with He would remedy the deficiencies of He, as described above.
6. The cited art further does not teach a momentum encoder updated as an exponentially moving average, as He’s update is part of an unsupervised instance-discrimination framework and applied to an unlabeled Queue.
Examiner submits that a POSITA would find it obvious to apply the queue of He to the technique of Ferdinand for the various advantages cited above, and that the queue could be adapted with minimal alteration.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Li (US 20220067506 A1) discloses out of domain adaptation based on a momentum model.
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 LIANG LI whose telephone number is (303)297-4263. The examiner can normally be reached Mon-Fri 9-12p, 3-11p MT (11-2p, 5-1a ET).
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. The examiner is available for interviews Mon-Fri 6-11a, 2-7p MT (8-1p, 4-9p ET).
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor Jennifer Welch can be reached on (571)272-7212. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from Patent Center and the Private Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from Patent Center or Private PAIR. Status information for unpublished applications is available through Patent Center or Private PAIR to authorized users only. Should you have questions about access to Patent Center or the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free).
/LIANG LI/
Primary examiner AU 2143
Prosecution Insights