DETAILED ACTION
This action is in response to the submission filed 22 April 2026 for application 18/314,132. Currently claims 1-30 are canceled. Claims 31-50 are newly added. Claims 31-50 are pending and have been examined.
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 .
Information Disclosure Statement
Information disclosure statements (IDS) were submitted on 22 August 2023 and 01 August 2024. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim 50 is rejected under 35 U.S.C. 102(a)(1) as being anticipated by Dorfman et al (DoCoFL: Downlink Compression for Cross-Device Federated Learning, 2023).
Regarding claim 50:
Dorfman teaches: An apparatus comprising: at least one processor; and at least one memory including computer program code ([Page 5, Figure 1] Note: Figure 1 shows a client mobile phone and computer both of which have processors and memory corresponding to at least one processor; and at least one memory including computer program code);
the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to: receive, by a network node from a user device, information indicating that the user device has a capability to perform at least one machine learning model update based on difference information ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Note: Client corresponds to user device. PS (parameter server) corresponds to network node. Correction shows that there is difference information);
transmit, by the network node to the user device, a machine learning model policy associated with model update, wherein the machine learning model policy comprises at least one of the following, for reducing overhead of with respect to transmitting a full machine learning model: difference information for a machine learning model with respect to model structure or model configuration parameters; difference information with respect to one or more weights or biases of the machine learning model; or difference information with respect to one or more layers of the machine learning model ([Page 5, Figure 1] Note: Figure 1 ( number 4 in green) shows PS sending correction to client corresponding to transmitting by the network node to the user device a difference information for a machine learning model with respect to model structure or model configuration parameters. [Page 14, Section C] To simplify mathematical notations and computations, we analyze a more general framework than DoCoFL, where at each round, each client can download any model from up to T rounds prior to their participation round as anchor. This generalized policy is described in Algorithm);
and receive, by the network node from the user device, an indication of the at least one machine learning model update being carried out based on the machine learning model policy ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. [Page 14, Section C] To simplify mathematical notations and computations, we analyze a more general framework than DoCoFL, where at each round, each client can download any model from up to T rounds prior to their participation round as anchor. This generalized policy is described in Algorithm 3. Note: Client Mobile phone corresponds to user device. PS (parameter server) corresponds to network node).
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 31 and 33-49 are rejected under 35 U.S.C. 103 as being unpatentable over Dorfman et al (DoCoFL: Downlink Compression for Cross-Device Federated Learning, 2023) in view of Barbieri et al (A Layer Selection Optimizer for Communication-Efficient Decentralized Federated Deep Learning, 2023).
Regarding claim 31:
Dorfman teaches: An apparatus comprising: at least one processor; and at least one memory including computer program code; the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to: ([Page 5, Figure 1] Note: Figure 1 shows a client mobile phone and computer both of which have processors and memory corresponding to at least one processor; and at least one memory including computer program code):
receive a machine learning model policy associated with a model update, wherein the machine learning model policy comprises vector difference information provided per layer of a machine learning model, the vector difference information including, for each of one or more layers, a mean squared error (MSE) value or a percentile cumulative distribution function (CDF) value indicating a difference between weights of a first version of the machine learning model and a second version of the machine learning model ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Also, see Figure 1 (number 7 in green) showing updating model weights. [Page 6, Column 1, Last but one Paragraph] Before we state our convergence result, we require an additional standard assumption about the correction compression operator Cc, namely, that it has a bounded Normalized Mean-Squared-Error (NMSE). Note: Also, see Equation (4) that shows weights. [Page 14, Section C] To simplify mathematical notations and computations, we analyze a more general framework than DoCoFL, where at each round, each client can download any model from up to T rounds prior to their participation round as anchor. This generalized policy is described in Algorithm 3. Note: Correction corresponds to the vector difference information);
However, Dorfman does not explicitly disclose: identify, based on the machine learning model policy, a subset of layers of the machine learning model indicated as updated; update only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information; determine, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model; compare, for each updated layer, the measured per-layer difference information to a threshold; and cause reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold.
