Partial quantization to implement fully quantized models on edge devices

By partially quantifying the weight and activation functions of the machine learning model on a portable device, combined with representative data set training, the complete quantization problem on a portable device is solved, and safe and flexible model training and transfer learning are achieved.

CN118940792APending Publication Date: 2024-11-12NXP USA INC
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202310532613.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-11
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing machine learning models are difficult to achieve full quantization on portable devices due to computing power and storage limitations, and require communication with cloud or more powerful processors, resulting in data security issues and limited operational flexibility.

Method used

Through partial quantization technology, the weights and activation functions of multiple layers of the machine learning model are quantified, frozen and updated quantized weights and activation functions are generated, and trained in combination with representative data sets to form a new quantized model, which is directly in A fully quantized model is realized on edge devices.

Benefits of technology

It realizes the training and quantification of machine learning models on edge devices without external resources, avoids data security risks, promotes transfer learning, and expands the applicability and flexibility of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940792A_ABST
    Figure CN118940792A_ABST
Patent Text Reader

Abstract

A method for partial quantization to implement a fully quantized model includes quantizing a plurality of weights and respective activation functions from each of a plurality of respective layers of an original machine learning model (MLM) to generate a quantized MLM including a plurality of frozen quantized weights. The plurality of frozen quantized weights are extracted from at least one frozen layer of the layers of the quantized MLM. The plurality of weights are quantized in accordance with at least one updated layer of the updated MLM to generate a plurality of updated quantized weights. The respective activation function of the at least one updated layer of the updated MLM is quantized as a function of a difference between the original MLM and the updated MLM to form a respective quantized activation function. A new quantized MLM is generated as a function of the frozen quantized weight, the updated quantized weight, and the respective quantized activation function.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates generally to neural networks, and more specifically, to partial quantization of neural networks to implement fully quantized models on edge devices. Background Art

[0002] Neural networks consist of a series of interconnected structures or layers, inspired by the neurons of the human brain. Each layer (e.g., neuron) may include the sum of multiple weighted inputs, which will produce an output if the sum exceeds the corresponding activation level. Functional models using neural networks can be built-in with sample data called training data to make predictions or decisions without explicit programming. Such models can also continue to be optimized with continuous training cycles that provide statistical learning. The trained model can then use this adaptive machine learning with subsequent data sets to predict appropriate responses.

[0003] Machine learning is increasingly used by portable devices such as cellular phones and wearable electronic devices. Due to the computing power and storage limitations of these portable devices, quantization can be used to reduce the complexity of the model. Typically, the training and quantization of machine learning models requires the portable device to communicate via the Internet cloud or with another processor with more computing resources, thus exposing the portable device to potential data security issues and generally limiting operational flexibility. Summary of the invention

[0004] As should be appreciated, at least some of the disclosed embodiments include at least the following. In one embodiment, a method for partial quantization to achieve a fully quantized model includes: quantizing a plurality of weights and corresponding activation functions from each of a plurality of corresponding layers of an original machine learning model (MLM) to generate a quantized MLM including a plurality of frozen quantized weights, each layer including a sum of a plurality of inputs, each input multiplied by a corresponding weight of the plurality of weights, and the sum gated by the corresponding activation function. Extracting the plurality of frozen quantized weights from at least one frozen layer of the layers of the quantized MLM. Quantizing the plurality of weights according to at least one updated layer of an updated MLM to generate a plurality of updated quantized weights, wherein the at least one updated layer of the updated MLM is generated by updating the corresponding layer of the original MLM. Quantizing the corresponding activation function of the at least one updated layer of the updated MLM according to a difference between the original MLM and the updated MLM to generate a corresponding quantized activation function. Generating a new quantized MLM according to the frozen quantized weights, the updated quantized weights, and the corresponding quantized activation function.

