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 .
Remarks
This Office Action is in response to applicant’s amendment filed on May 26, 2026, under which claims 1-31 are pending and under consideration.
Response to Arguments
Applicant’s amendments have overcome the previous claim rejections. However, upon further consideration, new and updated grounds of rejection have been made.
Applicant’s arguments have been fully considered but are not persuasive in distinguishing over the cited references, as currently applied in the updated rejections. Although the claim rejections have been updated to account for the amended claim language and have been updated to use an additional reference for the independent claims, the previously cited references are still currently applied.
Applicant argues:
First, the "training signal" 1116 in FIG. 13 of Ma does not include "control signaling that indicates, based on the layer ID information defined in the RRC configuration whether at least a subset of layers is to be frozen during training of the neural network," as in claim 20. In contrast to the "control signaling" of claim 20, the "training signal" in Ma is "a training sequence or training data." See paragraph [0142] of Ma. The distinction is further clarified, as the "training signal" in Ma does not indicate "whether at least a subset of layers is to be frozen during training of the neural network," as in claim 20.
…
As Tan does not disclose "wireless transmission" of the selection, and Ma does not disclose either the "layer identifier (ID) information" or "indicate to the UE whether at least a subset of layers are to be frozen during training of the neural network" even if combined in the proposed combination, Ma and Tan fail to disclose or suggest a "first wireless transmission" of an RRC configuration that "defines layer identifier (ID) information for each of a plurality of layers of a neural network for at least one of wireless channel compression at the UE, wireless channel measurement at the UE, wireless interference measurement at the UE, UE positioning, or wireless waveform determination at the UE," and the "second wireless transmission" includes "control signaling that indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of layers is to be frozen during training of the neural network." (Emphasis added).
(Applicant’s response, page 13).
In response, the Examiner agrees that neither Ma nor Tan alone teaches the entirety of the limitation of “control signaling that indicates, based on the layer ID information defined in the RRC configuration whether at least a subset of layers is to be frozen during training of the neural network.”
However, Ma teaches control signaling in general. While Ma teaches the example of sending a “training sequence or training data,” this transmission is a control signal because it affects the behavior of the UE. Furthermore, Ma also teaches the information 1020 to update AI/ML parameters (Ma, [0129]) and “AI/ML related information…sent separately from the training request” (Ma, [0140]), any or the combination of which can be regarded as a control signaling generically.
While Tan does not teach wireless transmissions, Tan teaches controlling a training process. Therefore, the difference between the claimed invention and Tan is that the claimed invention is a split system in which there is a server/BS and a UE. However, such a split system is already taught in Ma. Therefore, for the reasons stated in the rejections below, the combination of Ma and Tan addresses the limitations at issue.
Since applicant’s remarks focus on whether Tan or UE alone teaches the entirety of the above limitation, the Examiner noes that in response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Next, applicant argues:
Furthermore, in claim 20, the "neural network [is] for at least one of wireless channel compression at the UE, wireless channel measurement at the UE, wireless interference measurement at the UE, UE positioning, or wireless waveform determination at the UE," and "the subset of layers is to be frozen based on one or more of a channel characteristic, a stable environment condition, or a stationary or mobility condition of the UE," which is not disclosed or suggested by Ma and Tan.
In connection with dependent claim 21, the Office Action asserts that "the capability of the UE [in Ma] corresponds to the alternative of 'a condition at the UE." See Office Action at page 21. In amended claim 20, the alternative is amended to "a stationary or mobility condition of the UE," which distinguishes the "AI/ML capability" cited in Ma.
(Applicant’s response, page 13).
In response, the Examiner submits that Ma does teach “wireless waveform determination at the UE” (see Ma, [0077]: “The UE 110 may have multiple receive antennas, and in such embodiments the AI/ML module 552 may be configured to process waveforms received from multiple receive antennas as part of the waveform recovery process.”) and “a channel characteristic” (see Ma, [0114]: “the information sent at 1010 may include information indicating an AI/ML capability type of the UE….As another example, the plurality of AI/ML capability types may include different types that indicate different combinations of air interface components that are optimizable by AI/ML.”).
For claim 21, the Examiner has cited a new reference for this claim, namely Want et al. Therefore, applicant’s arguments for this claim are moot. Furthermore, regarding applicant’s observations for claim 23, the Examiner has likewise cited a new reference for this claim.
Claim Rejections - 35 USC § 103
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 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.
1. Claims 1-8, 10, 17-20, 23, and 28-31 are rejected under 35 U.S.C. 103 as being unpatentable over Ma et al. (US 2021/0160149 A1) (“Ma”) in view of Kulkarni et al., “Layer-wise training of deep networks using kernel similarity,” arXiv:1703.07115v1 [cs.LG] 21 Mar 2017 (“Kulkarni”) and Tan et al. (US 2021/0232909 A1) (“Tan”).
As to claim 1, Ma teaches an apparatus for wireless communication at a user equipment (UE) comprising: [[0003]: “a user equipment (UE) (also commonly referred to as a mobile station, a subscriber, a user, a terminal, a phone, and the like).” [0034]: “The EDs 110 are configured to operate, communicate, or both, in the wireless system 100. For example, the EDs 110 may be configured to transmit, receive, or both via wireless or wired communication channels. Each ED 110 represents any suitable end user device for wireless operation and may include such devices (or may be referred to) as a user equipment/device (UE).”]
memory; [[0182]: “a computer readable storage medium operatively coupled to the processor, the computer readable storage medium storing programming for execution by the processor.” [0043]: “The memory 208 stores instructions and data used, generated, or collected by the ED 110.”] and
at least one processor coupled to the memory, the at least one processor configured to: [[0182]: “a computer readable storage medium operatively coupled to the processor…” [0040]: “As shown in FIG. 2, the ED 110 includes at least one processing unit 200. The processing unit 200 implements various processing operations of the ED 110.”]
receive, in a first wireless transmission, a radio resource control (RRC) configuration from a wireless network entity that defines […] information for each of a plurality of layers of a neural network [[0136]: “the BS sends a training request to the UE at 1112 to trigger a training phase 1150. … In some embodiments, the training request may be set to the UE via RRC signaling...the training request may include initial training setting(s)/parameter(s), such as initial NN weights.” See also [0137]-[0139]: “the BS may also send AI/ML related information to the UE to facilitate joint training such as: Information indicating which AI/ML module is to be trained if there…”; [0140]: “the AI/ML related information may include an instruction for the UE to download initial AI/ML algorithm(s) and/or setting(s)/parameter(s).” That is, the “training setting(s)/parameter(s)” in the training request and the additional AI/ML related information includes information describing the model. See also [0142]: “the BS notifies the UE which AI/ML module(s)/component(s) is/are to be trained by including information in the training request that identifies one or more AI/ML modules/components.” The “AI/ML” in this reference is a “neural network” as described in [0129] (“AI/ML components, such as a neural network, is trained”), [0145], and [0146]. In regards to a plurality of layers, see [0152]: “where an AI/ML component is implemented with a deep neural network (DNN)… standardization may include a standard definition of the type(s) of neural network to be used, and certain parameters of the neural network (e.g., number of layers, number of neurons in each layer, etc.).” Note that a deep neural network, by definition, is a neural network with a plurality of layers.] for at least one of wireless channel compression at the UE, wireless channel measurement at the UE, wireless interference measurement at the UE, UE positioning, or wireless waveform determination at the UE [[0077]: “The UE 110 may have multiple receive antennas, and in such embodiments the AI/ML module 552 may be configured to process waveforms received from multiple receive antennas as part of the waveform recovery process.” [0076]: “The AI/ML module 552 of the UE 110 includes a joint waveform recovery, demodulator and source and channel decoder component 554.” That is, noting that this is an alternative expression reciting a list of items, the particular item of “wireless waveform determination at the UE” is taught.]