Barbieri teaches, in an analogous system: identify, based on the machine learning model policy, a subset of layers of the machine learning model indicated as updated ([Page 22159, FIGURE 2] Layer selection process: each node runs a layer optimizer that sorts the layers according to their normalized squared gradient. [Page 22159, Column 2, Paragraph 1] Nodes receiving the model updates are then able to retrieve 0i(t) from ai(t) and fuse the received parameters according to (8). Note that the layer selection process is performed locally at each device and does not require any information from neighbors. In particular, we keep track of the gradients estimated during the local optimization step and sort them in descending order. Intuitively, layers exhibiting higher gradients convey more information about the local data and should be therefore selected and propagated to neighbors);
update only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors. Layer selection is implemented on each FL training round: the devices run an optimizer that sorts the layers of the NN model according to their expected contributions to the learning performance (e.g., measured by the squared norm of the gradients));
determine, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model ([Page 22157, Column 2, Paragraphs 1-4] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wTℓ hℓ−1 + bℓ), (1) where wℓ and bℓ are the weights and biases of layer ℓ, while for ℓ = 0 we have h0 = xh and hL = yh for ℓ = L. The weights and the biases of each layer can be conveniently aggregated into the matrix. W = [p1 · · · pL]T , (2) where pℓ = [wT ℓ , bT ℓ ]T ∈ RPℓ×1 with ℓ = 1, . . . , L is a compact vectorized representation of the weights and biases of layer ℓ with Pℓ being the overall number of trainable parameters for the ℓ-th layer. The goal of FL is to learn a global model W∞, shared across all interconnected agents);
compare, for each updated layer, the measured per-layer difference information to a threshold ([Page 22163, Column 2, Section C, Paragraph 1] In this section, we show that a partial random selection of the layers, regardless of gradient sorting, allows to train higher quality models. The validation focuses on three different threshold probability pr values namely pr = 0.2, pr = 0.6 and pr = 1.0, and considers the connectivity patterns defined in Sec. V-A);
and cause reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold ([Page 22168, Column 2, Paragraph 1] Carefully selecting the probability threshold pr taking into account M is beneficial for balancing the communication resources and the model quality. For example, choosing pr = 0.6 when M = 2 over pr = 0.2 for M = 6 allows to substantially reduce the number of transmitted data without introducing accuracy penalties. Opting for pr = 1 when M = 6 rather than pr = 0.6 for M = 12 heavily reduces the communication footprint without significant accuracy drops. CFL-LS approaches the performance of conventional CFL and FL tools but with a much lower communication burden as shown for pr = 1 and M = {6, 12}. On the other hand, convergence rates shown by CFL and FL are now slightly superior compared to CFL-LS schemes).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri to identify, based on the machine learning model policy, a subset of layers of the machine learning model indicated as updated; update only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information; determine, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model; compare, for each updated layer, the measured per-layer difference information to a threshold; and cause reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 33:
The system of Dorfman and Barbieri teaches: The apparatus of claim 32 (as shown above).
Dorfman further teaches: wherein the vector difference information further comprises, for each layer, a layer index and an associated difference metric pair including MSE value or the percentile CDF value ([Page 8, Figure 2] Note: Figure 2 shows CNN which is a convolutional neural network with layers associated with NMSE which is a Normalized Mean-Squared-Error).
Regarding claim 34:
The system of Dorfman and Barbieri teaches: The apparatus of claim 33 (as shown above).
However, Dorfman is not relied upon to teach: wherein updating only the identified subset of layers comprises scaling weights of each identified layer by an amount derived from the corresponding vector difference information.
