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 .
Response to Arguments
Applicant's arguments filed 2/12/2026 have been fully considered but they are not persuasive.
Regarding applicants arguments for U.S.C. 103 rejection on page 14-5 “As agreed upon during the telephone interview, the amended claims are patentable over the cited references for the following reasons.
The cited references to not teach or suggest at least "obtaining a cost function to quantify a difference between an expected output and the quantized output; computing gradients with respect to parameters from a last layer to a first layer of the machine learning model based on the cost function," as recited in amended independent claim 1.” No agreement was reached regarding obtaining the cost function. It was agreed that further consideration and searching would be required to determine any allowable subject matter. As the reference in question during the interview was Yong et al CNQ: Compressor Based Non-Uniform Quantization to Deep Neural Networks. Chinese Journal of Electronics, 2020, 29(6) 1126-1133 (“Yong”). The discussion was centered on the issued if Yong thought a backpropagation and a cost function as discussed the applicant agenda (see Examiner Interview Summary Record data 2/13/2023). The examiner stated that further search and consideration would be required to determine any allowability. No agreement was made regarding allowability over the cited art only that the amendments would require further consideration.
Applicant furth argues in page 15 that “The Office alleges that page 1129 of Yong teaches computing gradients with respect to parameters from a last layer to a first layer of the machine learning model based on the quantized output. (Office Action, page 13). However, page 1129 of Yong describes optimizing quantized weights, and updating weights based on the gradients of the quantized weights. ….
For at least these reasons, amended independent claim 1 and its dependent claims are patentable over the cited references. Amended independent claims 9 and 16 and their dependent claims are patentable over the cited references for similar reasons as amended independent claim 1. Accordingly, Applicant respectfully requests withdrawal of the rejections under 35 U.S.C. § 103.” However the applicant argues amended limitations that have not been examined and thus the argument is moot and not convincing.
Claim Objections
Claims 4 and 12 are objected to because of the following informalities:
Claim 4 has the following informalities:
Claim 4 depends on claim 1 and the limitation “the edge node is a vehicle” a vehicle is already mentioned in claim 1 thus it should be the vehicle. The limitation “a server” should be the server.
(Examiner Note: Claim 4 limitation of a vehicle is found in claim 1. The amendments to claim 1 cause issues with claim limitations of claim 4 and fail to further limit the claims. The issue could lead to indefinite issues as the two recitations of a vehicle exist in the claims.)
Claim 12 has the following informalities:
Claim 12 depends on claim 9 and the limitation “a server” in claims should the server as it already mentioned in claim 9.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(d):
(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
The following is a quotation of pre-AIA 35 U.S.C. 112, fourth paragraph:
Subject to the following paragraph [i.e., the fifth paragraph of pre-AIA 35 U.S.C. 112], a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
Claim 4 and 12 are rejected under 35 U.S.C. 112(d) or pre-AIA 35 U.S.C. 112, 4th paragraph, as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends, or for failing to include all the limitations of the claim upon which it depends. Claim 4 depends on claims 1 as the amended limitation now includes:
PNG
media_image1.png
125
701
media_image1.png
Greyscale
The amended limitations on claim 1 are found in claim 4. Claim 4 does not further limit the limitations mentioned above as the limitation of the edge node is a vehicle causes problems with the mentioned vehicle in the amended section of claim 1. The only limitation impose by claim 4 is that the edge node is a vehicle however claim 1 mentions a vehicle that is the intended to be an edge node. Thus claim 4 does not further limit claim 1.
Claim 12 is analogous to claim 4 and has a similar issued where the controller is to perform the same limitations mentioned in the independent claim 9 of
PNG
media_image2.png
129
702
media_image2.png
Greyscale
. Claim 12 performs the same operations as claim 9. Applicant may cancel the claim(s), amend the claim(s) to place the claim(s) in proper dependent form, rewrite the claim(s) in independent form, or present a sufficient showing that the dependent claim(s) complies with the statutory requirements.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 2, 4, 6, 8-10, 12, 13, 15-17 are rejected under 35 U.S.C. 103 as being unpatentable over JAEHONG YOON, et al., "Bitwidth Heterogeneous Federated Learning with Progressive Weight Dequantization," Proceedings of the 39th International Conference on Machine Learning (PMLR), July 2022, vol. 162 (“Yoon”) in view of Jung, Sangil, et al. "Learning to quantize deep networks by optimizing quantization intervals with task loss." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019 (“Juang”) and further in view of Choi et al. (US20220245527A1) and Tu et al. (US12223734B2) (“Tu”).