receive, in a second wireless transmission, control signaling that […] [[0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE. In some embodiments, the BS may send a training sequence/training data to the UE after a certain predefined time gap following transmission of the training request act 1012.” Furthermore, Ma also teaches the information 1020 to update AI/ML parameters (Ma, [0129]) and “AI/ML related information…sent separately from the training request” (Ma, [0140]). Any one or the combination of the above can be regarded as a second wireless transmission.] […] and
train the neural network based on the RRC configuration and the control signaling received from the wireless network entity. [[0141]: “after the UE has received the training request and initial training information from the network, the UE may send a response to the training request to the BS, as indicated at 1114 in FIG. 13. This response may confirm that the UE has entered a training mode.” [0142]: “As noted above, in some embodiments the BS notifies the UE which AI/ML module(s)/component(s) is/are to be trained by including information in the training request that identifies one or more AI/ML modules/components… By doing so, the BS informs the UE which AI/ML modules(s)/component(s) is/are to be trained…” [0145]: “training of an AI/ML module that includes one or more AI/ML components takes place jointly in the network and at the UE, as indicated at 1119 in FIG. 13.” [0146]: “In other embodiments, the UE and/or the BS may be able to update the training setup and parameters autonomously based on their own training process.” That is, the training may be at the UE or at both the UE and in the network, either of which reads on the instant claim limitation.]
Ma does not explicitly teach:
(1) The limitation that the information is “layer identifier (ID) information.” [The examiner notes that Ma teaches that the information includes “setting(s)/parameter(s), such as initial NN weights,” but this is not seen as an explicit teaching of layer ID information.]
(2) The limitation that the control signaling “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network” and “wherein the subset of layers is to be frozen” based on the one or more of the channel characteristic, stable environment condition or stationary or mobility condition of the UE.
Kulkarni teaches “layer identifier (ID) information.” [§ III.B, paragraph 2: “An operation at the kth layer of a MLP can be represented as follows … Wk denotes the weight matrix for the kth layer.” That is, the layer identifier k is defined in the information on the parameters, since k is part of the parameters Wk, as also expressed in equation (1) in this part of the reference.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of Ma with the teachings of Kulkarni by implementing the information to also define “layer identifier (ID) information.” Doing so would have enabled the layers of a neural network to be to be defined in the parameter information such that parameters are associated with layers. Furthermore, since the instant limitation at issue only pertains to specific information included in a transmission of information, doing so would also have been a simple combination of prior art elements according to known methods to yield predicable results (MPEP § 2143(I)(A)) since the addition of a known type of information to a base prior art reference that only differs from the lack of an explicit teaching of this type of information by known methods, and that in combination, each element merely performs the same function as it does separately, and one of ordinary skill in the art would have recognized that the results of the combination were predictable, namely the result of including information in the form of layer identifier (ID) information.
The combination of references thus far does not teach limitations (2) identified above.
Tan teaches control signaling that “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network.” [Abstract: “The computer executable components can include: an assessment component that identifies units of a neural network, a selection component that selects a subset of units of the neural network, and a freeze-out component that freezes the selected subset of units of the neural network so that weights of output connections from the frozen subset of units will not be updated for a training run.” The selected subset of units can be on the basis entire layers, as disclosed in [0033]: “the selection component 106 can select a subset of units of the neural network comprising one or more entire layers of units”; [0035]: “In another example, the freeze-out component 108 freezes one or more layers of the neural network selected by the selection component 106 so that weights of output connections from the one or more frozen layers will not be updated for a training run.” Regarding the limitation of “based on the layer ID information defined in the RRC configuration,” the “layer ID information defined in the RRC configuration” is already taught by the existing modification of Ma. Furthermore, Tan is consistent with “based on” such a layer ID because Tan teaches identifying the layers. See [0045] teaches: “Block 902 represents a first act that includes identifying units of a neural network (e.g., via the assessment component 104). At 904, a subset of units of the neural network are selected (e.g., via the selection component 106). At 906, the selected subset of units of the neural network are frozen so that weights of output connections from the frozen subset of units will not be updated for a training run (e.g., via the freeze-out component 108).” In general, Tan implies the use of “layer identifier information” in order for the subset to be identified, selected, and used by different components of the system. See also [0040] and FIG. 5A, which illustrates four layers of a neural network 502, 504, 506 and 508. In this example, the entire layer 506 and all units therein are randomly frozen for a training run.” That is, a particular layer is identified for being frozen, which also teaches that there is layer identifier information in order for that layer (506) to be identified.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Ma with the teachings of Tan by implementing the layer freezing technique of Tan in a system in which a UE performs training based on signals from a base station, particularly by implementing the control signaling such that it “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network” and such that “the subset of layers is to be frozen” based on The motivation for doing so would have been to mitigate overfitting and improve the neural network’s ability to generalize, as suggested by Tan (see [0028]: “Regularization refers to techniques to solve the overfitting problem by making slight modifications to the learning algorithm, enabling the neural network model to more accurately generalize to new situations or data sets”; [0030]: “Freeze-out provides an improved regularization technique as it eliminates the need to update the weights of output connections.”).
