KR-102961232-B1 - Method for on-device training of a machine learning network of an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set, and an on-device training device using the same

KR102961232B1KR 102961232 B1KR102961232 B1KR 102961232B1KR-102961232-B1

Abstract

The present invention relates to a method for on-device training of a machine learning network through multi-stage training using adaptive hyperparameters, comprising: (a) a step of classifying the current training into a first stage training to an n-th stage training when new training data satisfies on-device training conditions, generating training data for the first stage to the n-th stage training, generating a candidate for a first hyperparameter set or a candidate for a hyperparameter set based on the default values of each of the hyperparameters, training the machine learning network, selecting the machine learning network with the highest performance, and generating a first adaptive hyperparameter set; (b) a step of generating a candidate for a (k_1) hyperparameter set to a candidate for a (k_h) hyperparameter set, training using the k-th stage training data, selecting the machine learning network trained at the (k-1)-th stage with the highest performance, and generating a k-th adaptive hyperparameter set; A method and an apparatus using the same are disclosed, comprising: (c) a step of completing the current learning by generating an n-th adaptive hyperparameter set and training the n-th stage.

Inventors

제홍모
김용중
유동규
권성안

Assignees

주식회사 스트라드비젼

Dates

Publication Date: 20260507
Application Date: 20210504
Priority Date: 20210413

Claims (16)