Regarding claim 1 and analogous claims 9 and 16, Yoon teaches a method for training a machine learning model in an edge node of a federated learning system, the method comprising (Yoon Page 2 Figure 1,
PNG
media_image3.png
371
467
media_image3.png
Greyscale
):
inputting a data point into the machine learning model including parameters quantized based on a first quantization level to obtain an output (Yoon Page 4, 3.2. Local Computation for Limited Bitwidth Clients, We assume that the local clients are edge devices that perform limited bitwidth computations according to their hardware specifications. Float32 clients perform regular fullprecision training, but quantized clients, such as Int6, Int8, and Int16, perform training using low-precision integer operations, following Wu et al. (2018).
Let
q
l
and
a
l
be s-bit quantized weights and activations at layer l for a client, respectively. We denote the convolution operator as *. Since convolution involves multiplication between the weights and the activations, naively computing
q
l
*
a
l
-
1
requires the hardware to support fast multiplication of two s-bit integers. To relax this requirement, we ternarize the quantized weights before the convolution operation: [including parameters quantized based on a first quantization level to obtain an output]
Page 6-7 5. Experiments Datasets, We validate our method against the relevant FL methods under several BHFL scenarios, with varying bitwidth configurations of participating clients. We use the widely used benchmark dataset for federated learning methods, CIFAR-10 to validate our method following the IID experimental settings of the existing works (Reisizadeh et al., 2020; Haddadpour et al., 2021). CIFAR-10 is a image classification dataset that consists of 10 object classes each of which has 5,000 training instances and 1,000 test instances [inputting a data point into the machine learning model].);
quantizing the output based on the first quantization level and a non-uniform quantization scheme ( Yoon Page 7-8, 5.1 Quantitative Evaluation, We validate our methods under multiple BHFL scenarios with heterogeneous proportions of bitwidths among the clients. We first report the experimental results with 50% of Int8 and 50% Float32 clients (Left) and 80% of Int8 and 20% Float32 clients (Right) in Table 5. Int8 models in FedAvg obtain superior performance to local training, where each client trains independently on its local task due to positive knowledge transfer from the full-precision weights. This is also evident in the poor performance of Int8 clients in FedGroupedAvg which demonstrates that FL only with low-bitwidth clients is highly limited due to the lack of information the low-bit weights provide. QPC-based FL methods, FedPAQ, FedCOM, and FedCOMGATE, communicate the quantized form of accumulated gradients at each round, and the server adds the accumulated gradients to the global model at each round, before broadcasting it to the clients [quantizing the output based on the first quantization level].
PNG
media_image4.png
321
550
media_image4.png
Greyscale
[and a non-uniform quantization scheme;]);
[[and]] updating the machine learning model using the quantized gradients (Yoon page 6, Algorithm 2,
PNG
media_image5.png
359
566
media_image5.png
Greyscale
[ using the quantized gradients]
Page 7-8, 5.1 Quantitative Evaluation, We validate our methods under multiple BHFL scenarios with heterogeneous proportions of bitwidths among the clients. We first report the experimental results with 50% of Int8 and 50% Float32 clients (Left) and 80% of Int8 and 20% Float32 clients (Right) in Table 5. Int8 models in FedAvg obtain superior performance to local training, where each client trains independently on its local task due to positive knowledge transfer from the full-precision weights. This is also evident in the poor performance of Int8 clients in FedGroupedAvg which demonstrates that FL only with low-bitwidth clients is highly limited due to the lack of information the low-bit weights provide. QPC-based FL methods, FedPAQ, FedCOM, and FedCOMGATE, communicate the quantized form of accumulated gradients at each round, and the server adds the accumulated gradients to the global model at each round, before broadcasting it to the clients [updating the machine learning model]));
transmitting the updated machine learning model to a server;
receiving an aggregated machine learning model from the server (Yoon page 3, 3.1. Problem statements, Federated Learning In a standard Federated Learning (FL) scenario (McMahan et al., 2017; Chen et al., 2019), each client trains the local model on the private data and periodically transmits the model parameters to the central server, where the models are aggregated and broadcasted back to the clients. Let N different clients C = {
c
1
,
…
,
c
N
}participate in an FL system. Given training samples
x
n
and its corresponding labels
y
n
, we suppose that a client
c
n
solves a local optimization problem
L
n
=
C
E
(
f
x
n
;
w
n
;
y
n
)
, where CE is a cross-entropy loss and
f
(
.