As to claim 2, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, further comprising a transceiver coupled to the at least one processor, [Ma, [0041]: “The ED 110 also includes at least one transceiver 202.” See FIG. 2 which shows that the transceiver 202 is coupled to the processor (processing unit 200).]
wherein the wireless network entity includes a base station, a transmission reception point (TRP), a core network component, a server or another UE, [Ma, [0035]: “In FIG. 1, the RANs 120 include base stations (BSs) 170 a-170 b (generically referred to as BS 170), respectively.” The instant limitation is an alternative expression, and the alternative of “base station” is taught as quoted above.] and wherein the control signaling in the second wireless transmission signals that the at least the subset of layers is to be frozen during training of the neural network [Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE.” (See also other forms of control signaling discussed in [0129] and [0140] as discussed in the rejection of the parent independent claim.) The signal function of “that at least the subset of layers is to be frozen” is taught by Tan for the reasons discussed in the parent independent claim and is also covered by the motivation given for Tan in the rejection of the parent independent claim.], wherein the subset of layers is frozen based on one or more of: a channel characteristic, a stable environment condition, or a stationary or mobility condition at the UE [Noting that this claim recites an alternative expression, the item of “a channel characteristic” is taught by Ma, [0114]: “the information sent at 1010 may include information indicating an AI/ML capability type of the UE….As another example, the plurality of AI/ML capability types may include different types that indicate different combinations of air interface components that are optimizable by AI/ML.” Here, a combination of optimizable air interface components is a channel characteristic, because “air interface component” refers to channel features. See Ma, [0073]: “For example, an air interface may include one or more components defining the waveform(s), frame structure(s), multiple access scheme(s), protocol(s), coding scheme(s) and/or modulation scheme(s) for conveying data over a wireless communications link.” Since this step occurs before the other transmissions, the other transmissions, including the training information and signal, are “based on” the condition at the UE. The Examiner notes that the instant claim merely recites “based on,” which only requires, for example, a relationship in which the existence of the subset of layers was in response to the condition of the UE, and does not require any specific relationships or algorithm for selecting the subset of layers.]
As to claim 3, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, as set forth above.
Kulkarni further teaches “wherein the control signaling indicates for the UE to perform a hierarchical training to train different layers in a particular order, wherein the particular order is identified based on the layer ID information from the RRC configuration.” [§ III.B, paragraph 1: “A MLP build a hierarchy of feature representation where a subsequent layer build a representation on the top of the features computed by previous layers.” As shown in Algorithm 1 (description of the heading), the kth layer is trained given the previously trained k-1 layers, which is therefore training the layers in a particular order, namely the order of the layers.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of Ma, Kulkarni, and Tan, including the above further teachings of Kulkarni, by further modifying Ma, as already modified thus far, to implement layerwise training such that “the control signaling indicates for the UE to perform a hierarchical training to train different layers in a particular order, wherein the particular order is identified based on the layer ID information from the RRC configuration.” The motivation for doing so is to implement a method for training that achieves good generalization performance (see Kulkarni, § II, paragraph 1: “It was demonstrated that the unsupervised training leads to a good generalization performance by appropriately initializing weights in a region near a good local minimum which provides feature representations that are high-level abstractions of the input.”).
As to claim 4, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, wherein the control signaling is received in at least one of: a medium access control (MAC) control element (CE) (MAC-CE), downlink control information (DCI), sidelink control information (SCI), or a sidelink message. [Ma, [0124]: “Dynamic control channel: When the number of bits required to send the training sequence/training data is less than a certain threshold, a dynamic control channel may be used to send the training sequence/training data. In some embodiment, several levels of bit lengths may be defined. The different bit lengths may correspond to different DCI formats or different DCI payloads. The same DCI can be used for carrying training sequences/data for different AI/ML modules. In some embodiments, a DCI field may contain information indicating an AI/ML module the training sequence/training data is to be used to train.” Note that “training sequences/data” refers to the transmission in step 1016 as described in [0122] (“At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE.”). Therefore, the alternative of using DCI is disclosed in Ma.]
As to claim 5, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 4, wherein, to receive the RRC configuration and the control signaling, the at least one processor is configured to:
receive multiple sets of neural network training parameters in higher-layer signaling; [Ma, [0136]: “In some embodiments, the training request may be set to the UE via RRC signaling.” Here, the RRC signaling is a type of higher-layer signaling in the absence of further limitations as to a more specific definition of this term. The Examiner interprets the instant limitation as being met if RRC is used for any part of the multiple sets of neural network training parameters.] and
receive an indication of one of the multiple sets of neural network training parameters in at least one of the MAC-CE, the DCI, or a combination thereof. [Ma, [0136]: “In some embodiments, the training request may be sent to the UE through DCI (dynamic signaling) on a downlink control channel or on a data channel. For example, in some embodiments the training request may be sent to the UE with UE specific or UE common DCI. For example, UE common DCI may be used to send a training request to all UEs or a group of UEs. In some embodiments, the training request may be set to the UE via RRC signaling. In some embodiments, the training request may include initial training setting(s)/parameter(s), such as initial NN weights.” Ma, [0210]: “Example Embodiment 131. …transmitting the AI/ML training request through downlink control information (DCI) on a downlink control channel or RRC signaling or the combination of the DCI and RRC signaling. …Example Embodiment 135…wherein the dynamic control channel includes a dynamic control information (DCI) field containing information indicating an AI/ML module that is to be trained.” That is Ma, [0210] teaches that DCI is used to indicate the training parameter of “an AI/ML module that is to be trained” and further teaches that the combination of DCI and RRC may be used.]
As to claim 6, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, wherein the control signaling in the second wireless transmission indicates one or more neural network training parameters to be used in the training of the neural network, [Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE” (and also Ma, [0129]: “At 1020, the BS sends information to the UE to update AI/ML parameters”; and [0140]: “AI/ML related information…sent separately from the training request”).]
Tan further teaches “wherein the one or more neural network training parameters includes at least one of: a channel state information reporting identifier, a channel state reference signal identifier, a component carrier identifier, a bandwidth part (BWP) identifier, a neural network identifier, a first indication of at least one layer to be trained, a second indication of the at least one layer to be frozen, a group of multiple layers to be trained, a subset of layers to be trained, or a combination thereof.” [Abstract: “The computer executable components can include: an assessment component that identifies units of a neural network, a selection component that selects a subset of units of the neural network, and a freeze-out component that freezes the selected subset of units of the neural network so that weights of output connections from the frozen subset of units will not be updated for a training run.” The selected subset of units can be on the basis entire layers, as disclosed in [0033]: “the selection component 106 can select a subset of units of the neural network comprising one or more entire layers of units”; [0035]: “In another example, the freeze-out component 108 freezes one or more layers of the neural network selected by the selection component 106 so that weights of output connections from the one or more frozen layers will not be updated for a training run.” That is, the alternative of “a second indication of the at least one layer to be frozen” is taught by Tan.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of the references combined thus far, including the above teachings of Tan, so to have also arrived at the claimed invention of the instant dependent claim. The motivation for doing so is covered by the one given for Tan in the rejection of the parent independent claim.
As to claim 7, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, wherein the at least one processor is further configured to:
receive a training command in the second wireless transmission, [Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE. In some embodiments, the BS may send a training sequence/training data to the UE after a certain predefined time gap following transmission of the training request at 1012.” Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE” (and also Ma, [0129]: “At 1020, the BS sends information to the UE to update AI/ML parameters”; and [0140]: “AI/ML related information…sent separately from the training request”). Since this training signal “starts the training phase,” it is a training command. See also Ma, [0129] (“At 1020, the BS sends information to the UE to update AI/ML parameters”) and [0140] (“AI/ML related information…sent separately from the training request”) which are also training commands.] wherein the UE applies one or more neural network training parameters indicated in the second wireless transmission to train the neural network at the UE in response to receiving the training command. [The training signal described in [0122] initiates the training phase (see [0129]: “For example, the parameters of an AI/ML module, such as neural network weights, may be updated/modified based on measurement results returned by the UE.”), which applies this information that was sent in the training request and also the training signal. The “training sequence or training data” described in [0122] corresponds to training parameters.]