Barbieri further teaches: wherein updating only the identified subset of layers comprises scaling weights of each identified layer by an amount derived from the corresponding vector difference information ([Page 22157, Column 2, Paragraph 1] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wTℓ hℓ−1 + bℓ), (1) where wℓ are the weights of layer ℓ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein updating only the identified subset of layers comprises scaling weights of each identified layer by an amount derived from the corresponding vector difference information. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 35:
The system of Dorfman and Barbieri teaches: The apparatus of claim 34 (as shown above).
However, Dorfman is not relied upon to teach: wherein the vector difference information further comprises separate difference information for weights and biases for each identified layer.
Barbieri further teaches: wherein the vector difference information further comprises separate difference information for weights and biases for each identified layer ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors. Layer selection is implemented on each FL training round: the devices run an optimizer that sorts the layers of the NN model according to their expected contributions to the learning performance (e.g., measured by the squared norm of the gradients) [Page 22157, Column 2, Paragraph 1] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wTℓ hℓ−1 + bℓ), (1) where wℓ and bℓ are the weights and biases of layer ℓ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the vector difference information further comprises separate difference information for weights and biases for each identified layer. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 36:
The system of Dorfman and Barbieri teaches: The apparatus of claim 35 (as shown above).
However, Dorfman is not relied upon to teach: wherein the machine learning model policy further comprises information indicating a correlation between weights of consecutive layers, and wherein updating the identified subset of layers comprises adjusting weights of a second layer based on an adjustment applied to a first layer according to the indicated correlation.
Barbieri further teaches: wherein the machine learning model policy further comprises information indicating a correlation between weights of consecutive layers, and wherein updating the identified subset of layers comprises adjusting weights of a second layer based on an adjustment applied to a first layer according to the indicated correlation ([Page 22157, Column 2, Paragraph 1] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wTℓ hℓ−1 + bℓ), (1) where wℓ are the weights of layer ℓ).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the machine learning model policy further comprises information indicating a correlation between weights of consecutive layers, and wherein updating the identified subset of layers comprises adjusting weights of a second layer based on an adjustment applied to a first layer according to the indicated correlation. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 37:
The system of Dorfman and Barbieri teaches: The apparatus of claim 36 (as shown above).
However, Dorfman is not relied upon to teach: wherein the threshold is layer-specific, and each measured per-layer difference information is compared to a corresponding threshold assigned to that layer.
Barbieri further teaches: wherein the threshold is layer-specific, and each measured per-layer difference information is compared to a corresponding threshold assigned to that layer ([Page 22163, Column 1, Section B, Paragraph 1] In Fig. 4 we compare two different strategies for gradient sorting and layer selection. The first one sorts the layers of the NN in a descending order (descending gradient sorting, DGS) while the second one uses an ascending order (AGS). Descending ordering prevents the transmission of layers with low gradients gt,i, while ascending ordering favors the transmission of these layers. These strategies are denoted as AGS pr = 0 and DGS pr = 0, as relying only on gradient sorting operations. The performances are also studied for the case where AGS and DGS integrate a randomized layer selection by using the strategy presented in Sec. III with probability threshold pr = 0.2, here referred to as AGS pr = 0.2 and DGS pr = 0.2. All strategies are compared with devices sharing M = 4 layers).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the threshold is layer-specific, and each measured per-layer difference information is compared to a corresponding threshold assigned to that layer. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 38:
The system of Dorfman and Barbieri teaches: The apparatus of claim 37 (as shown above).
However, Dorfman is not relied upon to teach: wherein transmitting only the measured per-layer difference information comprises transmitting an indication of layer indices for which the threshold is exceeded without transmitting full weight values of the layers.