;
w
n
is a neural network of client
c
n
parameterized by
w
n
. At each communication round r, clients
C
(
r
)
⊆
C
send the model parameters to the central server, and the server aggregates received weights, for exampling by averaging their weights.
page 5, Figure 3,
PNG
media_image6.png
289
958
media_image6.png
Greyscale
(i.e. at Figure 3 (a) [transmitting the updated machine learning model to a server; ])) [receiving an aggregated machine learning model from the server;]);
Yoon does not explicitly teach obtaining a cost function to quantify a difference between an expected output and the quantized output;
computing gradients with respect to parameters from a last layer to a first layer of the machine learning model based on the quantized output cost function;
quantizing the gradients based on a second quantization level and the non-uniform quantization scheme;
and operating a vehicle to drive autonomously by inputting data sensed by the vehicle to the aggregated machine learning model to detect objects and adjusting vehicle parameters based on the detected objects.
Jung teaches obtaining a cost function to quantify a difference between an expected output and the quantized output (Jung Page 4352,
PNG
media_image7.png
109
1140
media_image7.png
Greyscale
page 4354 3.2. Trainable quantization interval para 1, To design the quantizers, we consider two operations: clipping and pruning (Fig. 1 (a)). The underlying idea of clipping is to limit the upper bound for quantization [27, 4]. Decreasing the upper bound increases the quantization resolution within the bound so that the accuracy of the low bit-width network can increase. On the other hand, if the upper bound is set too low, accuracy may decrease because too many values will be clipped. Thus, setting a proper clipping threshold is crucial for maintaining the performance of the networks. Pruning removes low-valued weight parameters [7]. Increasing pruning threshold helps to increase the quantization resolution and reduce the model complexity, while setting pruning threshold too high can cause the performance degradation due to the same reason as the clipping scheme does.
Page 4354, Algorithm 1
PNG
media_image8.png
655
555
media_image8.png
Greyscale
));
computing gradients with respect to parameters from a last layer to a first layer of the machine learning model based on the cost function (Jung Page 4352,
PNG
media_image7.png
109
1140
media_image7.png
Greyscale
page 4354 3.2. Trainable quantization interval para 1, To design the quantizers, we consider two operations: clipping and pruning (Fig. 1 (a)). The underlying idea of clipping is to limit the upper bound for quantization [27, 4]. Decreasing the upper bound increases the quantization resolution within the bound so that the accuracy of the low bit-width network can increase. On the other hand, if the upper bound is set too low, accuracy may decrease because too many values will be clipped. Thus, setting a proper clipping threshold is crucial for maintaining the performance of the networks. Pruning removes low-valued weight parameters [7]. Increasing pruning threshold helps to increase the quantization resolution and reduce the model complexity, while setting pruning threshold too high can cause the performance degradation due to the same reason as the clipping scheme does.
Page 4354, Algorithm 1
PNG
media_image8.png
655
555
media_image8.png
Greyscale
[ based on the quantized output cost function]
Page 4354 3.2. Trainable quantization interval para 5, We use stochastic gradient descent for optimizing the parameters of both the weights and the quantizers. The transformers are piece-wise differentiable, and thus we can compute the gradient with respect to the interval parameters
d
∆
and γ. We use straight-through-estimator [2, 27] for the gradient of the discretizers [computing gradients with respect to parameters from a last layer to a first layer of the machine learning mode]);
Yoon and Jang are considered to be analogous to the claim invention because they are in the same field of quantized machine learning methods. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Yoon to incorporate the teachings of Jung and disclose using a cost function to determine loss between the expected output and the quantized out. Doing so to minimize the task loss (Jang page 4357 5. Conclusion, We proposed a novel trainable quantizer with parameterized quantization intervals for training low bit-width networks. Our trainable quantizer performs simultaneous pruning and clipping for both weights and activations, while maintaining the accuracy of the full-precision network by learning appropriate quantization intervals. Instead of minimizing the quantization error with respect to the weights/activations of the full-precision networks as done in previous work, we train the quantization parameters jointly with the weights by directly minimizing the task loss. As a result, we achieved very promising results on the large scale ImageNet classification dataset.)
Choi teaches quantizing the gradients based on a second quantization level and the non-uniform quantization scheme (Choi Para 0083 In some cases, to efficiently transmit gradient data to the server 205, the workers 215 may compress the gradient data according to some quantization level (e.g., quantization level of 32 bits, 64, bits, 128 bits, ... , 1024 bits, or some other quantity of bits). However, channel conditions (e.g., link budgets, channel bandwidths, channel qualities, or other channel conditions) between the workers 215 and the server 205 may differ. For example, the worker 215-a may have relatively better channel conditions (e.g., higher link budget, channel bandwidth, channel quality, or some other charmel condition) than the worker 215-b [the non-uniform quantization scheme].