As to claim 8, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 7, wherein the training command is a group common command, and the group common command is received over a group common downlink control information (DCI). [Ma, [0121]: “For example, in some embodiments the training request may be sent to the UE as UE specific or UE common DCI. For example, UE common DCI may be used to send a training request to all UEs or a group of UEs.” Here, “UE common DCI” refers to a “group common DCI” in the sense of common to the group, since the context is a group of UE.]
As to claim 10, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 7, wherein the at least one processor is further configured to:
apply the one or more neural network training parameters indicated in the second wireless transmission to train one or more neural networks identified in the training command. [ [Ma, [0125]: “The DCI used to schedule such a data channel can carry the information required for decoding the data channel and AI/ML module indicator(s) to indicate which AI/ML module(s) the training sequence/data is for.” Ma [0128]: “In some embodiments, the training response message may include feedback information indicating an updated training sequence for an iterative training process (e.g., for autoencoder based ML) or certain type(s) of measurement results to help Tx/Rx to further train or refine the training of a NN, e.g., for enforcement learning.” See also Ma, [0129], which states that the communication at step 1020 is “to update AI/ML parameters, such as neural network weights” (i.e., a neural network identified in the training command.)]
As to claims 17-19, these claims are directed to a method comprising the same or substantially the same operations as those of claims 1, 2, and 7. Therefore, the rejections made to claims 1, 2, and 7 are applied to claims 17-19, respectively.
As to claim 20, Ma teaches an apparatus for wireless communication, comprising:
memory; [[0044]: “As shown in FIG. 3, the base station 170 includes … at least one memory 258.” [0045]: “the memory 258 could store software instructions or modules configured to implement some or all of the functionality and/or embodiments described herein and that are executed by the processing unit(s) 250.”] and
at least one processor coupled to the memory, the at least one processor configured to: [[0044]: “The processing unit 250 can also be configured to implement some or all of the functionality and/or embodiments described in more detail herein.” [0045]: “the memory 258 could store software instructions or modules…that are executed by the processing unit(s) 250.”]
transmit, in a first wireless transmission to the UE, a radio resource control (RRC) configuration that defines […] information for each of a plurality of layers of a neural network for wireless communication by the UE, [[0136]: “the BS sends a training request to the UE at 1112 to trigger a training phase 1150. … In some embodiments, the training request may be set to the UE via RRC signaling...the training request may include initial training setting(s)/parameter(s), such as initial NN weights.” See also [0137]-[0139]: “the BS may also send AI/ML related information to the UE to facilitate joint training such as: Information indicating which AI/ML module is to be trained if there…”; [0140]: “the AI/ML related information may include an instruction for the UE to download initial AI/ML algorithm(s) and/or setting(s)/parameter(s).” That is, the “training setting(s)/parameter(s)” in the training request and the additional AI/ML related information includes information describing the model. See also [0142]: “the BS notifies the UE which AI/ML module(s)/component(s) is/are to be trained by including information in the training request that identifies one or more AI/ML modules/components.” The “AI/ML” in this reference is a “neural network” as described in [0129] (“AI/ML components, such as a neural network, is trained”), [0145], and [0146]. In regards to a plurality of layers, see [0152]: “where an AI/ML component is implemented with a deep neural network (DNN)… standardization may include a standard definition of the type(s) of neural network to be used, and certain parameters of the neural network (e.g., number of layers, number of neurons in each layer, etc.).” Note that a deep neural network, by definition, is a neural network with a plurality of layers.] for at least one of wireless channel compression at the UE, wireless channel measurement at the UE, wireless interference measurement at the UE, UE positioning, or wireless waveform determination at the UE [[0077]: “The UE 110 may have multiple receive antennas, and in such embodiments the AI/ML module 552 may be configured to process waveforms received from multiple receive antennas as part of the waveform recovery process.” [0076]: “The AI/ML module 552 of the UE 110 includes a joint waveform recovery, demodulator and source and channel decoder component 554.” That is, noting that this is an alternative expression reciting a list of items, the particular item of “wireless waveform determination at the UE” is taught.] and
transmit, in a second wireless transmission, control signaling that […] [[0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE. In some embodiments, the BS may send a training sequence/training data to the UE after a certain predefined time gap following transmission of the training request act 1012.” The training signal is “for” the training parameters because it is for the use of those training parameters, and the training signal (and transmission thereof) is “based on” the earlier information because the signal was based on the UE’s request (step 1014) received in response to the earlier training request. Furthermore, Ma also teaches the information 1020 to update AI/ML parameters (Ma, [0129]) and “AI/ML related information…sent separately from the training request” (Ma, [0140]). Any one or the combination of the above can be regarded as a second wireless transmission.] […] based on one or more of a channel characteristic, a stable environment condition , or a stationary or mobility condition of the UE. [Noting that this claim recites an alternative expression, the item of “a channel characteristic” is taught by Ma, [0114]: “the information sent at 1010 may include information indicating an AI/ML capability type of the UE….As another example, the plurality of AI/ML capability types may include different types that indicate different combinations of air interface components that are optimizable by AI/ML.” Here, a combination of optimizable air interface components is a channel characteristic, because “air interface component” refers to channel features. See Ma, [0073]: “For example, an air interface may include one or more components defining the waveform(s), frame structure(s), multiple access scheme(s), protocol(s), coding scheme(s) and/or modulation scheme(s) for conveying data over a wireless communications link.” Since this step occurs before the other transmissions, the other transmissions, including the training information and signal, are “based on” the condition at the UE. The Examiner notes that the instant claim merely recites “based on,” which only requires, for example, a relationship in which the existence of the subset of layers was in response to the condition of the UE, and does not require any specific relationships or algorithm for selecting the subset of layers.]
Ma does not explicitly teach:
(1) The limitation that the information is “layer identifier (ID) information.” [The examiner notes that Ma teaches that the information includes “setting(s)/parameter(s), such as initial NN weights,” but this is not seen as an explicit teaching of layer ID information.]
(2) The limitation that the control signaling “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network” and “wherein the subset of layers is to be frozen” based on the one or more of the channel characteristic, stable environment condition or stationary or mobility condition of the UE.