Barbieri further teaches: wherein transmitting only the measured per-layer difference information comprises transmitting an indication of layer indices for which the threshold is exceeded without transmitting full weight values of the layers ([Abstract] Devices independently select fragments of the DNN to be shared with neighbors on each training round. Selection is based on a local optimizer that trades model quality improvement with sidelink communication resource savings. [Page 22158, Column 1, Paragraph 3] S0, the path loss index υ, and the average SNR γ¯0 at reference distance d0. A link is assigned as potential edge (k, i) ∈ E, ∀k ∈ Ni of the graph G if γk,i(t) > β where β is the receiver-side sensitivity threshold. Note: The greater than symbol > corresponds to exceeding).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein transmitting only the measured per-layer difference information comprises transmitting an indication of layer indices for which the threshold is exceeded without transmitting full weight values of the layers. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 39:
The system of Dorfman and Barbieri teaches: The apparatus of claim 38 (as shown above).
Dorfman further teaches: wherein the machine learning model policy further comprises information indicating a change in at least one hyper-parameter of the machine learning model, including a learning rate, and wherein updating the machine learning model further comprises updating the at least one hyper-parameter based on the machine learning model policy ([Page 4, Column 2, Algorithm 1] Input: learning rate η [Page 15, Algorithm 3] Input: learning rate η).
Regarding claim 40:
The system of Dorfman and Barbieri teaches: The apparatus of claim 39 (as shown above).
Dorfman further teaches: wherein the apparatus is further configured to: transmit, to the network node, a capability indication specifying support for vector difference information at a per-layer level ([Page 5, Figure 1] Note: Figure 1 (numbers 3, 5, and 6 in blue) shows client sending update to PS. Note: PS corresponds to network node and the steps in 3, 5, and 6 show support for the information coming from PS vector difference information).
and receive the machine learning model policy based on the capability indication ([Page 5, Figure 1] Note: Figure 1 also shows both PS and client receiving).
Regarding claim 41:
The system of Dorfman and Barbieri teaches: The apparatus of claim 40 (as shown above).
Dorfman further teaches: wherein the apparatus is further configured to: transmit, to the network node, a request for the machine learning model policy; and receive the machine learning model policy based on the request ([Page 5, Figure 1] Note: Figure 1 also shows both PS and client transmitting and receiving. [Page 14, Section C] Proof. To simplify mathematical notations and computations, we analyze a more general framework than DoCoFL, where at each round, each client can download any model from up to T rounds prior to their participation round as anchor. This generalized policy is described in Algorithm 3).
Regarding claim 42:
The system of Dorfman and Barbieri teaches: The apparatus of claim 41 (as shown above).
However, Dorfman is not relied upon to teach: wherein the difference information with respect to one or more layers comprises difference information indicated for each of one or more layers, including a first difference information for weights of the layer, and a second difference information for biases of the layer of the machine learning model.
Barbieri further teaches: wherein the difference information with respect to one or more layers comprises difference information indicated for each of one or more layers, including a first difference information for weights of the layer, and a second difference information for biases of the layer of the machine learning model ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the difference information with respect to one or more layers comprises difference information indicated for each of one or more layers, including a first difference information for weights of the layer, and a second difference information for biases of the layer of the machine learning model. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 43:
The system of Dorfman and Barbieri teaches: The apparatus of claim 42 (as shown above).
However, Dorfman is not relied upon to teach: wherein the difference information comprises an indication of one or more layers of the machine learning model that are updated.
Barbieri further teaches: wherein the difference information comprises an indication of one or more layers of the machine learning model that are updated ([Page 22159, FIGURE 2] Layer selection process: each node runs a layer optimizer that sorts the layers according to their normalized squared gradient. [Page 22159, Column 2, Paragraph 1] Nodes receiving the model updates are then able to retrieve 0i(t) from ai(t) and fuse the received parameters according to (8). Note that the layer selection process is performed locally at each device and does not require any information from neighbors. In particular, we keep track of the gradients estimated during the local optimization step and sort them in descending order. Intuitively, layers exhibiting higher gradients convey more information about the local data and should be therefore selected and propagated to neighbors).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the difference information comprises an indication of one or more layers of the machine learning model that are updated. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 44:
The system of Dorfman and Barbieri teaches: The apparatus of claim 43 (as shown above).