Choi Para 0162, In some examples, the duration component 1040 may be configured as or otherwise support a means for calculating a duration associated with the communicating of the compressed gradient data. In some examples, the quantization component 1025 may be configured as or otherwise support a means for determining a second quantization level for second gradient data output by the machine learning model based on the duration satisfying a threshold duration [quantizing the gradients based on a second quantization level].);
Yoon and Choi are considered to be analogous to the claim invention because they are in the
same field of machine learning using quantized methods. Therefore, it would have been obvious to
someone of ordinary skill in the art before the effective filling date of the claimed invention to have
modified Yoon to incorporate the teachings of Choi and disclose quantizing gradients based on a second
quantization level. Doing so to reduce latency associated with training the global model (Choi para 0077
line 1-8, To reduce latency associated with training the global model and to ensure global convergence
of the global model, the server may adaptively select and indicate quantization levels for each UE 115 to
use to compress respective gradient data. For example, the server may determine a quantization level
for a UE 115 of the set of UEs 115 to use to compress gradient data output by a local model of the UE
115.)
Tu teaches and operating a vehicle to drive autonomously by inputting data sensed by the vehicle to the aggregated machine learning model to detect objects and adjusting vehicle parameters based on the detected objects (Tu
PNG
media_image9.png
778
1070
media_image9.png
Greyscale
Col 18 line 65-67 Col 19 line 1-5, Furthermore, although aspects of the present disclosure focus on the application of training techniques described herein to object detection models utilized in autonomous vehicles, the systems and methods of the present disclosure can be used to train any machine-learned model. Thus, for example, the systems and methods of the present disclosure can be used to train machine-learned models configured for image processing, labeling, etc.
Col 22 line 6-18, In some implementations, the sensor data 155 can be indicative of one or more objects within the surrounding environment of the vehicle 105. The object(s) can include, for example, vehicles, pedestrians, bicycles, and/or other objects. The object(s) can be located in front of, to the rear of, to the side of, above, below the vehicle 105, etc. The sensor data 155 can be indicative of locations associated with the object(s) within the surrounding environment of the vehicle 105 at one or more times. The object(s) can be static objects (e.g., not in motion) and/or dynamic objects/actors (e.g., in motion or likely to be in motion) in the vehicle's environment. The sensor(s) 135 can provide the sensor data 155 to the autonomy computing system 140.
Col 24 line 10-20, he autonomy computing system 140 can perform various functions for autonomously operating the vehicle 105. For example, the autonomy computing system 140 can perform the following functions: perception 170A, prediction 170B, and motion planning 170C. For example, the autonomy computing system 130 can obtain the sensor data 155 via the sensor(s) 135, process the sensor data 155 (and/or other data) to perceive its surrounding environment, predict the motion of objects within the surrounding environment, and generate an appropriate motion plan through such surrounding environment [and operating a vehicle to drive autonomously].
Col 24 line 49-67 Col 25 line 1-12, The planned vehicle motion trajectories can indicate the path the vehicle 105 is to follow as it traverses a route from one location to another. Thus, the vehicle computing system 110 can take into account a route/route data when performing the motion planning function 170C.
The motion planning system 180 can implement an optimization algorithm, machine-learned model, etc. that considers cost data associated with a vehicle action as well as other objective functions ( e.g., cost functions based on speed limits, traffic lights, etc.), if any, to determine optimized variables that make up the motion plan. The vehicle computing system 110 can determine that the vehicle 105 can perform a certain action (e.g., pass an object, etc.) without increasing the potential risk to the vehicle 105 and/or violating any traffic laws ( e.g., speed limits, lane boundaries, signage, etc.). For instance, the vehicle computing system 110 can evaluate the predicted motion trajectories of one or more objects during its cost data analysis to help determine an optimized vehicle trajectory through the surrounding environment. The motion planning system 180 can generate cost data associated with such trajectories. In some implementations, one or more of the predicted motion trajectories and/or perceived objects may not ultimately change the motion of the vehicle 105 (e.g., due to an overriding factor). In some implementations, the motion plan may define the vehicle's motion such that the vehicle 105 avoids the object(s ), reduces speed to give more leeway to one or more of the object(s), proceeds cautiously, performs a stopping action, passes an object, queues behind/in front of an object, etc [by inputting data sensed by the vehicle to the aggregated machine learning model to detect objects and adjusting vehicle parameters based on the detected objects]).
Yoon and Tu are considered to be analogous to the claim invention because they are in the
same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Yoon to incorporate the teachings of Tu and disclose deploy the model to train a model for object detection used in autonomous vehicles. Doing so to train any machine learning configured for image processing and labeling and increase the safety of passengers of autonomous vehicles (Col 2 line 48-53, The autonomous vehicle technology described herein can help improve the safety of passengers of an autonomous vehicle, improve the safety of the surroundings of the autonomous vehicle, improve the experience of the rider and/or operator of the autonomous vehicle, as well as provide other improvements as described herein.