Kulkarni teaches “layer identifier (ID) information.” [§ III.B, paragraph 2: “An operation at the kth layer of a MLP can be represented as follows … Wk denotes the weight matrix for the kth layer.” That is, the layer identifier k is defined in the information on the parameters, since k is part of the parameters Wk, as also expressed in equation (1) in this part of the reference.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of Ma with the teachings of Kulkarni by implementing the information to also define “layer identifier (ID) information.” Doing so would have enabled the layers of a neural network to be to be defined in the parameter information such that parameters are associated with layers. Furthermore, since the instant limitation at issue only pertains to specific information included in a transmission of information, doing so would also have been a simple combination of prior art elements according to known methods to yield predicable results (MPEP § 2143(I)(A)) since the addition of a known type of information to a base prior art reference that only differs from the lack of an explicit teaching of this type of information by known methods, and that in combination, each element merely performs the same function as it does separately, and one of ordinary skill in the art would have recognized that the results of the combination were predictable, namely the result of including information in the form of layer identifier (ID) information.
The combination of references thus far does not teach limitations (2) identified above.
Tan teaches control signaling that “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network” and “wherein the subset of layers is to be frozen” [Abstract: “The computer executable components can include: an assessment component that identifies units of a neural network, a selection component that selects a subset of units of the neural network, and a freeze-out component that freezes the selected subset of units of the neural network so that weights of output connections from the frozen subset of units will not be updated for a training run.” The selected subset of units can be on the basis entire layers, as disclosed in [0033]: “the selection component 106 can select a subset of units of the neural network comprising one or more entire layers of units”; [0035]: “In another example, the freeze-out component 108 freezes one or more layers of the neural network selected by the selection component 106 so that weights of output connections from the one or more frozen layers will not be updated for a training run.” Regarding the limitation of “based on the layer ID information defined in the RRC configuration,” the “layer ID information defined in the RRC configuration” is already taught by the existing modification of Ma. Furthermore, Tan is consistent with “based on” such a layer ID because Tan teaches identifying the layers. See [0045] teaches: “Block 902 represents a first act that includes identifying units of a neural network (e.g., via the assessment component 104). At 904, a subset of units of the neural network are selected (e.g., via the selection component 106). At 906, the selected subset of units of the neural network are frozen so that weights of output connections from the frozen subset of units will not be updated for a training run (e.g., via the freeze-out component 108).” In general, Tan implies the use of “layer identifier information” in order for the subset to be identified, selected, and used by different components of the system. See also [0040] and FIG. 5A, which illustrates four layers of a neural network 502, 504, 506 and 508. In this example, the entire layer 506 and all units therein are randomly frozen for a training run.” That is, a particular layer is identified for being frozen, which also teaches that there is layer identifier information in order for that layer (506) to be identified.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Ma with the teachings of Tan by implementing the layer freezing technique of Tan in a system in which a UE performs training based on signals from a base station, particularly by implementing the control signaling such that it “indicates, based on the layer ID information defined in the RRC configuration, whether at least a subset of the plurality of layers is to be frozen during training of the neural network” such that “the subset of layers is to be frozen” based on the one or more of the channel characteristic, stable environment condition or stationary or mobility condition of the UE. The motivation for doing so would have been to mitigate overfitting and improve the neural network’s ability to generalize, as suggested by Tan (see [0028]: “Regularization refers to techniques to solve the overfitting problem by making slight modifications to the learning algorithm, enabling the neural network model to more accurately generalize to new situations or data sets”; [0030]: “Freeze-out provides an improved regularization technique as it eliminates the need to update the weights of output connections.”).
As to claim 23, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 20, as set forth above.
Kulkarni further teaches “wherein the control signaling indicates for the UE to perform a hierarchical training to train different layers in a particular order, wherein the particular order is identified based on the layer ID information from the RRC configuration.” [§ III.B, paragraph 1: “A MLP build a hierarchy of feature representation where a subsequent layer build a representation on the top of the features computed by previous layers.” As shown in Algorithm 1 (description of the heading), the kth layer is trained given the previously trained k-1 layers, which is therefore training the layers in a particular order, namely the order of the layers.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of Ma, Kulkarni, and Tan, including the above further teachings of Kulkarni, by further modifying Ma, as already modified thus far, to implement layerwise training such that “the control signaling indicates for the UE to perform a hierarchical training to train different layers in a particular order, wherein the particular order is identified based on the layer ID information from the RRC configuration.” The motivation for doing so is to implement a method for training that achieves good generalization performance (see Kulkarni, § II, paragraph 1: “It was demonstrated that the unsupervised training leads to a good generalization performance by appropriately initializing weights in a region near a good local minimum which provides feature representations that are high-level abstractions of the input.”).
As to claims 28-30, these claims are directed to a method comprising the same or substantially the same operations as those of claims 20 and 22-23. Therefore, the rejections made to claims 20 and 22-23 are applied to claims 28-30, respectively.
Additionally, the preamble recitation of “at a base station” is taught by the parts of Ma cited in the rejection of claim 20 which discuss a base station (BS) (e.g., [0044]: “As shown in FIG. 3, the base station 170”).
As to claim 31, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, as set forth above.
Tan further teaches wherein the configuration includes:
a first indication of at least one layer of the plurality of layers to be trained, wherein remaining layers are to be frozen during the training of the neural network,
a second indication of the at least one layer to be frozen, wherein the remaining layers are to be trained during the training of the neural network, or
both the first indication of the at least one layer to be trained and the second indication of the at least one layer to be frozen during the training of the neural network. [Noting that the instant claim recites an “or”-delimited alternative expression, the second alternative of “a second indication of the at least one layer to be frozen, wherein the remaining layers are to be trained during the training of the neural network” is taught by Tan, since the “subset of units of the neural network” to be frozen an indication of those elements to be frozen. See also FIG. 7, where the “specified node set FN which can be frozen out” is an indication that is an input to the algorithm. Additionally, as shown in the middle of the algorithm in FIG. 7, the layers not in the subset (as stated in the line “if node i does not belong to FN”) correspond to remaining layers to be trained.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Man and Tan to have further arrived at the limitations of the instant dependent claim. The motivation for doing so is the same as the motivation already given for the teachings of Tan in the rejection of the parent independent claim, since the teachings of Tan discussed above for the instant dependent claim are part of the layer freezing technique of Tan.
2. Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Ma in view of Kulkarni and Tan, and further in view of Zhang et al. (US 2019/0147344 A1) (“Zhang”).
As to claim 9, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 7, wherein the control signaling indicates for the UE to apply the configuration to train […] of multiple neural networks at the UE in response to receiving the training command. [Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE. In some embodiments, the BS may send a training sequence/training data to the UE after a certain predefined time gap following transmission of the training request at 1012.” Since this training signal “starts the training phase,” it is an indication to train the neural network based on the settings and parameters previously transmitted.]
Ma as modified thus far does not explicitly teach the limitation that “each layer” of the multiple neural networks is trained. However, this is a conventional feature of neural network training.
Zhang teaches training “each layer” of a neural network [[0005]: “a neural network has two different computing phases, namely a training phase and an inference phase. The training phase is used to adjust the weights of each layer to make the neural network fit a specific function.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Zheng by training each layer of the multiple neural networks. The motivation for doing so would have been to implement a standard method for training a neural network to fit a specific function, as suggested by Zhang.