A method for on-device training of a machine learning network for an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set, (a) When the on-device learning condition is satisfied while acquiring new training data during the operation of an autonomous vehicle, the on-device learning device (i) refers to a basic hyperparameter set to classify the current learning into a first stage learning to the nth stage learning - where n is an integer greater than or equal to 2 - stage learning, and generates first stage learning data to nth stage learning data for the first stage learning to the nth stage learning using the new training data and the previous training data used in the previous learning, and (ii) generates a first hyperparameter set candidate to a (1_h) hyperparameter set candidate - where h is an integer greater than or equal to 2 - hyperparameter set candidate by combining each of the first candidate values within a preset range based on the default values of each hyperparameter included in the basic hyperparameter set, and (iii) applies the basic hyperparameter set and each of the first hyperparameter set candidate to the (1_h) hyperparameter set candidate to train the machine learning network using the first stage learning data, respectively, and (iv) the basic hyperparameter set and Evaluating the performance of each of the machine learning networks trained by applying each of the first hyperparameter set candidate to the (1_h) hyperparameter set candidate, and selecting the machine learning network with the highest performance as the first stage trained machine learning network; (v) generating a first adaptive hyperparameter set of hyperparameters applied to the training of the first stage trained machine learning network; (b) the on-device learning device increases k from 2 to (n-1), (i) generates a (k_1) hyperparameter set candidate to a (k_h) hyperparameter set candidate by combining each of the k candidate values within the preset range based on the (k-1) adaptive values of each of the hyperparameters included in the (k-1) adaptive hyperparameter set, (ii) trains a machine learning network trained at the (k-1) stage by applying the (k-1) adaptive hyperparameter set and each of the (k_1) hyperparameter set candidate to the (k_h) hyperparameter set candidate using the k-stage training data, and (iii) evaluates the performance of each of the machine learning networks trained at the (k-1) stage by applying the (k-1) adaptive hyperparameter set and each of the (k_1) hyperparameter set candidate to the (k_h) hyperparameter set candidate. Selecting the machine learning network trained at the (k-1) stage with the highest performance as the machine learning network trained at the k-th stage, and (iv) generating the set of hyperparameters applied to the training of the machine learning network trained at the k-th stage as the k-th adaptive hyperparameter set; and (c) the on-device learning device, (i-1) generates an n-th adaptive hyperparameter set using an optimization function constructed by referring to each of the first adaptive hyperparameter set to the (n-1)-th adaptive hyperparameter set and (i-2) the performance evaluation results of each of the machine learning network learned in the first stage learning to the machine learning network learned in the (n-1)-th stage learning, and (ii) completes the current learning by applying the n-th adaptive hyperparameter set to train the machine learning network learned in the (n-1)-th stage using the n-th stage learning data. A method characterized by including
In paragraph 1, (d) The on-device learning device comprises: (i) a process of not updating the machine learning network if the performance of the machine learning network in the current state of completed learning is not improved by a certain threshold, and allowing the autonomous vehicle to operate using the machine learning network until the next on-device learning condition is satisfied; and (ii) a process of updating the machine learning network to the currently learned machine learning network if the performance of the machine learning network in the current state of completed learning is improved by a certain threshold, and allowing the autonomous vehicle to operate using the currently learned machine learning network until the next on-device learning condition is satisfied; A method characterized by further including
In paragraph 1, A method in which, when the performance of the machine learning network trained at the k-th stage is higher than the performance of the machine learning network trained at the (k-1)-th stage, each of the (k+1) adaptive hyperparameter set to the nth adaptive hyperparameter set is set to be the same as the k-th adaptive hyperparameter set.
In paragraph 1, In step (b) above, The above-described on-device learning device is a method for maintaining the same adaptive value of at least one hyperparameter included in the first adaptive hyperparameter set to the nth adaptive hyperparameter set.
In paragraph 1, A method characterized by the above-described on-device learning device completing the current learning by setting the k-th stage learning to the n-th stage learning when the performance of the machine learning network learned in the k-th stage learning is higher than the performance of the machine learning network learned in the (k-1)-th stage learning.
In paragraph 1, In step (a) above, When sensing data is acquired by sensors mounted on the autonomous vehicle, the machine learning network analyzes the sensing data and generates output data regarding the driving information of the autonomous vehicle, A method characterized in that the above-described on-device learning device inputs the sensing data and output data corresponding to each of the sensing data into a data selection network, thereby causing the data selection network to select specific sensing data to be used for training the machine learning network by referring to the output data, and to store the selected specific sensing data as the new training data.
In paragraph 1, A method characterized in that the above hyperparameters include at least one of a learning algorithm setting, a mini-batch size, a maximum stage, and a maximum epoch for each stage.
In paragraph 1, In step (a) above, A method characterized by the above-determined on-device learning device selecting a pre-determined set of hyperparameters as the base model of the machine learning network as the basic hyperparameter set, or selecting a best set of hyperparameters generated in previous learning as the basic hyperparameter set.
In a learning device for on-device training of a machine learning network of an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set, At least one memory for storing instructions; and It includes one or more processors configured to execute the above instructions, The above processor, (I) when the on-device learning condition is satisfied while acquiring new training data during the operation of an autonomous vehicle, refers to a basic hyperparameter set to classify the current learning into a first stage learning to the nth stage learning - where n is an integer greater than or equal to 2 - stage learning, and generates first stage learning data to nth stage learning data for the first stage learning to the nth stage learning using the new training data and the previous training data used in the previous learning, and generates a first hyperparameter set candidate to a (1_h) hyperparameter set candidate - where h is an integer greater than or equal to 2 - by combining each first candidate value within a preset range based on the default value of each hyperparameter included in the basic hyperparameter set, and applies the basic hyperparameter set and each of the first hyperparameter set candidate to the (1_h) hyperparameter set candidate to train the machine learning network using the first stage learning data, respectively, and (iv) the basic hyperparameter set and the first hyperparameter set candidate to the (1_h) Evaluating the performance of each machine learning network trained by applying each of the hyperparameter set candidates and selecting the machine learning network with the highest performance as the machine learning network trained in the first stage; (v) A process of generating a hyperparameter set applied to the training of the machine learning network trained in the first stage as a first adaptive hyperparameter set; (II) Increasing k from 2 to (n-1), generating a hyperparameter set candidate from the (k_1) hyperparameter set candidate to the (k_h) hyperparameter set candidate by combining each of the k candidate values within the preset range based on the (k-1) adaptive values of each of the hyperparameters included in the (k-1) adaptive hyperparameter set; and training the machine learning network trained in the (k-1) stage by applying each of the (k_1) adaptive hyperparameter set candidate to the (k_h) hyperparameter set candidate using the k-stage training data, and A process of evaluating the performance of each machine learning network trained at the (k-1) stage by applying the (k-1) adaptive hyperparameter set and the candidate for the (k_1) hyperparameter set to the candidate for the (k_h) hyperparameter set, selecting the machine learning network trained at the (k-1) stage with the highest performance as the machine learning network trained at the (k-1) stage, and generating the hyperparameter set applied to the training of the machine learning network trained at the (k-1) stage as the k-th adaptive hyperparameter set, and (III) generating the n-th adaptive hyperparameter set using an optimization function constructed by referring to the performance evaluation results of each of the first adaptive hyperparameter set to the (n-1) adaptive hyperparameter set and the machine learning network trained at the first stage to the machine learning network trained at the (n-1) stage, and completing the current training by training the machine learning network trained at the (n-1) stage using the n-th stage training data by applying the n-th adaptive hyperparameter set. A learning device that performs a process.
In Paragraph 9, (IV) A learning device that further performs the following processes: if the performance of the machine learning network in the current state of completed learning is not improved by a certain threshold, the machine learning network is not updated, and the autonomous vehicle is operated using the machine learning network until the next on-device learning condition is satisfied; and if the performance of the machine learning network in the current state of completed learning is improved by a certain threshold, the machine learning network is updated to the currently learned machine learning network, and the autonomous vehicle is operated using the currently learned machine learning network until the next on-device learning condition is satisfied.
In Paragraph 9, A learning device characterized by the processor setting each of the (k+1) adaptive hyperparameter set to the (k+1) adaptive hyperparameter set to be the same as the (k) adaptive hyperparameter set when the performance of the machine learning network trained at the k-th stage is higher than the performance of the machine learning network trained at the (k-1)th stage.
In Paragraph 9, In the above (II) process, The above processor is a learning device that maintains the same adaptive value of at least one hyperparameter included in the first adaptive hyperparameter set to the nth adaptive hyperparameter set.
In Paragraph 9, A learning device characterized by the processor completing the current learning by setting the k-th stage learning to the n-th stage learning when the performance of the machine learning network learned at the k-th stage is higher than the performance of the machine learning network learned at the (k-1)-th stage.
In Paragraph 9, In the above (I) process, When sensing data is acquired by sensors mounted on the autonomous vehicle, the machine learning network analyzes the sensing data based on deep learning to generate output data regarding the driving information of the autonomous vehicle, A learning device characterized by the processor inputting the sensing data and output data corresponding to each of the sensing data into a data selection network, causing the data selection network to select specific sensing data to be used for learning the machine learning network by referring to the output data, and storing the selected specific sensing data as the new learning data.
In Paragraph 9, A learning device characterized in that the above hyperparameters include at least one of a learning algorithm setting, a mini-batch size, a maximum stage, and a maximum epoch for each stage.
In Paragraph 9, In the above (I) step, A learning device characterized by the processor selecting a pre-determined set of hyperparameters as the base model of the machine learning network as the basic hyperparameter set, or selecting a best set of hyperparameters generated in previous learning as the basic hyperparameter set.

Description

Method for on-device training of a machine learning network of an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set, and an on-device training device using the same The present invention claims priority to U.S. Patent Application No. 63/020,101 filed May 5, 2020 and U.S. Patent Application No. 17/229,350 filed April 13, 2021, which are incorporated herein by reference. The present invention relates to a method for on-device learning of a machine learning network of an autonomous vehicle and a device using the same. More specifically, the invention relates to a method for on-device learning of a machine learning network installed in an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set on an embedded system, and an on-device learning device using the same. In order to enable the machine learning network applied to an autonomous vehicle to adapt to a new driving environment that has not been previously trained on, it is necessary to retrain the said machine learning network using information acquired through sensors, such as cameras, LiDAR, and radar, while the autonomous vehicle is driving. Since it is difficult to perform annotation for generating ground truth (GT) for learning using data acquired from autonomous vehicles on the autonomous vehicle itself, a method is proposed to train the machine learning network of an autonomous vehicle by creating a base model of the machine learning network or by constructing a training dataset that combines a portion of the training data used in previous training with the data to be used for training from the data acquired in real-time from the autonomous vehicle. In particular, compared to conventional methods, on-device learning technology is emerging that utilizes OTA (over the air) technology to transmit data required for training machine learning networks in autonomous vehicles to the cloud, trains the machine learning network on the server side using the data transmitted via the cloud, and then transmits only the model updated through training back to the autonomous vehicle. However, conventional on-device learning methods utilizing such OTA technology have limitations, such as slow update cycles and the inability to use them in situations where an OTA connection with a cloud server is impossible. Therefore, there is a need for technology that can train machine learning networks using an embedded system within autonomous vehicles with limited computing power, without the need for a process of connecting to a cloud server and OTA. The drawings attached below for use in describing embodiments of the present invention are merely some of the embodiments of the present invention, and other drawings can be obtained based on these drawings without inventive work by a person skilled in the art to which the present invention pertains (hereinafter “person skilled in the art”). FIG. 1 schematically illustrates an on-device learning device for on-device learning of a machine learning network of an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set according to an embodiment of the present invention. FIG. 2 schematically illustrates a method for on-device training a machine learning network of an autonomous vehicle through multi-stage learning using an adaptive hyperparameter set according to an embodiment of the present invention. FIG. 3 schematically illustrates the configuration of a hyperparameter set according to one embodiment of the present invention, and FIG. 4 schematically illustrates a method for generating an adaptive hyperparameter set through each stage learning according to an embodiment of the present invention and multi-stage learning a machine learning network using the generated adaptive hyperparameter set. The following detailed description of the present invention refers to the accompanying drawings, which illustrate specific embodiments in which the present invention can be practiced in order to clarify the objects, technical solutions, and advantages of the present invention. These embodiments are described in sufficient detail to enable a person skilled in the art to practice the present invention. Furthermore, throughout the detailed description and claims of the invention, the word “comprising” and its variations are not intended to exclude other technical features, additions, components, or steps. Other objects, advantages, and characteristics of the invention will become apparent to a person skilled in the art, in part from this description and in part from the practice of the invention. The following examples and drawings are provided by way of example and are not intended to limit the invention. Furthermore, the present invention encompasses all possible combinations of the embodiments set forth in this specification. It should be understood that various embodiments of the present invention are different but need not be mutually exclusive. For ex