Col 18 line 65-67 Col 19 line 1-5, Furthermore, although aspects of the present disclosure focus on the application of training techniques described herein to object detection models utilized in autonomous vehicles, the systems and methods of the present disclosure can be used to train any machine-learned model. Thus, for example, the systems and methods of the present disclosure can be used to train machine-learned models configured for image processing, labeling, etc).
Regarding claim 2 and analogous claims 10 and 17, Yoon in view of Jang, Choi and Tu teach the computer-implemented method of claim 1 and analogous 9 and 16.
Yoon, Jang, Choi and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon teaches wherein the first quantization level is determined based on at least one of a memory footprint of the edge node, a computation power of the edge node, and a communication bandwidth between the edge node and a server (Yoon Bitwidth page 1-2 1. Introduction para 3 line 9-19, The scenario allows the participation of local clients with various hardware infrastructures, some of them using models built under lightweight devices based on low-bitwidth hardware operations using FPGA, ASIC, Raspberry Pi, or Edge GPUs. Here, BHFL enables local devices with different hardware specifications to participate in a single federated learning framework without the need for uniformity of the infrastructure, enhancing the pool of devices that could participate in collaborative learning. We refer to this practical FL scenario as Bitwidth Heterogeneous Federated Learning (BHFL), which we illustrate in Figure 1.
page 3, Bitwidth heterogeneous federated learning
The majority of existing FL methods assume full-precision operations for local clients, even when they consider device heterogeneity. However, the participating devices have largely hetero generous bitwidths according to their hardware specifications To this end, we introduce a practical FL scenario, named Bitwidth Heterogeneous Federated Learning (BHFL), in which we relax the strong assumption that all clients are capable of full-precision floating-point operations. We represent n-th local client as a tuple of the model weights and
w
n
the corresponding hardware bitwidth information ,
s
n
ϵ
S
where S is a set of available bitwidth specifications for local S hardware devices. Then, the set of clients, can be represented as follows: g. At each
C
=
{
w
1
,
s
1
,
…
,
(
w
N
,
s
N
}
round of communication, a central server receives client tuples and aggregates model parameters which might be quantized in various levels depending on hardware specifications (i.e. wherein the first quantization level is determined based on a computation power of the edge node)).
Regarding claim 4 and analogous claim 12, , Yoon in view of Jang, Choi and Tu teach the computer-implemented method of claim 1 and analogous 9.
Yoon, Jang and Choi and Ying are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Choi teaches wherein the edge node is a vehicle, and the method further comprises: transmitting the updated machine learning model to a server; receiving an aggregated machine learning model from the server (Choi para 0056, A UE 115 may include or may be referred to as a mobile device, a wireless device, a remote device, a handheld device, or a subscriber device, or some other suitable terminology, where the "device" may also be referred to as a unit, a station, a terminal, or a client, among other examples. AUE 115 may also include or may be referred to as a personal electronic device such as a cellular phone, a personal digital assistant (PDA), a tablet computer, a laptop computer, or a personal computer. In some examples, a UE 115 may include or be referred to as a wireless local loop (WLL) station, an Internet of Things (IoT) device, an Internet of Everything (IoE) device, or a machine type communications (MTC) device, among other examples, which may be implemented in various objects such as appliances, or vehicles [wherein the edge node is a vehicle]
Para 0093 At 318, the server 305 may transmit, to the worker 310, information for updating parameters ( e.g., estimates or weights) of the local model at a first time. In some cases, the information may be a first estimate (x_i) for the machine learning model at the worker 310. For example, the information may include the weights and/ or parameters corresponding to the current global model [receiving an aggregated machine learning model from the server;].
Para 0097, At 326, the worker 310 may transmit the compressed gradient data to the server 305. In some examples, the worker 310 may transmit an indication of the quantization level used to compress the gradient data [transmitting the updated machine learning model to a server;] (Examiner Note: The worker transmits the gradients of its updated model to the server as done in Federated Learning methods)));
However Tu teaches operating the vehicle to drive autonomously using the aggregated machine learning model (Tu Col 24 line 49-67 Col 25 line 1-12, The planned vehicle motion trajectories can indicate the path the vehicle 105 is to follow as it traverses a route from one location to another. Thus, the vehicle computing system 110 can take into account a route/route data when performing the motion planning function 170C.