3. Claims 11-12 and 24-25 are rejected under 35 U.S.C. 103 as being unpatentable over Ma in view of Kulkarni and Tan, and further in view of Wang et al. (US 2021/0182658 A1) (“Wang”).
As to claim 11, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 7, wherein the memory and the at least one processor are configured to receive training command […] and to train the neural network […]. [As shown in FIG. 13, the UE and BS communicate wirelessly. See Ma, [0037]: The BSs 170 communicate with one or more of the EDs 110 over one or more air interfaces 190 a using wireless communication links (e.g. radio frequency (RF), microwave, infrared (IR), etc.). This includes the training request (which is sent via RRC channel, a downlink control channel or on a data channel as described see [0136]), and also the training signal on which the training is performed (as described in Ma, [0142]: “At 1116 the BS starts the training phase 1150 by sending a training signal that includes a training sequence or training data to the UE… Non-limiting examples of channels that may be used by the BS to send training sequences or training data to UE include those discussed above with reference to FIG. 12, namely a dynamic control channel, a data channel and/or RRC channel.”). The AI/ML module is trained on a channel for communication between the UE and the base station, in addition to being trained using a channel for such communication. See Ma, [0074]: “an AI/ ML module 502,552 that is trainable in order to provide a tailored personalized air interface between the base station 170 and UE 110.”]
Ma as modified thus far does not explicitly teach the limitations that the training command is received “in a first frequency range or a first frequency band” and that the neural network is trained “on a second frequency range or a second frequency band.”
Wang teaches, “in a first frequency range or a first frequency band” and “on a second frequency range or a second frequency band” [[0175]: “the base station 120 (and/or the core network server 302 by way of the base station 120) communicates the configuration of the DNN to the UE 110 using a first component carrier, where the DNN configuration corresponds to forming a DNN for processing a second component carrier of the carrier aggregation.” That is, the first component carrier is used for communication of the configuration, which is analogous to the training command of the instant claim, while the second component carrier is for the function of the DNN, which is analogous for a function for which the neural network of the instant claim is being configured for on. See [0025]: “training a DNN on transmitter and/or receiver processing chain operations.” Note that the two component carriers have different frequency bands. See [0156]: “the first component carrier resides in a licensed band and the second component carrier resides in an unlicensed band”; [0144]: “Image 1020 depicts a non-contiguous, inter-band configuration, where at least one component carrier used in the carrier-aggregation-communications resides in a different frequency band. For example, in various implementations, band 1 of image 1020 corresponds to licensed bands allocated to the base station 120 and band 2 of image 1020 corresponds to unlicensed bands used by the base station 120, such as frequency bands accessed through Licensed-Assisted Access (LAA).”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Wang by implementing the use of the different component carriers as taught in Wang for communication and DNN functionality, so as to arrive at the limitations of receiving the training command “in a first frequency range or a first frequency band” and having the neural network be trained “on a second frequency range or a second frequency band.” The motivation would have been to enable the neural network to process information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier (see Wang, [0026]: “at least one deep neural network (DNN) configuration for processing information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier.”).
As to claim 12, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 7, wherein the memory and the at least one processor are configured to receive the training command […] and to train the neural network […]. [As shown in FIG. 13, the UE and BS communicate wirelessly. See Ma, [0037]: The BSs 170 communicate with one or more of the EDs 110 over one or more air interfaces 190 a using wireless communication links (e.g. radio frequency (RF), microwave, infrared (IR), etc.). This includes the training request (which is sent via RRC channel, a downlink control channel or on a data channel as described see Ma, [0136]), and also the training signal on which the training is performed (as described in [0142]: “At 1116 the BS starts the training phase 1150 by sending a training signal that includes a training sequence or training data to the UE… Non-limiting examples of channels that may be used by the BS to send training sequences or training data to UE include those discussed above with reference to FIG. 12, namely a dynamic control channel, a data channel and/or RRC channel.”). The AI/ML module is trained on a channel for communication between the UE and the base station, in addition to being trained using a channel for such communication. See Ma, [0074]: “an AI/ ML module 502,552 that is trainable in order to provide a tailored personalized air interface between the base station 170 and UE 110.”]
Ma as modified thus far does not explicitly teach the limitations that the training command is received “in a first component carrier” and that the neural network is trained “on a second component carrier.”
Wang teaches, “in a first component carrier” and “on a second component carrier” [[0175]: “the base station 120 (and/or the core network server 302 by way of the base station 120) communicates the configuration of the DNN to the UE 110 using a first component carrier, where the DNN configuration corresponds to forming a DNN for processing a second component carrier of the carrier aggregation.” That is, the first component carrier is used for communication of the configuration, which is analogous to the training command of the instant claim, while the second component carrier is for the function of the DNN, which is analogous for a function for which the neural network of the instant claim is being configured for on.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Wang by implementing the use of the different component carriers as taught in Wang for communication and DNN functionality, so as to arrive at the limitations of receiving the training command “in a first component carrier” and having the neural network be trained “on a second component carrier.” The motivation would have been to enable the neural network to process information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier (see Wang, [0026]: “at least one deep neural network (DNN) configuration for processing information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier.”).
As to claim 24, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 20, wherein the at least one processor is further configured to transmit a training command […] that indicates for the UE to apply one or more neural network training parameters to the UE to train the neural network […]. [As shown in FIG. 13, the UE and BS communicate wirelessly. See [0037]: The BSs 170 communicate with one or more of the EDs 110 over one or more air interfaces 190 a using wireless communication links (e.g. radio frequency (RF), microwave, infrared (IR), etc.). This includes the training request (which is sent via RRC channel, a downlink control channel or on a data channel as described see [0136]), and also the training signal on which the training is performed (as described in [0142]: “At 1116 the BS starts the training phase 1150 by sending a training signal that includes a training sequence or training data to the UE… Non-limiting examples of channels that may be used by the BS to send training sequences or training data to UE include those discussed above with reference to FIG. 12, namely a dynamic control channel, a data channel and/or RRC channel.”). The AI/ML module is trained on a channel for communication between the UE and the base station, in addition to being trained using a channel for such communication. See [0074]: “an AI/ ML module 502,552 that is trainable in order to provide a tailored personalized air interface between the base station 170 and UE 110.” The training may be at the UE or at both the UE and in the network (see [0142], [0145]-[0146]), either of which reads on the instant claim limitation.]
Ma as modified thus far does not explicitly teach the limitations that the training command is transmitted “in a first frequency range or a first frequency band” and that the neural network is trained “on a second frequency range or a second frequency band.”