[0005] An alternative embodiment of a method for partial quantization to achieve a fully quantized model includes one of the following features or any combination thereof. The difference between the original MLM and the updated MLM is determined by determining an updated scale, wherein determining the updated scale includes adding a ratio to the original scale of the original MLM, wherein the ratio is determined by dividing the difference in delta between the maximum delta and the minimum delta by the maximum integer minus one of the integer range of the corresponding activation function used to quantize the updated MLM. The maximum delta is determined by subtracting the original maximum value of the original floating point range of the activation function of each layer of the original MLM from the updated maximum value of the updated floating point range of the activation function of each layer of the updated MLM. The minimum delta is determined by subtracting the original minimum value of the original floating point range of the activation function of each layer of the original MLM from the updated minimum value of the updated floating point range of the activation function of each layer of the updated MLM. The difference between the original MLM and the updated MLM is determined by adding an offset to the original integer zero point of the original MLM to determine an updated integer zero point, wherein the offset is determined by subtracting the updated minimum value of the updated floating point range of the activation function of each layer of the updated MLM from the original minimum value of the original floating point range of the activation function of each layer of the original MLM divided by the original scale of the original MLM. The original integer zero point is determined by subtracting the original minimum value of the original floating point range of the activation function of each layer of the original MLM from the integer minimum value of the integer range of the corresponding activation function used to quantize the quantized MLM, divided by the original scale. The original scale is determined by subtracting the original minimum value of the original floating point range from the original maximum value of the original floating point range, divided by subtracting the integer minimum value of the integer range from the integer maximum value of the integer range. The plurality of weights from at least one updated layer of the updated MLM are quantized using the original MLM, the quantized MLM, the updated MLM, and a representative data set collected on an edge device, wherein the representative data set is a subset of a data set used to train the updated MLM. The quantized MLM further includes quantizing the plurality of weights for each of a plurality of tensor inputs from the original MLM. Generating the quantized MLM further includes quantizing the plurality of weights for each of a plurality of channels from the original MLM.

[0006] In another embodiment, a method for partial quantization to achieve a fully quantized model includes extracting a plurality of frozen quantized weights from each frozen layer of a quantized machine learning model (MLM), wherein the quantized MLM is generated by quantizing an original MLM. Quantizing a plurality of updated weights from an updated MLM to generate a plurality of updated quantized weights, wherein the updated MLM is generated by updating at least one layer of the original MLM to form at least one updated layer. Quantizing a corresponding activation function from each updated layer of the updated MLM based on a difference between the original MLM and the updated MLM to generate a quantized activation function, wherein the difference includes an updated scale and an updated integer zero point. Generate a new quantized MLM based on the frozen quantized weights, the updated quantized weights, and the corresponding quantized activation function of each updated layer.

[0007] An alternative embodiment of the method for partial quantization to achieve a fully quantized model includes one of the following features or any combination thereof. The updated scale is determined by adding a ratio to the original scale of the original MLM, wherein the ratio is determined by the difference in delta between the maximum delta and the minimum delta divided by the maximum integer minus one of the integer range of the corresponding quantized activation function used to quantize the updated MLM. The maximum delta is determined by subtracting the original maximum value of the original floating point range of the activation function of each layer of the original MLM from the updated maximum value of the updated floating point range of the activation function of each layer of the updated MLM, and the minimum delta is determined by subtracting the original minimum value of the original floating point range of the activation function of each layer of the original MLM from the updated minimum value of the updated floating point range of the activation function of each layer of the updated MLM. The updated integer zero point is determined by adding an offset to the original integer zero point of the original MLM, wherein the offset is determined by subtracting the updated minimum value of the updated floating point range of the activation function of each layer of the updated MLM from the original minimum value of the original floating point range of the activation function of each layer of the original MLM divided by the original scale of the original MLM. The original integer zero point is determined by subtracting the quantized minimum value of the quantized floating point range of the activation function of each layer of the quantized MLM from the integer minimum value of the integer range of the corresponding activation function used to quantize the quantized MLM divided by the original scale, and the original scale is determined by subtracting the original minimum value of the original floating point range from the original maximum value of the original floating point range divided by subtracting the integer minimum value of the integer range from the integer maximum value of the integer range.

[0008] In another embodiment, a method for partial quantization to achieve a fully quantized model includes extracting a plurality of frozen quantized weights from each frozen layer of a quantized model, wherein the quantized model is generated by quantizing an original model. Quantizing a plurality of updated weights from an updated model to generate a plurality of updated quantized weights, wherein the updated model is generated by updating at least one layer of the original model to form at least one updated layer. Quantizing a corresponding activation function from each updated layer of the updated model to generate a quantized activation function based on a difference between the original model and the updated model, wherein the difference includes an updated scale and an updated integer zero point, wherein the updated scale modifies the original scale used to quantize the original model at a ratio proportional to the floating point range difference between the updated model and the original model, and the updated integer zero point modifies the original integer zero point used to quantize the original model by a scaled difference between the floating point range minimum of the updated model and the original model. A new quantized model is generated based on the frozen quantized weights and quantized activation functions, the updated quantized weights, and the corresponding quantized activation function of each updated layer.

[0009] An alternative embodiment of the method for partial quantization to achieve a fully quantized model includes one or any combination of the following features. The updated model is trained with a new data set stored on an edge device. A plurality of weights and corresponding activation functions of each of a plurality of corresponding layers from the original model are quantized with a representative data set, wherein the representative data set is a subset of the new data set. The updated model is adapted to the new data set via transfer learning. Inference is performed with a neural processing unit and the new quantized model to perform transfer learning on an edge device. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The present invention is illustrated by way of example and not limitation in the accompanying figures, in which like reference numerals indicate similar elements. Elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale.

[0011] Figure 1 is a schematic diagram of an embodiment of a neuron that forms the basis of a neural network.

[0012] Figure 2 is a diagrammatic view of an embodiment of an affine mapping of real numbers to integers.

[0013] Figure 3 is a flowchart representation of a method for partial quantization to achieve a fully quantized model according to an embodiment of the present disclosure.

[0014] Figure 4 is a flowchart representation of using a partial quantization tool to implement a fully quantized model according to an embodiment of the present disclosure.

[0015] Figure 5 is a flowchart representation of a training and quantization system according to an embodiment of the present disclosure.

[0016] Figure 6 is a flowchart representation of a method for partial quantization to implement a fully quantized model on an edge device according to an embodiment of the present disclosure.

[0017] Figure 7 is a flowchart representation of another method for partial quantization to implement a fully quantized model on an edge device according to an embodiment of the present disclosure.

[0018] Figure 8 is a flowchart representation of another method for partial quantization to implement a fully quantized model on an edge device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0019] Embodiments described herein provide for training and quantizing machine learning models (MLMs) on edge devices without requiring the resources of a more powerful host machine, nor requiring a connection to a host or similar cloud-based computing platform. Edge devices as defined herein include hardware that controls data flow at the boundaries of a network, such as cellular phones, wearable fitness tracking devices, portable devices, and the like. Edge devices typically have limited processing power and memory storage compared to host machines.

[0020] Partially training and quantizing the MLM directly on the edge device avoids the typical requirement to provide a full training data set on the device, where the full data set may include proprietary information. In addition, by quantizing the MLM directly on the device, security issues associated with the host (including, for example, malware, data theft, and cloning) are avoided. Additional personalization of the edge device can also be achieved without typical data security concerns.

[0021] Partial training and quantization on the device may also facilitate "transfer learning," whereby a device trained by a set of data may be extended or modified to learn with related data by modifying only a subset of many layers of the MLM. In one non-limiting example of transfer learning, a device is trained to recognize sedans. Via transfer learning, the device is subsequently taught to recognize trucks or buses. A new MLM with updated layers may thus be quantized by maintaining quantization of weights associated with frozen layers (e.g., those layers that have not been updated due to transfer learning), quantizing weights associated with updated layers, and quantizing activation functions based on changes to scaling and integer zeros between the original MLM and the updated MLM (with updated layers).

[0022] Figure 1An example of a neuron that forms the basis of an MLM is shown. In one example, a neuron may include a plurality of inputs (e.g., tensor inputs) 10a, 10b, and 10c (collectively 10). Each input 10 is modified by a corresponding weight 12a, 12b, and 12c (collectively 12) and summed using a biased summation 14. If the result of the biased summation exceeds a threshold defined by an activation function 16, an output 18 is generated. Figure 1 An example neuron may define a single layer from multiple interconnected layers of an MLM.

[0023] Figure 2 An example of an affine mapping for quantization of floating point numbers to integers is shown. Specifically, a floating point range 20 includes a floating point minimum value (Rmin) 22 , a floating point maximum value (Rmax) 24 , and a floating point zero 26 . Figure 2 The affine mapping of further includes an integer range 30. The integer range 30 includes an integer minimum value (Qmin) 32, an integer maximum value (Qmax) 34, and an integer zero point (Z) 36. Figure 2 The example of shows asymmetric quantization of floating point numbers to integers based on the scaling difference between Rmax 24 - Rmin 22 and Qmax 34 - Qmin 32, and potential shifting of floating point zeros 26 to integer zeros 36 within the integer range 30. Figure 2 The affine mapping reduces the numbers represented by floating point values ​​to integer values, which will require less memory storage and require less computing resources during subsequent processing of the integer values.

[0024] Figure 3 It is a flowchart representation of a method for partial quantization to achieve a fully quantized model according to an embodiment 40 of the present disclosure. In one embodiment, the representative data set 42 is a subset of the training data set used to train the original 32-bit floating point (FP32) model 44 to create the original quantized INT8 model 46. Therefore, the model 46 is a pre-quantized model created from the original FP32 model 44. The original quantized INT8 model 46 has been quantized using more than one image before being deployed to the edge device. The representative data set 42 may be a subset of the data generated or produced on the edge device itself and used to train the original FP32 model 44 via transfer learning. Specifically, only one image from the representative data set 42 is required to calculate R1max, R1min, R2max and R2min in the updated layer 52 through the original FP32 model 44 and the updated FP32 model 48, and the one image ultimately reconstructs the new quantized model 64.

[0025] It should be understood that Figure 3The examples of use FP32 and INT8 models for ease of illustration, but should not be considered limited to these numerical ranges. The representative data set 42 can also be used to train (or update) FP32 models 48 on a quantization device for transfer learning. The updated model 48 includes a frozen layer 50, which remains the same as the corresponding layer of the original FP32 model 44. The updated model 48 further includes an updated layer 52 that is different from the corresponding layer of the original FP32 model 44.

[0026] In embodiment 40, quantized weights are extracted 54 from frozen layer 50 of model 46 to form frozen quantized weights. Weights of updated layer 52 of updated model 48 are quantized 56 to generate updated quantized weights. Figure 2 and Figure 3 At 60, a floating point maximum value (Rmax) and a floating point minimum value (Rmin) (R1max and R1min, respectively) for the range of activated floating point values ​​for the original FP32 model 44 are determined, and a floating point maximum value (Rmax) and a floating point minimum value (Rmin) (R2max and R2min, respectively) for the range of activated floating point values ​​for the updated FP32 model 48 are determined. The maximum and minimum delta values ​​are determined at 62 according to the following equations [1] and [2]:

[0027] [1] Maximum difference = R2max - R1max

[0028] [2] Minimum difference = R2min - R1min

[0029] At 64, a new quantized model is generated, wherein the frozen quantized weights and activated quantized parameters are extracted from 54, the updated quantized weights are quantized from 56, and the updated quantized activation function of the updated model 48 is quantized with the maximum delta and minimum delta values ​​from 62. In one embodiment, the activated quantized parameters include scaling and integer zero values.

[0030] Figure 4 is a flowchart representation of using a partial quantization tool to implement a fully quantized model according to embodiment 70 of the present disclosure. Figure 4 , Example 70 is further defined for reconstructing Figure 3 The method of the new quantized model shown in Example 40 of . Similar to Figure 3 , the partial quantization tool 72 receives input from the original FP32 model 44, the quantized INT8 model 46, and the updated FP32 model 48, and generates a new quantized model 74 from the model reconstruction step 64. Figure 2 and Figure 4 , the partial quantization tool 72 determines the original scale (S1) and the original integer zero point (Z1) at 76 according to the following equations [3] and [4]:

[0031] [3]S=(Rmax-Rmin) / (Qmax-Qmin)

[0032] [4]Z=Qmin-Rmin / S

[0033] The partial quantization tool 72 determines at 78 an updated scale (S2) and an updated integer zero point (Z2) according to the following equations [5] and [6]:

[0034] [5]S2=S1 and (Delta-max-Delta-min) / 255

[0035] [6]Z2=Z1+(R1min / S1)-(R2min / S2)

[0036] Equation [5] is an example based on quantizing a floating point number to an eight-bit integer value (INT8), so the denominator of 255 represents the maximum integer value of the integer range (256) minus 1. For embodiments based on different integer ranges, the denominator should be appropriately modified to match the new integer range minus one.