However, Dorfman is not relied upon to teach: wherein the difference information comprises a per parameter difference information for a subset of one or more weights or biases of the machine learning model.
Barbieri further teaches: wherein the difference information comprises a per parameter difference information for a subset of one or more weights or biases of the machine learning model ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the difference information comprises a per parameter difference information for a subset of one or more weights or biases of the machine learning model. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 45:
The system of Dorfman and Barbieri teaches: The apparatus of claim 44 (as shown above).
Dorfman further teaches: wherein the difference information comprises a difference or change in a state of the machine learning model, including an updated state or a change in a state of one or more parameters of the machine learning model, to be used by the apparatus for training or updating the machine learning model ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Also, see Figure 1 (number 7 in green) showing updating model weights).
Regarding claim 46:
The system of Dorfman and Barbieri teaches: The apparatus of claim 45 (as shown above).
However, Dorfman is not relied upon to teach: wherein the difference information comprises information indicating a relationship between two or more consecutive layers of the machine learning model that have been updated.
Barbieri further teaches: wherein the difference information comprises information indicating a relationship between two or more consecutive layers of the machine learning model that have been updated ([Page 22157, Column 2, Paragraphs 1 and 2] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wT ℓ hℓ−1 + bℓ), (1) where wℓ and bℓ are the weights and biases of layer ℓ, while for ℓ = 0 we have h0 = xh and hL = yh for ℓ = L. The weights and the biases of each layer can be conveniently aggregated into the matrix).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the difference information comprises information indicating a relationship between two or more consecutive layers of the machine learning model that have been updated. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 47:
The system of Dorfman and Barbieri teaches: The apparatus of claim 46 (as shown above).
However, Dorfman is not relied upon to teach: wherein the apparatus is further configured to: transmit, to the network node, information on at least one change to the machine learning model caused by the updating; and transmit, to the network node, information indicating the following: that one or more layers of the machine learning model were updated based on the difference information; an indication of one or more weights and biases that were updated based on the difference information; an amount that one or more weights and biases of the machine learning model were changed; an indication that weights and biases of the machine learning model were changed by more than a threshold; and an indication of one or more layers of the machine learning model for which weights and biases of the layer were changed by more than a threshold.
Barbieri further teaches: wherein the apparatus is further configured to: transmit, to the network node, information on at least one change to the machine learning model caused by the updating ([Page 22156, Figure 1] Decentralized FL system: cooperating agents autonomously share Machine Learning (ML) model fragments with neighbors and implement an average consensus strategy for model aggregation);
and transmit, to the network node, information indicating the following: that one or more layers of the machine learning model were updated based on the difference information ([Page 22157, Column 2, Paragraphs 1 and 2] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wT ℓ hℓ−1 + bℓ), (1));
an indication of one or more weights and biases that were updated based on the difference information ([Page 22157, Column 2, Paragraphs 1 and 2] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wT ℓ hℓ−1 + bℓ), (1) where wℓ and bℓ are the weights and biases of layer ℓ, while for ℓ = 0 we have h0 = xh and hL = yh for ℓ = L. The weights and the biases of each layer can be conveniently aggregated into the matrix);
an amount that one or more weights and biases of the machine learning model were changed ([Page 22157, Column 2, Paragraph 2] The weights and the biases of each layer can be conveniently aggregated into the matrix);
an indication that weights and biases of the machine learning model were changed by more than a threshold ([Page 22156, Column 1, Paragraph 1] FL preserves the privacy as it requires only the exchange of locally trained parameters of a ML model, namely the weights and biases of a Neural Network (NN));
and an indication of one or more layers of the machine learning model for which weights and biases of the layer were changed by more than a threshold ([Abstract] Devices independently select fragments of the DNN to be shared with neighbors on each training round. Selection is based on a local optimizer that trades model quality improvement with sidelink communication resource savings. [Page 22156, Column 1, Paragraph 1] FL preserves the privacy as it requires only the exchange of locally trained parameters of a ML model, namely the weights and biases of a Neural Network (NN). [Page 22158, Column 1, Paragraph 3] S0, the path loss index υ, and the average SNR γ¯0 at reference distance d0. A link is assigned as potential edge (k, i) ∈ E, ∀k ∈ Ni of the graph G if γk,i(t) > β where β is the receiver-side sensitivity threshold. Note: The greater than symbol > corresponds to exceeding).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri wherein the apparatus is further configured to: transmit, to the network node, information on at least one change to the machine learning model caused by the updating; and transmit, to the network node, information indicating the following: that one or more layers of the machine learning model were updated based on the difference information; an indication of one or more weights and biases that were updated based on the difference information; an amount that one or more weights and biases of the machine learning model were changed; an indication that weights and biases of the machine learning model were changed by more than a threshold; and an indication of one or more layers of the machine learning model for which weights and biases of the layer were changed by more than a threshold. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 48:
The system of Dorfman and Barbieri teaches: The apparatus of claim 47 (as shown above).
Dorfman further teaches: wherein the apparatus is further configured to: transmit, to the network node, a capabilities indication that indicates the following: that the apparatus has a capability to perform the updating of the machine learning model ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Also, see Figure 1 (number 7 in green) showing updating model weights);
that the apparatus has a capability to perform the updating of the machine learning model based on difference information ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Also, see Figure 1 (number 7 in green) showing updating model weights. Note: Correction corresponds to the vector difference information);
that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes the difference information for the machine learning model with respect to a structure of the model or model configuration parameters ([Page 2, Column 1, Paragraph 4] The primary challenge addressed by DoCoFL is that clients must download model parameters (i.e., weights) instead of model updates. Unlike updates, which are proportional to gradients and thus their norm is expected to decrease during training, the model parameters do not decay, rendering low-bit compression methods undesirable. As a result, and since clients can only download the updated model weights during their designated participation round, this can lead to a network bottleneck for low-resourced clients. [Page 2, Column 1, Paragraph 5] To address this bottleneck, DoCoFL decomposes the download burden by utilizing previous models, referred to as anchors, which clients can download prior to their participation round. Then, at the designated participation round, clients only need to download the correction, i.e., the difference between the updated model and the anchor);
However, Dorfman is not relied upon to teach: that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more layers of the machine learning model; and that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more biases or weights of the machine learning model.
Barbieri further teaches: that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more layers of the machine learning model ([Abstract] This paper proposes a communication-efficient design of consensus-driven FL optimized for training of Deep Neural Networks (DNNs). Devices independently select fragments of the DNN to be shared with neighbors on each training round. Selection is based on a local optimizer that trades model quality improvement with sidelink communication resource savings. The proposed technique is validated on a vehicular cooperative sensing use case characterized by challenging real-world datasets and complex DNNs typically employed in autonomous driving with up to 40 trainable layers. The impact of layer selection is analyzed under different distributed coordination configurations. The results show that it is better to prioritize the DNN layers possessing few parameters);
and that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more biases or weights of the machine learning model ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors. [Page 22157, Column 2, Paragraph 2] The weights and the biases of each layer can be conveniently aggregated into the matrix).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more layers of the machine learning model; and that the apparatus has a capability to perform the updating of the machine learning model based on the machine learning model policy that includes difference information with respect to one or more biases or weights of the machine learning model. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Regarding claim 49:
Dorfman teaches: A method performed by a user equipment, the method comprising: receiving a machine learning model policy associated with a model update, wherein the machine learning model policy comprises vector difference information provided per layer of the machine learning model, the vector difference information including, for each of one or more layers, a mean squared error (MSE) value or a percentile cumulative distribution function (CDF) value indicating a difference between weights of a first version of the machine learning model and a second version of the machine learning model ([Page 5, Figure 1] Note: Figure 1 (number 6 in blue) shows client sending update to PS based on correction. Also, see Figure 1 (number 7 in green) showing updating model weights. [Page 6, Column 1, Last but one Paragraph] Before we state our convergence result, we require an additional standard assumption about the correction compression
operator Cc, namely, that it has a bounded Normalized Mean-Squared-Error (NMSE). Note: Also, see Equation (4) that shows weights. [Page 14, Section C] To simplify mathematical notations and computations, we analyze a more general framework than DoCoFL, where at each round, each client can download any model from up to T rounds prior to their participation round as anchor. This generalized policy is described in Algorithm 3. Note: Client corresponds to user equipment. PS (parameter server) corresponds to network node. Correction corresponds to the vector difference information);
However, Dorfman is not relied upon to teach: identifying, based on the machine learning model policy, a subset of layers of a machine learning model indicated as updated; updating only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information; determining, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model; comparing, for each updated layer, the measured per-layer difference information to a threshold; and causing reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold.
Barbieri further teaches: identifying, based on the machine learning model policy, a subset of layers of a machine learning model indicated as updated ([Page 22159, FIGURE 2] Layer selection process: each node runs a layer optimizer that sorts the layers according to their normalized squared gradient. [Page 22159, Column 2, Paragraph 1] Nodes receiving the model updates are then able to retrieve 0i(t) from ai(t) and fuse the received parameters according to (8). Note that the layer selection process is performed locally at each device and does not require any information from neighbors. In particular, we keep track of the gradients estimated during the local optimization step and sort them in descending order. Intuitively, layers exhibiting higher gradients convey more information about the local data and should be therefore selected and propagated to neighbors);
updating only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information ([Page 22156, Column 2, Section B] We consider the decentralized FL system in Fig. 1 where a set of networked peer devices (or learners) cooperate to train a deep NN model. Each learner independently selects a subset of the NN layers and transmits the related trainable parameters (i.e., weights and biases) to the neighbors. Layer selection is implemented on each FL training round: the devices run an optimizer that sorts the layers of the NN model according to their expected contributions to the learning performance (e.g., measured by the squared norm of the gradients));
determining, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model ([Page 22157, Column 2, Paragraphs 1-4] The DNN model is composed by L layers, with outputs hℓ at layer ℓ computed by applying a non-linear (activation) function fℓ(.) to the weighted sum of the outputs of the previous layer hℓ−1 as hℓ = fℓ(wTℓ hℓ−1 + bℓ), (1) where wℓ and bℓ are the weights and biases of layer ℓ, while for ℓ = 0 we have h0 = xh and hL = yh for ℓ = L. The weights and the biases of each layer can be conveniently aggregated into the matrix. W = [p1 · · · pL]T , (2) where pℓ = [wT ℓ , bT ℓ ]T ∈ RPℓ×1 with ℓ = 1, . . . , L is a compact vectorized representation of the weights and biases of layer ℓ with Pℓ being the overall number of trainable parameters for the ℓ-th layer. The goal of FL is to learn a global model W∞, shared across all interconnected agents);
comparing, for each updated layer, the measured per-layer difference information to a threshold ([Page 22163, Column 2, Section C, Paragraph 1] In this section, we show that a partial random selection of the layers, regardless of gradient sorting, allows to train higher quality models. The validation focuses on three different threshold probability pr values namely pr = 0.2, pr = 0.6 and pr = 1.0, and considers the connectivity patterns defined in Sec. V-A.);
and causing reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold ([Page 22168, Column 2, Paragraph 1] Carefully selecting the probability threshold pr taking into account M is beneficial for balancing the communication resources and the model quality. For example, choosing pr = 0.6 when M = 2 over pr = 0.2 for M = 6 allows to substantially reduce the number of transmitted data without introducing accuracy penalties. Opting for pr = 1 when M = 6 rather than pr = 0.6 for M = 12 heavily reduces the communication footprint without significant accuracy drops. CFL-LS approaches the performance of conventional CFL and FL tools but with a much lower communication burden as shown for pr = 1 and M = {6, 12}. On the other hand, convergence rates shown by CFL and FL are now slightly superior compared to CFL-LS schemes).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Dorfman to incorporate the teachings of Barbieri identifying, based on the machine learning model policy, a subset of layers of a machine learning model indicated as updated; updating only the identified subset of layers of the machine learning model by adjusting weights of each identified layer according to the corresponding vector difference information; determining, after the updating, a measured per-layer difference information for each updated layer as a difference between weights of the updated machine learning model and the first version of the machine learning model; comparing, for each updated layer, the measured per-layer difference information to a threshold; and causing reduced signaling overhead relative to transmitting a full updated version of the machine learning model by transmitting, to a network node, only the measured per-layer difference information for layers whose measured per-layer difference information exceeds the threshold. One would have been motivated to do this modification because doing so would give the benefit of allowing to train higher quality models as taught by Barbieri [Page 22163, Column 2, Section C, Paragraph 1].
Claim 32 is rejected under 35 U.S.C. 103 as being unpatentable over Dorfman et al (DoCoFL: Downlink Compression for Cross-Device Federated Learning, 2023) in view of Barbieri et al (A Layer Selection Optimizer for Communication-Efficient Decentralized Federated Deep Learning, 2023) and further in view of Ali (Towards Autonomous and Efficient Machine Learning Systems A, 2021).
Regarding claim 32:
The system of Dorfman and Barbieri teaches: The apparatus of claim 32 (as shown above).
However, the system of Dorfman and Barbieri is not relied upon to teach: wherein the percentile CDF value comprises a 95th percentile value representing a difference between weights of the first version and the second version of the machine learning model.
Ali further teaches: wherein the percentile CDF value comprises a 95th percentile value representing a difference between weights of the first version and the second version of the machine learning model ([Page 111, Paragraph 1] The second column depicts the complementary CDF (CCDF) of latencies (normalized over the SLO and note the log scale of the y-axis). Here, we opted for showing the CCDF to better illustrate the performance of tail latencies. The vertical solid line is the SLO and the horizontal dashed line is the 95th percentile. [Page 60, Paragraph 1] After the weights are received and the global model is updated, the global model is evaluated on every client for every tier on their respective Test Data and their resulting accuracies are stored as the corresponding tier t’s accuracy for that round r. [Page 64, Section 3.4.1, Paragraph 1] In each training round, 5 clients are selected to train on their own data and send the trained weights to the server which aggregates them and updates the global model).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the combined system of Dorfman and Barbieri to incorporate the teachings of Ali wherein the percentile CDF value comprises a 95th percentile value representing a difference between weights of the first version and the second version of the machine learning model. One would have been motivated to do this modification because doing so would give the benefit of to better illustrating the performance as taught by Ali [Page 111, Paragraph 1].
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Sun et al (ML MODEL POLICY WITH DIFFERENCE INFORMATION FOR ML MODEL UPDATE FOR WIRELESS NETWORKS) discloses Executing Machine-Learning Models.
Alabbasi et al (WO 2021121585 A1) discloses METHODS FOR CASCADE FEDERATED LEARNING FOR TELECOMMUNICATIONS NETWORK PERFORMANCE AND RELATED APPARATUS.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to CHAITANYA RAMESH JAYAKUMAR whose telephone number is (571)272-3369. The examiner can normally be reached Mon-Fri 9am-1pm.
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, Omar Fernandez Rivas can be reached at (571)272-2589. 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.
/C.R.J./Examiner, Art Unit 2128
/OMAR F FERNANDEZ RIVAS/Supervisory Patent Examiner, Art Unit 2128
Prosecution Insights