Wang teaches, “in a first frequency range or a first frequency band” and “on a second frequency range or a second frequency band” [[0175]: “the base station 120 (and/or the core network server 302 by way of the base station 120) communicates the configuration of the DNN to the UE 110 using a first component carrier, where the DNN configuration corresponds to forming a DNN for processing a second component carrier of the carrier aggregation.” That is, the first component carrier is used for communication of the configuration, which is analogous to the training command of the instant claim, while the second component carrier is for the function of the DNN, which is analogous for a function for which the neural network of the instant claim is being configured for on. See [0025]: “training a DNN on transmitter and/or receiver processing chain operations.” Note that the two component carriers have different frequency bands. See [0156]: “the first component carrier resides in a licensed band and the second component carrier resides in an unlicensed band”; [0144]: “Image 1020 depicts a non-contiguous, inter-band configuration, where at least one component carrier used in the carrier-aggregation-communications resides in a different frequency band. For example, in various implementations, band 1 of image 1020 corresponds to licensed bands allocated to the base station 120 and band 2 of image 1020 corresponds to unlicensed bands used by the base station 120, such as frequency bands accessed through Licensed-Assisted Access (LAA).”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Wang by implementing the use of the different component carriers as taught in Wang for communication and DNN functionality, so as to arrive at the limitations of transmitting the training command “in a first frequency range or a first frequency band” for the neural network to be trained “on a second frequency range or a second frequency band.” The motivation would have been to enable the neural network to process information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier (see Wang, [0026]: “at least one deep neural network (DNN) configuration for processing information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier.”).
As to claim 25, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 20, wherein the at least one processor is futher configured to transmit a training command […] that indicates for the UE to apply one or more neural network training parameters to train the neural network […]. [As shown in FIG. 13, the UE and BS communicate wirelessly. See Ma, [0037]: The BSs 170 communicate with one or more of the EDs 110 over one or more air interfaces 190 a using wireless communication links (e.g. radio frequency (RF), microwave, infrared (IR), etc.). This includes the training request (which is sent via RRC channel, a downlink control channel or on a data channel as described see Ma, [0136]), and also the training signal on which the training is performed (as described in Ma, [0142]: “At 1116 the BS starts the training phase 1150 by sending a training signal that includes a training sequence or training data to the UE… Non-limiting examples of channels that may be used by the BS to send training sequences or training data to UE include those discussed above with reference to FIG. 12, namely a dynamic control channel, a data channel and/or RRC channel.”). The AI/ML module is trained on a channel for communication between the UE and the base station, in addition to being trained using a channel for such communication. See Ma, [0074]: “an AI/ ML module 502,552 that is trainable in order to provide a tailored personalized air interface between the base station 170 and UE 110.” The training may be at the UE or at both the UE and in the network (see Ma, [0142], [0145]-[0146]), either of which reads on the instant claim limitation.]
Ma as modified thus far does not explicitly teach the limitations that the training command is received “in a first component carrier” and that the neural network is trained “on a second component carrier.”
Wang teaches, “in a first component carrier” and “on a second component carrier” [[0175]: “the base station 120 (and/or the core network server 302 by way of the base station 120) communicates the configuration of the DNN to the UE 110 using a first component carrier, where the DNN configuration corresponds to forming a DNN for processing a second component carrier of the carrier aggregation.” That is, the first component carrier is used for communication of the configuration, which is analogous to the training command of the instant claim, while the second component carrier is for the function of the DNN, which is analogous for a function for which the neural network of the instant claim is being configured for on.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings the references combined thus far with the teachings of Wang by implementing the use of the different component carriers as taught in Wang for communication and DNN functionality, so as to arrive at the limitations of receiving the training command “in a first component carrier” for the neural network to be trained “on a second component carrier.” The motivation would have been to enable the neural network to process information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier (see Wang, [0026]: “at least one deep neural network (DNN) configuration for processing information exchanged with a user equipment (UE) over a wireless communication system using carrier aggregation that includes at least a first component carrier and a second component carrier.”).
4. Claims 13-16 and 26-27 are rejected under 35 U.S.C. 103 as being unpatentable over Ma in view of Kulkarni, Tan, and further in view of Kobayashi et al. (US 2017/0061329 A1) (“Kobayashi”).
As to claim 13, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 1, but does not teach the further limitation of the instant dependent claim.
Kobayashi teaches “wherein the RRC configuration indicates a period of time associated with one or more neural network training parameters for training the neural network.” [[0230]: “the user may wish to stop execution of a learning step that takes much time by setting a time limit.” [0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.” [0127]: “The time limit input unit 131 acquires information about the time limit of machine learning and notifies the learning control unit 135 of the time limit. The information about the time limit may be inputted by a user via the input device 112. The information about the time limit may be read from a setting file held in the RAM 102 or the HDD 103.” Note that the feature of “the neural network” is already disclosed in Ma, and Kobayashi’s techniques are applicable to machine learning models generically.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Kobayashi by modifying the configuration to include a settable time limit (and thus period of time) associated with the training, so as to arrive at the limitations of the instant claim. The motivation would have been to enable control of the learning time such that execution of learning is stopped if the training is taking too much time, as suggested by Kobayashi (see [0230] quoted above).
As to claim 14, the combination of Ma, Kulkarni, Tan, and Kobayashi teaches the apparatus of claim 13, as set forth above.
Kobayashi further teaches “wherein the RRC configuration indicates an action for the UE to perform when the period of time expires.” [[0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far so as to further arrive at the limitations of the instant dependent claim. Since the parts of Kobayashi cited above for this claim are the same as those cited in the rejection of the parent claim, the motivation for doing so is the same as the motivation given for Kobayashi in the rejection of the parent claim.
As to claim 15, the combination of Ma, Kulkarni, Tan, and Kobayashi teaches the apparatus of claim 13, as set forth above.
Kobayashi further teaches “wherein, when the period of time expires, the at least one processor is further configured to perform at least one of: cease training the neural network based on the one or more neural network training parameters, freeze layers of the neural network, or resume the training of one or more layers of the neural network.” [The first alternative of “cease training the neural network” is disclosed. See [0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far so as to further arrive at the limitations of the instant dependent claim. Since the parts of Kobayashi cited above for this claim are the same as those cited in the rejection of the parent claim 13, the motivation for doing so is the same as the motivation given for Kobayashi in the rejection of the parent claim 13.
As to claim 16, the combination of Ma, Kulkarni, Tan, and Kobayashi teaches the apparatus of claim 13, as set forth above.
Kobayashi further teaches “wherein the period of time is a periodic time, semi-persistent time, or aperiodic time for training the neural network, and wherein the memory and the at least one processor are further configured to periodically or aperiodically train the neural network based on the period of time.” [[0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.” That is, the time limit is aperiodic since it does not repeat.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far so as to further arrive at the limitations of the instant dependent claim. Since the parts of Kobayashi cited above for this claim are the same as those cited in the rejection of the parent claim 13, the motivation for doing so is the same as the motivation given for Kobayashi in the rejection of the parent claim 13.
As to claim 26, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 20, but does not teach the further limitation of the instant dependent claim.
Kobayashi teaches “wherein the RRC configuration indicates a period of time associated with the one or more neural network training parameters for training the neural network.” [[0230]: “the user may wish to stop execution of a learning step that takes much time by setting a time limit.” [0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.” [0127]: “The time limit input unit 131 acquires information about the time limit of machine learning and notifies the learning control unit 135 of the time limit. The information about the time limit may be inputted by a user via the input device 112. The information about the time limit may be read from a setting file held in the RAM 102 or the HDD 103.” Note that the feature of “the neural network” is already disclosed in Ma, and Kobayashi’s techniques are applicable to machine learning models generically.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Kobayashi by modifying the configuration to include a settable time limit (and thus period of time) associated with the training, so as to arrive at the limitations of the instant claim. The motivation would have been to enable control of the learning time such that execution of learning is stopped if the training is taking too much time, as suggested by Kobayashi (see [0230] quoted above).
As to claim 27, the combination of Ma, Kulkarni, Tan, and Kobayashi teaches the apparatus of claim 26, as set forth above.
Kobayashi further teaches “wherein the period of time is a periodic time, semi-persistent time, or aperiodic time for the UE to train the neural network.” [[0157]: “(S27) The learning control unit 135 determines whether the time that has elapsed since the start of the machine learning has exceeded the time limit specified by the time limit input unit 131. If the elapsed time has exceeded the time limit, the operation proceeds to step S28.” [0107]: “the learning time is limited and the machine learning is stopped before its completion.” That is, the time limit is aperiodic since it does not repeat.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far so as to further arrive at the limitations of the instant dependent claim. Since the parts of Kobayashi cited above for this claim are the same as those cited in the rejection of the parent claim, the motivation for doing so is the same as the motivation given for Kobayashi in the rejection of the parent claim.
5. Claims 21-22 are rejected under 35 U.S.C. 103 as being unpatentable over Ma in view of Kulkarni and Tan, and further in view of Wang et al., “Deep Learning-based CSI Feedback Approach for Time-varying Massive MIMO Channels,” arXiv:1807.11673v1 [cs.IT] 31 Jul 2018 (“Wang”).
As to claim 21, the combination of Ma, Kulkarni, and Tan teaches the apparatus of claim 20, wherein the apparatus includes a network entity for a wireless communication system or another UE, [The base station described above is a network entity for wireless communication. See also Ma, [0035]: “In FIG. 1, the RANs 120 include base stations (BSs) 170 a-170 b (generically referred to as BS 170), respectively. Each BS 170 is configured to wirelessly interface with one or more of the EDs 110 to enable access to any other BS 170, the core network 130, the PSTN 140, the internet 150, and/or the other networks 160.”] and wherein the control signaling in the second wireless transmission signals that the at least the subset of layers is to be frozen during training of the neural network [Ma, [0122]: “At 1016 the BS starts the training phase 1050 by sending a training signal that includes a training sequence or training data to the UE.” See also other forms of control signaling discussed in [0129] and [0140] as discussed in the rejection of the parent independent claim. The signal function of “that at least the subset of layers is to be frozen” is taught by Tan for the reasons discussed in the parent independent claim and is also covered by the motivation given for Tan in the rejection of the parent independent claim.].
The combination of references thus far does not teach the remaining limitation of that the subset of layers is frozen “based on the stationary or mobility condition of the UE.”
Wang teaches “based on the stationary or mobility condition of the UE.” [Second page, left column, “Observation 2” section: “UE motion during communication results in a Doppler spread, i.e., time-varying characteristics of wireless channels. With the maximum movement velocity denoted as v, coherence time can be calculated as… The CSI within ∆t is considered correlated with one other. Therefore, instead of independently recovering CSI, the BS can combine the feedback and previous channel information for the subsequent reconstruction. We set the feedback time interval as δt and place T adjacent instantaneous angular delay domain channel matrices into a channel group… The group exhibits correlation property, as long as T satisfies 0 ≤ δt · T ≤ ∆t.” The next paragraph states: In this article, we design an encoder, st = fen(Ht’’), at the UE to compress each complex-valued Ht’’ of {Ht’’}Tt=1 into an M-dimensional real-valued codeword vector st (M < N)…” In other words, the velocity of the vehicle is a factor in designing the model, specifically the hyperparameter T. This variability in the model’s design is analogous to selection of the model in Ma (see Ma, [0125]: “AI/ML module indicator(s) to indicate which AI/ML module(s) the training sequence/data is for.”) before performing subsequent operations of training; thus, by taking the velocity into account in the selection of the model, the subset of frozen layers would be “based on” the model selection in the combination of Wang with the references already cited thus far. The Examiner notes that the term “based on” is used broadly in the instant claim, without requiring a specific relationship between how different conditions affect different layers to be frozen.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention combined the teachings of the references combined thus far with the teachings of Wang by implementing the use of a velocity measure of the UE such that the model that is being trained and thus the subset of layers to be frozen in the training are “based on the stationary or mobility condition of the UE.” The motivation for doing so would have been to take into account the effect of doppler spread, which affects the operation of the model (see Wang, part cited above).
As to claim 22, the combination of Ma, Kulkarni, Tan, and Wang teaches the apparatus of claim 21, wherein the at least one processor is configured to:
transmit multiple sets of parameters for neural network training in a higher-layer signaling; [Ma, [0136]: “In some embodiments, the training request may be set to the UE via RRC signaling.” Here, the RRC signaling is a type of the limitation of “higher-layer signaling” in the absence of further limitations as to a more specific definition of this term. The Examiner interprets the instant limitation as being met if RRC is used for any part of the multiple sets of neural network training parameters.] and
transmit an indication of one of the multiple sets of parameters in at least one of a MAC-CE, DCI, or a combination thereof. [Ma, [0136]: “In some embodiments, the training request may be sent to the UE through DCI (dynamic signaling) on a downlink control channel or on a data channel. For example, in some embodiments the training request may be sent to the UE with UE specific or UE common DCI. For example, UE common DCI may be used to send a training request to all UEs or a group of UEs. In some embodiments, the training request may be set to the UE via RRC signaling. In some embodiments, the training request may include initial training setting(s)/parameter(s), such as initial NN weights.” Ma, [0210]: “Example Embodiment 131. …transmitting the AI/ML training request through downlink control information (DCI) on a downlink control channel or RRC signaling or the combination of the DCI and RRC signaling. …Example Embodiment 135…wherein the dynamic control channel includes a dynamic control information (DCI) field containing information indicating an AI/ML module that is to be trained.” That is Ma, [0210] teaches that DCI is used to indicate the training parameter of “an AI/ML module that is to be trained” and further teaches that the combination of DCI and RRC may be used.]
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. The following documents depict the state of the art.
US 20050265436 A1 teaches the use of different encoder models depending on the speed of the UE (see [0061]).
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 nonprovisional extension fee (37 CFR 1.17(a)) 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 mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to YAO DAVID HUANG whose telephone number is (571)270-1764. The examiner can normally be reached Monday - Friday 9:00 am - 5:30 pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Miranda Huang can be reached at (571) 270-7092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Y.D.H./Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124
Prosecution Insights