The motion planning system 180 can implement an optimization algorithm, machine-learned model, etc. that considers cost data associated with a vehicle action as well as other objective functions ( e.g., cost functions based on speed limits, traffic lights, etc.), if any, to determine optimized variables that make up the motion plan. The vehicle computing system 110 can determine that the vehicle 105 can perform a certain action (e.g., pass an object, etc.) without increasing the potential risk to the vehicle 105 and/or violating any traffic laws ( e.g., speed limits, lane boundaries, signage, etc.). For instance, the vehicle computing system 110 can evaluate the predicted motion trajectories of one or more objects during its cost data analysis to help determine an optimized vehicle trajectory through the surrounding environment. The motion planning system 180 can generate cost data associated with such trajectories. In some implementations, one or more of the predicted motion trajectories and/or perceived objects may not ultimately change the motion of the vehicle 105 (e.g., due to an overriding factor). In some implementations, the motion plan may define the vehicle's motion such that the vehicle 105 avoids the object(s ), reduces speed to give more leeway to one or more of the object(s), proceeds cautiously, performs a stopping action, passes an object, queues behind/in front of an object, etc [operating the vehicle to drive autonomously]).
Regarding claim 6 and analogous 13, Yoon in view of Jang, Choi and Tu teach the computer-implemented method of claim 1 and analogous 9 and 16.
Yoon, Jang, Choi and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon further teaches wherein the machine learning model is a convolutional neural network (Yoon Page 8 5.2 Qualitative Analysis, The effect of progressive weight dequantization To further analyze the role of the progressive weight dequantizer in our method, we visualize stepwise distributions of reconstructed weights using our dequantizer. For ease of interpretation, we use the initial input as the quantized weight from the last convolution layer in the Int8 client for training on CIFAR-10, where the result is illustrated in Figure 6. As low-bitwidth clients transmit ternarized model weights to the central server, the distribution of input weights to the dequantizer is visualized in three peaks.
Page 13 C. Training of Progressive Weight Dequantizer, Construction of the weights dataset. Given the local model weights w, we construct the weight datasets to learn the progressive weight dequantizer. Since layers in deep neural networks is often composed of the weights with different dimensionality each other, we split them into the uniformly-sized subweights. As following common structures for CNN models that mostly composed of a number of convolution layers, such as VGG (Simonyan & Zisserman, 2015) and ResNet (He et al., 2016), we basically utilize convolution weights with a filter size of 3 _ 3 and input dimension is 64 or larger to construct the weight dataset. That is, we use all convolution weights except the weights from the first layer (input dimension is the channel of Image, 3). We split the weights at each convolution layer to partial modules with the shape of 64_64_3_3, (e.g., the weights with the shape of 256_128_3_3 is splitted to 4_2 = 8 different modules. Next, we reshape each module sized by 24_24 with 64 channels (i.e., 64_24_24) to as illustrated in Figure 8. [convolutional neural network]).
Regarding claim 8 and analogous 15, Yoon in view of Jang, Choi and Tu teach the computer-implemented method of claim 1 and analogous 9 and 16.
Yoon, Jang, Choi and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Choi teaches further comprising: quantizing parameters of the updated machine learning model according to a third quantization level; and transmitting the quantized parameters of the updated machine learning model to a server (Choi Figure 3
PNG
media_image10.png
807
456
media_image10.png
Greyscale
Para 0107, At 340, the server 305 may transmit an indication of the second quantization level to the worker 310. In some examples, the indication may identify the second quantization level from the set of quantization levels transmitted at 338. In some examples, the indication may identify a subset of quantization levels of the set of quantization levels, for example, if the server 305 determined a set of quantization levels at 334, where the subset of quantization levels corresponds to the determined set of quantization levels. In some cases, the server 305 may transmit the next information, the set of quantization levels, and the indication of the second quantization level in any combination of one or more messages.
Para 0110 line 1-4, At 346, the worker 310 may transmit the compressed second gradient data to the server 305. In some examples, the worker 310 may transmit an indication of each quantization level used to compress the gradient data [transmitting the quantized parameters of the updated machine learning model to a server.].
para 0112, The server 305 and the worker 310 may continue to train the global model using one or more of the above operations. For example, the server 305 and the worker 310 may repeat any combination of 330 through 348 to iteratively train and update the global model. (i.e. the process will continue and a to a third quantization level will be used to quantized parameters of the updated machine learning model)).
Claim(s) 3, 11, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Yoon in view of Jang, Choi, Tu and further in view of Li et al. (US20220374689A1) (“Li”).
Regarding claim 3 and analogous claims 11 and 18, Yoon in view of Jang, Choi, and Tu teach the computer-implemented method of claim 1 and analogous 9 and 16.
Yoon, Jang, Choi, and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon does not explicitly teach wherein the second quantization level is determined based on at least one of a memory footprint of the edge node, and a computation power of the edge node.
Li teaches wherein the second quantization level is determined based on at least one of a memory footprint of the edge node, and a computation power of the edge node (Li para 0103, Compared with the previous first quantization scheme (in which a single shared zero point value z.sub.w, is shared by weights and input values), using separate zero point values z.sub.x and z.sub.w in this second quantization scheme may help to reduce to quantization precision loss and help to improve accuracy, at the expense of an increase in computations (which may be acceptable in some applications, and which is generally still more efficient than computing the quantized inner product operator) [and a computation power of the edge node]).