[0037] Figure 5 8 is a flowchart representation of a training and quantization system according to an embodiment 80 of the present disclosure. A new data set is provided to train an updated model 48 having an updated layer. In one example, the new data set 82 is used to train the updated model 48 using a tensorflow (TF) lite program 84. Figure 3 and Figure 5 , the representative data set 42 (derived as a subset of the new data set 82) is used together with the original quantized model 46, the updated model 48 and the original floating point model 44 to use Figure 4 The partial quantization tool generates a new quantized model 74. In one embodiment, the representative data set 42, the original quantized model 46, the updated model 48, and the original floating point model 44 are all stored on the edge device without connecting to other computing resources (e.g., a host or cloud connection). The new quantized model 74 can then be used by the Neural Processing Unit (NPU) for inference to perform transfer learning.

[0038] Figure 6 An embodiment 90 of a method for partial quantization to achieve a fully quantized model is shown. Figure 3 and Figure 4, at 92, a plurality of weights and corresponding activation functions of the original MLM 44 are quantized to generate a quantized MLM 46. At 94, frozen quantized weights and quantized activation parameters are extracted 54 from the quantized MLM. At 96, weights of the updated layer 52 from the updated MLM 48 are quantized 56. At 98, corresponding activation functions of the updated MLM 48 are quantized based on the difference between the original MLM 44 and the updated MLM 48. At 100, a new quantized MLM 64 is generated based on the frozen weights, the quantized weights from the updated MLM 48, and the quantized activation functions of the updated MLM 48.

[0039] Figure 7 An embodiment 110 of a method for partial quantization to achieve a fully quantized model is shown. Figure 3 and Figure 4 At 112, frozen quantized weights and quantized activation parameters are extracted 54 from at least one frozen layer 50 of the quantized MLM 46 generated by the original MLM 44. At 114, weights from the updated layer 52 of the updated MLM 48 are quantized 56. At 116, the weights are quantized 56 according to the updated scale S2 and the updated integer zero point Z2 (see Figure 4 At 118, a new quantized MLM 64 is generated based on the frozen weights, the quantized weights from the updated MLM 48, and the quantized activation function of the updated MLM 48.

[0040] Figure 8 An embodiment 120 of a method for partial quantization to achieve a fully quantized model is shown. Figure 3 , Figure 4 and Figure 5 At 122, frozen quantized weights are extracted 54 from at least one frozen layer 50 of the quantized model 46 generated by the original model 44. At 124, weights from the updated layer 52 of the updated model 48 are quantized 56. At 126, the weights are quantized according to the updated scale S2 and the updated integer zero point Z2 (see Figure 4 78) quantizes the corresponding activation function of the updated layer 52 of the updated model 48, wherein the updated scale S2 modifies the original scale S1 used to quantize the original model 44, and the updated zero integer zero point Z2 modifies the original integer zero point Z1 used to quantize the original model 44. At 128, a new quantized model 64 is generated based on the frozen weights, the quantized weights from the updated model 48, and the quantized activation function of the updated model 48.

[0041] Although the present invention is described herein with reference to specific embodiments, various modifications and changes may be made without departing from the scope of the present invention as set forth in the appended claims. Therefore, the description and drawings should be regarded as illustrative rather than restrictive, and all such modifications are intended to be included within the scope of the present invention. It is not intended that any advantages, advantages, or solutions to problems described herein with respect to specific embodiments be construed as key, essential, or requisite features or elements of any or all claims.

[0042] Unless stated otherwise, terms such as "first" and "second" are used to arbitrarily distinguish between the elements these terms describe. Therefore, these terms are not necessarily intended to indicate temporal or other prioritization of such elements.

Claims

1. A method for partial quantization to achieve a fully quantized model, characterized in that include: quantizing a plurality of weights and a corresponding activation function from each of a plurality of corresponding layers of an original machine learning model (MLM) to generate a quantized MLM comprising a plurality of frozen quantized weights, each layer comprising a summation of a plurality of inputs, each input being multiplied by a corresponding weight from the plurality of weights, and the summation being gated by the corresponding activation function; extracting the plurality of frozen quantized weights from at least one frozen layer of the layers of the quantized MLM; quantizing the plurality of weights from at least one updated layer of an updated MLM to generate a plurality of updated quantized weights, wherein the at least one updated layer of the updated MLM is generated by updating the corresponding layer of the original MLM; quantizing the corresponding activation function of the at least one updated layer of the updated MLM according to a difference between the original MLM and the updated MLM to generate a corresponding quantized activation function; as well as A new quantized MLM is generated according to the frozen quantized weights, the updated quantized weights and the corresponding quantized activation function.

2. The method according to claim 1, characterized in that Further comprising determining the difference between the original MLM and the updated MLM by determining an updated scale, wherein determining the updated scale comprises adding a ratio to the original scale of the original MLM, wherein the ratio is determined by dividing the difference between the maximum difference and the minimum difference by the maximum integer minus one of the integer range of the corresponding activation function used to quantize the updated MLM.

3. The method according to claim 1, characterized in that Further comprising determining the difference between the original MLM and the updated MLM by adding an offset to the original integer zero point of the original MLM to determine an updated integer zero point, wherein the offset is determined by subtracting an updated minimum value of the updated floating point range of the activation function of each layer of the updated MLM divided by the updated scale from the original minimum value of the original floating point range of the activation function of each layer of the original MLM divided by the original scale of the original MLM.

4. The method according to claim 1, characterized in that: Further comprising quantizing the plurality of weights from the at least one updated layer of the updated MLM using the original MLM, the quantized MLM, the updated MLM, and a representative dataset collected on an edge device, wherein the representative dataset is a subset of a dataset used to train the updated MLM.

5. A method for partial quantization to achieve a fully quantized model, characterized in that include: Extracting a plurality of frozen quantized weights from each frozen layer of a quantized machine learning model (MLM), wherein the quantized MLM is generated by quantizing an original MLM; quantizing a plurality of updated weights from an updated MLM to generate a plurality of updated quantized weights, wherein the updated MLM is generated by updating at least one layer of the original MLM to form at least one updated layer; quantizing a corresponding activation function from each updated layer of the updated MLM according to a difference between the original MLM and the updated MLM to generate a quantized activation function, wherein the difference includes an updated scale and an updated integer zero point; as well as A new quantized MLM is generated according to the frozen quantized weights, the updated quantized weights, and the corresponding quantized activation function of each updated layer.

6. The method according to claim 5, characterized in that The updated scale is determined by adding a ratio to an original scale of the original MLM, wherein the ratio is determined by a delta difference between a maximum delta and a minimum delta divided by a maximum integer minus one of an integer range of the corresponding quantized activation function used to quantize the updated MLM.

7. The method according to claim 5, characterized in that The updated integer zero point is determined by adding an offset to the original integer zero point of the original MLM, wherein the offset is determined by subtracting the updated minimum value of the updated floating-point range of the activation function of each layer of the updated MLM divided by the updated scale from the original minimum value of the original floating-point range of the activation function of each layer of the original MLM divided by the original scale of the original MLM.

8. A method for partial quantization to achieve a fully quantized model, characterized in that include: extracting a plurality of frozen quantized weights from each frozen layer of a quantized model, wherein the quantized model is generated by quantizing an original model; quantizing a plurality of updated weights from an updated model to generate a plurality of updated quantized weights, wherein the updated model is generated by updating at least one layer of the original model to form at least one updated layer; quantizing a corresponding activation function from each updated layer of the updated model according to a difference between the original model and the updated model to generate a quantized activation function, wherein the difference comprises an updated scale and updated integer zeros, the updated scale modifies an original scale used to quantize the original model by a ratio proportional to a difference in floating point ranges between the updated model and the original model, and the updated integer zeros modify the original integer zeros used to quantize the original model by a scaled difference between the floating point range minimums of the updated model and the original model; as well as A new quantized model is generated according to the frozen quantized weights, the updated quantized weights, and the corresponding quantized activation function of each updated layer.

9. The method according to claim 8, characterized in that Further comprising training the updated model with a new dataset stored on the edge device.

10. The method according to claim 9, characterized in that Further comprising quantizing a plurality of weights and corresponding activation functions of each of a plurality of corresponding layers from the original model using a representative dataset, wherein the representative dataset is a subset of the new dataset.

Citation Information

Cited By

  • Method and system for training deep learning model

    WO2026222943A1