Yoon and Li are considered to be analogous to the claim invention because they are in the same field of machine learning methods. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Yoon to incorporate the teachings of Li and disclose a second quantization level based on computation power. Doing so to help reduce the quantization precision loss and improve accuracy (Li para 0103, Compared with the previous first quantization scheme (in which a single shared zero point value z.sub.w, is shared by weights and input values), using separate zero point values z.sub.x and z.sub.w in this second quantization scheme may help to reduce to quantization precision loss and help to improve accuracy, at the expense of an increase in computations (which may be acceptable in some applications, and which is generally still more efficient than computing the quantized inner product operator).
Claim(s) 5 are rejected under 35 U.S.C. 103 as being unpatentable over Yoon in view of Jang, Choi, Tu and further in view of Z. Yang, X. Zhang, D. Wu, R. Wang, P. Zhang and Y. Wu, "Efficient Asynchronous Federated Learning Research in the Internet of Vehicles," in IEEE Internet of Things Journal, vol. 10, no. 9, pp. 7737-7748 (“Yang”).
Regarding claim 5, Yoon in view of Jang, Choi, and Tu teach the computer-implemented method of claim 1 and analogous claims 9 and 16.
Yoon, Jang and Choi and Ying are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon does not explicitly teach wherein the edge node is an edge server, and the method further comprises: transmitting the updated machine learning model to a cloud server; receiving an aggregated machine learning model from the cloud server; and transmitting the aggregated machine learning model to one or more vehicles.
However Yang teaches wherein the edge node is an edge server, and the method further comprises: transmitting the updated machine learning model to a cloud server; receiving an aggregated machine learning model from the cloud server; and transmitting the aggregated machine learning model to one or more vehicles (Yang page 7737 1. Introduction para 2 line 6-13,
In a typical two-layer (cloud server and client) FL system the cloud server act as the only parameter aggregation server receiving local updates from clients distributed in different regions. After receiving all the parameter updates, the cloud server aggregates the parameters by using the defined method and returns the result to the clients for the next iteration. This process is sequentially looped until the global model converges or a predefined accuracy threshold is reached [receiving an aggregated machine learning model from the cloud server;].
Page 7740,
PNG
media_image11.png
325
575
media_image11.png
Greyscale
[transmitting the aggregated machine learning model to one or more vehicles] [wherein the edge node is an edge server]
Page 7743 B. Hierarchical Asynchronous Aggregation Scheme Para 12, After receiving the gradient parameters uploaded by the client, the edge server starts to perform asynchronous aggregation based on the fresh weight factor, and the cloud server waits for the gradient transmission from the edge server to arrive and performs synchronous aggregation [transmitting the updated machine learning model to a cloud server]).
Yoon and Yang are considered to be analogous to the claim invention because they are in the same field of Federated Learning methods. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Yoon to incorporate the teachings of Yang and disclose a federated learning environment using clod and edge servers. Doing so to allow multiple edge server to perform partial model aggregation for faster model training and better communication-computations tradeoffs (Yang page 7738 para 5 line 6-15, A device-edge-cloud layered system allows multiple edge servers to perform partial model aggregation for faster model training and better communication-computation tradeoffs [14], [15]. That is, in the case of IoV, hierarchical FL (HFL) is a more suitable training model with low delay and high reliability. In HFL, the vehicle as a client node uses local data for model training, the road-side unit (RSU) as an edge server for local aggregation, and the base station (BS) as a cloud server for global aggregation.
Regarding claim 19, Yoon in view of Jang, Choi, and Tu, and Yang teach the computer-implemented method of claim 1 and analogous claims 9 and 16.
Yoon, Jang, Choi, and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon and Yang are combine in the same rational as set forth above with respect to claim 5.
Yang teaches wherein the plurality of edge nodes are a plurality of edge servers (Page 7743 A. Simulation Setting para 1, To evaluate the performance of the EHAFL method on real data sets, we conduct experiments in a hierarchical federation consisting of cloud server, five edge servers [edge nodes are a plurality of edge servers], and 100 vehicle nodes. In real-world scenarios, the high-speed mobility of vehicle nodes make the edge nodes interacting with it may change with the increase of communication rounds).
Regarding claim 20, Yoon in view of Jang, Choi, and Tu are combine in the same rational as set forth above with respect to claim 19.
Yoon, Jang, Choi, and Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon and Yang are combine in the same rational as set forth above with respect to claim 5.
Yang further teaches wherein each of the plurality edge servers communicate with a plurality of vehicles, and each of the plurality of vehicles includes a controller programmed to train another machine learning model received from corresponding edge server (Yang Page 7740 Figure 2,
PNG
media_image12.png
323
580
media_image12.png
Greyscale
[wherein each of the plurality edge servers communicate with a plurality of vehicles])
Yang Page 7740, IV. SCHEME DESIGN, The essential requirement to reduce the communication costs of FL is to reduce the amount of data uploaded/downloaded as much as possible while ensuring the performance of the global model. To this end, we propose an efficient hierarchical asynchronous FL (EHAFL) method with variable bit length, including DAQC and AHAFL. The DAQC occurs during the quantization and encoding process before the client transmits, and the AHAFL is the asynchronous weighted average. Fig. 2 shows the specific interaction process between the client and the edge server. Since the relative positions of the edge server and the cloud server are fixed, and the quality of the downlink is better than that of the uplink and more stable, so we only consider the communication uplink from the client to the edge server [and each of the plurality of vehicles includes a controller programmed to train another machine learning model received from corresponding edge server]. (Examiner Note: Each client independently runs and trains its own machine learning model using a controller program. The client also received from an edge server another ml model.))
Claim(s) 7 and analogous 14 are rejected under 35 U.S.C. 103 as being unpatentable over Yoon in view of Jang, Choi, Tu and further in view of Jiawei Jiang, Fangcheng Fu, Tong Yang, and Bin Cui. 2018. SketchML: Accelerating Distributed Machine Learning with Data Sketches. In Proceedings of the 2018 International Conference on Management of Data (SIGMOD '18). Association for Computing Machinery, New York, NY, USA, 1269–1284 (“Jiang”).
Regarding claim 7 and analogous 14, Yoon in view of Jang and Choi teach the computer-implemented method of claim 1 and analogous 9.
Yoon in view of Jang, Choi, Tu are combine in the same rational as set forth above with respect to claim 1 and analogous claims 9 and 16.
Yoon does not explicitly teach wherein the non-uniform quantization scheme quantizes the output based on quantile values.
However Jiang teaches wherein the non-uniform quantization scheme quantizes the output based on quantile values (Jiang Page 1271, Quantile sketch uses a small data structure to approximate the exact distribution of item value in a single pass over the items. The main component of quantile sketch is the quantile summary which consists of a small number of points from the original items [12]. Two major operations, merge and prune, are defined for quantile summary. The merge operation combines two summaries into a merged summary, while the prune operation reduces the number of summaries to avoid exceeding the maximal size. Since there are m quantile summaries in a quantile sketch, the computation complexity is O(N) and the space complexity is O(m). In contrast to the brute-force sorting, the total cost is reduced significantly. Meanwhile, the existing quantile sketches also provide solid error bounds. For example, Yahoo DataSketches [41] guarantees 99% correctness when m = 256. Once a quantile sketch is built for these one billion items, the quantile summaries are used to give approximate answers to any quantile query q ∈ [0, 1]. For example, a query of 0.5 refers to the median value of the items, and the quantile sketch returns an estimated value for the item ranking 0.5 billion. With the same manner, a query of 0.01 returns an estimated value for the item ranking 10 million [quantizes the output based on quantile values].
Page 1273 3.1 Overview of The Framework, Summary. Through an in-depth anatomy of existing quantification methods, we find that they cannot capture the distribution property of gradients. We therefore investigate nonuniform quantification methods. By designing a technique that combines quantile sketch and bucket sort, we successfully encode gradient values to small binary numbers and achieve self-adaption to data nonuniformity. The key-value pair (kj ,vj ) is encoded to (kj ,b(vj )) where b(vj ) denotes the binary bucket index. In practice, we find that q = 256 is often enough to obtain comparable prediction accuracy [wherein the non-uniform quantization scheme]).
Yoon and Jiang are considered to be analogous to the claim invention because they are in the same field of machine learning methods. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Yoon to incorporate the teachings of Jiang and use a non-uniform quantization scheme based on quantile values. Doing so to provide good ability of scalability (Jiang Page 11, In summary, SketchML outperforms Adam and ZipML on a range of ML algorithms and datasets. SketchML consumes remarkably less time to execute an epoch. Although it needs more epoch to get converged, the overall performance still surpasses the other two competitors. Besides, SketchML reveals a good ability of scalability.
Pertinent Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Hubara et al. "Quantized neural networks: Training neural networks with low precision weights and activations." journal of machine learning research 18.187 (2018): 1-30 (“Hubara”) teaches a quantization method with a cost function by performing a backward pass gradient calculation (see Algorithm 1).
Conclusion
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 ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
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, Michael J. Huntley can be reached at (303) 297-4307. 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.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights