Guided training data acquisition for machine learning
By dividing the input data set into multiple categories and generating regression curve equations, the data acquisition quota is optimized, and the problem of low data acquisition efficiency in vehicle machine learning algorithm training is solved, and a more efficient training process is achieved.
Patent Information
- Application Number
- CN202311401644.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-05-08
- Filing Date
- 2023-10-25
- Publication Date
- 2024-11-08
AI Technical Summary
Existing vehicle machine learning algorithm training requires a lot of time and resources, and it is difficult to optimize data acquisition to improve training efficiency.
By dividing the input dataset into multiple categories and generating regression curve equations based on exploratory training courses, determining data acquisition quotas, optimizing data acquisition to reduce model losses, and using power law equations and predetermined equations to adjust the data acquisition task.
It effectively reduces training time and resource use, improves the training efficiency of machine learning algorithms, and optimizes the data acquisition process.
Smart Images

Figure CN118916691A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to systems and methods for training machine learning algorithms. Background Art
[0002] To improve occupant awareness and convenience, vehicles may be equipped with driver assistance systems and / or automated driving systems. Driver assistance systems may use inputs from multiple vehicle sensors to determine information about the vehicle's surroundings and determine recommended or optimal behaviors for the vehicle and / or occupants. In order to utilize large amounts of input sensor data to determine information about complex environments under dynamic conditions, driver assistance systems may utilize machine learning algorithms that take inputs from vehicle sensors and determine recommended or optimal behaviors. However, machine learning algorithms must be trained using large amounts of input data. Collecting the input data may require additional time and / or resources. In addition, in order to optimally train a machine learning algorithm, specific types or categories of training data with specific characteristics may be required.
[0003] Therefore, while current vehicle machine learning algorithms achieve their intended purposes, there is still a need for new and improved systems and methods for training vehicle machine learning algorithms. Summary of the invention
[0004] In several aspects, the present disclosure provides a method for training a machine learning algorithm. The method includes performing at least one exploratory training session of the machine learning algorithm using an input data set. The input data set includes a plurality of first input data samples. The method also includes dividing the input data set into a plurality of input data categories. The method also includes determining a regression curve equation for each of the plurality of input data categories based at least in part on the at least one exploratory training session. The method also includes collecting a plurality of second input data samples based at least in part on the regression curve equation for each of the plurality of input data categories. The method also includes training the machine learning algorithm using the plurality of second input data samples and the input data set.
[0005] In another aspect of the present disclosure, dividing the input data set into a plurality of input data categories may further include identifying at least one data set parameter by which the input data set is classified.Dividing the input data set into a plurality of input data categories may further include dividing the input data set into a plurality of input data categories based on at least one data set parameter.
[0006] In another aspect of the present disclosure, determining a regression curve equation for each of a plurality of input data categories may further include generating an exploratory learning curve plot of a model loss of a machine learning algorithm relative to the number of input data samples trained for each of the plurality of input data categories in at least one exploratory training session. Determining a regression curve equation for each of a plurality of input data categories may further include determining a regression curve equation for each of a plurality of input data categories based at least in part on the exploratory learning curve plot for each of the plurality of input data categories.
[0007] In another aspect of the present disclosure, determining the regression curve equation for each of the plurality of input data categories may also include determining the regression curve equation for each of the plurality of input data categories based at least in part on an exploratory learning curve graph for each of the plurality of input data categories. The regression curve equation for one of the plurality of input data categories is a power law equation having the following form:
[0008]
[0009] Where ε is the model loss of the machine learning algorithm for the one of the multiple input data categories; α is the first constant factor for the one of the multiple input data categories; m is the number of input data samples trained for the one of the multiple input data categories; β g is the steepness of the regression curve described by the regression curve equation of the one of the multiple input data categories; and γ is the lower bound model loss of the machine learning algorithm for the one of the multiple input data categories.
[0010] In another aspect of the present disclosure, collecting the plurality of second input data samples may further include identifying a first subset of the plurality of input data categories based at least in part on the number of input data samples in each of the plurality of input data categories. Collecting the plurality of second input data samples may further include determining a plurality of data collection quotas based at least in part on a regression curve equation for each of the plurality of input data categories. Each of the plurality of data collection quotas corresponds to an input data category in the first subset of the plurality of input data categories. Collecting the plurality of second input data samples may further include collecting the plurality of second input data samples based at least in part on the plurality of data collection quotas.
[0011] In another aspect of the present disclosure, identifying a first subset of the plurality of input data categories may further include comparing the number of input data samples for each of the plurality of input data categories to a previous data collection quota. The previous data collection quota is one of the plurality of data collection quotas determined during a previous execution of the method. Identifying the first subset of the plurality of input data categories may further include determining each of the plurality of input data categories having a number of input data samples less than the previous data collection quota as an input data category in the first subset of the plurality of input data categories.
[0012] In another aspect of the present disclosure, determining a data collection quota from among the plurality of data collection quotas that corresponds to an input data category from a first subset of the plurality of input data categories may further include identifying a second subset of the plurality of input data categories. The second subset of the plurality of input data categories includes each of the plurality of input data categories that is not in the first subset of the plurality of input data categories. Determining a data collection quota from among the plurality of data collection quotas that corresponds to an input data category from the first subset of the plurality of input data categories may further include determining an average steepness of a regression curve equation for each input data category from the second subset of the plurality of input data categories based at least in part on a regression curve equation for each input data category from the second subset of the plurality of input data categories. Determining a data collection quota from among the plurality of data collection quotas corresponding to an input data category in the first subset of the plurality of input data categories may also include: determining a constant factor α′ and a lower limit model loss γ′ of the input data category in the first subset of the plurality of input data categories based at least in part on a regression curve equation of the input data category in the first subset of the plurality of input data categories. Determining a data collection quota from among the plurality of data collection quotas corresponding to an input data category in the first subset of the plurality of input data categories may also include: determining a constant factor α′ and a lower limit model loss γ′ of the input data category in the first subset of the plurality of input data categories based at least in part on an average steepness The constant factor α′ and the lower bound model loss γ′ determine the one of the plurality of data collection quotas using a predetermined equation.
[0013] In another aspect of the present disclosure, determining one of the plurality of data collection quotas using a predetermined equation may also include determining the one of the plurality of data collection quotas using a predetermined equation. The predetermined equation includes:
[0014]
[0015] Where m i+1 is the one data collection quota among multiple data collection quotas; m iis the number of input data samples of the one input data category in the first subset of the plurality of input data categories; α′ is a first constant factor of the one input data category in the first subset of the plurality of input data categories; γ′ is a lower bound model loss for the one input data category in the first subset of the plurality of input data categories; and is the average steepness of the regression curve equation for each input data category in the second subset of the plurality of input data categories.
[0016] In another aspect of the present disclosure, collecting a plurality of second input data samples may further include determining a number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories using the following equation:
[0017] s=m i+1 -m i
[0018] Where s is the number of additional input data samples to be collected for the one input data category in the first subset of the plurality of input data categories; m i+1 is a data collection quota for the one input data category in the first subset of the plurality of input data categories in the plurality of data collection quotas; and m i is the number of input data samples of the one input data category in the first subset of the plurality of input data categories. Collecting the plurality of second input data samples may also include sending a data sample collection task to the vehicle. The data sample collection task includes at least the number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories.
[0019] In another aspect of the present disclosure, training a machine learning algorithm using a plurality of second input data samples and an input data set may also include receiving a plurality of second input data samples from a vehicle. Training a machine learning algorithm using a plurality of second input data samples and an input data set may also include generating an updated input data set. The updated input data set includes a plurality of second input data samples and an input data set. Training a machine learning algorithm using a plurality of second input data samples and an input data set may also include performing the method using an updated input data set.
[0020] In several aspects, the present disclosure provides a system for training a machine learning algorithm for a vehicle. The system includes a server system including a server storage device, a server communication system, and a server controller in electrical communication with the server storage device and the server communication system. The server controller is programmed to perform at least one exploratory training course of the machine learning algorithm using an input data set. The input data set includes a plurality of first input data samples. The input data set is stored on the server storage device. The server controller is also programmed to divide the input data set into a plurality of input data categories. The server controller is also programmed to generate an exploratory learning curve graph of a model loss of the machine learning algorithm relative to the number of input data samples trained for each of the plurality of input data categories in at least one exploratory training course. The server controller is also programmed to determine a regression curve equation for each of the plurality of input data categories based at least in part on the exploratory learning curve graph for each of the plurality of input data categories. The regression curve equation for one of the plurality of input data categories is a power law equation having the following form:
[0021]
[0022] Where ε is the model loss of the machine learning algorithm for the one of the multiple input data categories; α is the first constant factor for the one of the multiple input data categories; m is the number of input data samples trained for the one of the multiple input data categories; β g is the steepness of the regression curve described by the regression curve equation for the one of the multiple input data categories; and γ is the lower bound model loss of the machine learning algorithm for the one of the multiple input data categories. The server controller is also programmed to collect a plurality of second input data samples from the vehicle using the server communication system. The plurality of second input data samples are based at least in part on the regression curve equation for each of the plurality of input data categories. The server controller is also programmed to train the machine learning algorithm using the plurality of second input data samples and the input data set.
[0023] In another aspect of the present disclosure, to collect a plurality of second input data samples from the vehicle using the server communication system, the server controller is further programmed to identify a first subset of the plurality of input data categories based at least in part on the number of input data samples in each of the plurality of input data categories. To collect a plurality of second input data samples from the vehicle using the server communication system, the server controller is further programmed to determine a plurality of data collection quotas based at least in part on a regression curve equation for each of the plurality of input data categories, wherein each of the plurality of data collection quotas corresponds to an input data category in the first subset of the plurality of input data categories. To collect a plurality of second input data samples from the vehicle using the server communication system, the server controller is further programmed to determine a number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories using the following equation:
[0024] s=m i+1 -m i
[0025] Where s is the number of additional input data samples to be collected for one input data category in the first subset of the plurality of input data categories; m i+1 is the one data collection quota in the plurality of data collection quotas for the one input data category in the first subset of the plurality of input data categories; and m i is the number of input data samples in the one input data category in the first subset of the plurality of input data categories. To collect a plurality of second input data samples from the vehicle using the server communication system, the server controller is further programmed to send a data sample collection task to the vehicle using the server communication system. The data sample collection task includes at least the number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories.
[0026] In another aspect of the present disclosure, in order to determine a data collection quota from among the plurality of data collection quotas corresponding to the one input data category from the first subset of the plurality of input data categories, the server controller is further programmed to identify a second subset of the plurality of input data categories. The second subset of the plurality of input data categories includes each of the plurality of input data categories that is not in the first subset of the plurality of input data categories. In order to determine a data collection quota from among the plurality of data collection quotas corresponding to the one input data category from the first subset of the plurality of input data categories, the server controller is further programmed to determine an average steepness of a regression curve equation for each of the second subset of the plurality of input data categories based at least in part on a regression curve equation for each of the second subset of the plurality of input data categories. In order to determine a data collection quota from among the plurality of data collection quotas corresponding to the one input data category from the first subset of the plurality of input data categories, the server controller is further programmed to determine a constant factor α′ and a lower limit model loss γ′ for the one input data category from the first subset of the plurality of input data categories based at least in part on a regression curve equation for the one input data category from the first subset of the plurality of input data categories. In order to determine a data collection quota from among the plurality of data collection quotas corresponding to the one input data category from the first subset of the plurality of input data categories, the server controller is further programmed to determine a constant factor α′ and a lower limit model loss γ′ for the one input data category from the first subset of the plurality of input data categories based at least in part on an average steepness The constant factor α′ and the lower-bound model loss γ′ determine the one data acquisition quota among the plurality of data acquisition quotas using the following predetermined equation.
[0027] In another aspect of the present disclosure, in order to determine one of the plurality of data collection quotas using a predetermined equation, the server controller is further programmed to determine the one of the plurality of data collection quotas using a predetermined equation. The predetermined equation includes:
[0028]
[0029] Where m i+1 is the one data collection quota among multiple data collection quotas; m i is the number of input data samples of the one input data category in the first subset of the plurality of input data categories; α′ is a first constant factor of the one input data category in the first subset of the plurality of input data categories; γ′ is a lower bound model loss for the one input data category in the first subset of the plurality of input data categories; and is the average steepness of the regression curve equation for each input data category in the second subset of the plurality of input data categories.
[0030] In another aspect of the present disclosure, in order to send the data sample collection task to the vehicle using the server communication system, the server controller is further programmed to send the data sample collection task to the vehicle using the server communication system. The data sample collection task includes a validation algorithm describing one of a plurality of input data categories and at least one of a task priority, an expected reduction in model loss, and a number of additional input data samples to be collected.
[0031] In another aspect of the present disclosure, the system also includes a vehicle system including at least one vehicle sensor, a vehicle communication system, and a vehicle controller in electrical communication with the at least one vehicle sensor and the vehicle communication system. The vehicle controller is programmed to receive a data sample collection task from a server system using the vehicle communication system. The vehicle controller is also programmed to prioritize the data sample collection task based at least in part on at least one of a task priority, a projected reduction in model loss, and a number of additional input data samples to be collected. The vehicle controller is also programmed to perform the data sample collection task using the at least one vehicle sensor.
[0032] In another aspect of the present disclosure, to perform the data sample collection task, the vehicle controller is further programmed to record a plurality of unverified input data samples using at least one vehicle sensor. To perform the data sample collection task, the vehicle controller is further programmed to determine a plurality of second input data samples based at least in part on the verification algorithm. The plurality of second input data samples are a subset of the plurality of unverified input data samples. To perform the data sample collection task, the vehicle controller is further programmed to send the plurality of second input data samples to the server communication system using the vehicle communication system.
[0033] In several aspects, the present disclosure provides a method for training a machine learning algorithm for a vehicle. The method includes performing at least one exploratory training session of the machine learning algorithm using an input data set. The input data set includes a plurality of first input data samples. The input data set is stored on a server storage device. The method also includes dividing the input data set into a plurality of input data categories. The method also includes generating an exploratory learning curve plot of a model loss of the machine learning algorithm relative to the number of input data samples trained for each of the plurality of input data categories in the at least one exploratory training session. The method also includes determining a regression curve equation for each of the plurality of input data categories based at least in part on the exploratory learning curve plot for each of the plurality of input data categories. The regression curve equation for one of the plurality of input data categories is a power law equation having the following form:
[0034]
[0035] Where ε is the model loss of the machine learning algorithm for the one of the multiple input data categories; α is the first constant factor for the one of the multiple input data categories; m is the number of input data samples trained for the one of the multiple input data categories; β gis the steepness of the regression curve described by the regression curve equation of the one input data category among the multiple input data categories; and γ is the lower bound model loss of the machine learning algorithm for the one input data category among the multiple input data categories. The method also includes sending a data sample collection task to a vehicle communication system of a vehicle using a server communication system. The method also includes receiving a data sample collection task using a vehicle communication system. The method also includes collecting a plurality of second input data samples using at least one vehicle sensor. The plurality of second input data samples are based at least in part on the regression curve equation of each of the plurality of input data categories. The method also includes sending the plurality of second input data samples from the vehicle communication system to the server communication system. The method also includes training a machine learning algorithm using the plurality of second input data samples and the input data set received from the vehicle communication system.
[0036] In another aspect of the present disclosure, collecting a plurality of second input data samples from the vehicle using the server communication system may also include identifying a first subset of the plurality of input data categories based at least in part on the number of input data samples in each of the plurality of input data categories. Collecting a plurality of second input data samples from the vehicle using the server communication system may also include determining a plurality of data collection quotas based at least in part on a regression curve equation for each of the plurality of input data categories. Each of the plurality of data collection quotas corresponds to an input data category in the first subset of the plurality of input data categories. Collecting a plurality of second input data samples from the vehicle using the server communication system may also include determining a number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories using the following equation:
[0037] s=m i+1 -m i
[0038] Where s is the number of additional input data samples to be collected for the one input data category in the first subset of the plurality of input data categories; m i+1 is a data collection quota for the one input data category in the first subset of the plurality of input data categories in the plurality of data collection quotas; and m i is the number of input data samples of the one input data category in the first subset of the plurality of input data categories. The data sample collection task is sent to the vehicle communication system using the server communication system. The data sample collection task includes at least the number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories.
[0039] In another aspect of the present disclosure, determining a data collection quota from among the plurality of data collection quotas that corresponds to an input data category from a first subset of the plurality of input data categories may further include identifying a second subset of the plurality of input data categories. The second subset of the plurality of input data categories includes each of the plurality of input data categories that is not in the first subset of the plurality of input data categories. Determining a data collection quota from among the plurality of data collection quotas that corresponds to an input data category from the first subset of the plurality of input data categories may further include: determining an average steepness of a regression curve equation for each input data category from the second subset of the plurality of input data categories based at least in part on a regression curve equation for each input data category from the second subset of the plurality of input data categories Determining one of the plurality of data collection quotas corresponding to one of the input data categories in the first subset of the plurality of input data categories may also include: determining a constant factor α′ and a lower bound model loss γ′ of the one of the first subset of the plurality of input data categories based at least in part on a regression curve equation of the one of the first subset of the plurality of input data categories. Determining one of the plurality of data collection quotas corresponding to one of the input data categories in the first subset of the plurality of input data categories may also include: determining the one of the plurality of data collection quotas using a predetermined equation. The predetermined equation includes:
[0040]
[0041] Where m i+1 is the one data collection quota among multiple data collection quotas; m i is the number of input data samples in the one input data category in the first subset of the plurality of input data categories; α′ is a first constant factor for the one input data category in the first subset of the plurality of input data categories; γ′ is a lower bound model loss for the one input data category in the first subset of the plurality of input data categories; and is the average steepness of the regression curve equation for each input data category in the second subset of the plurality of input data categories.
[0042] Further applicability will become apparent from the description provided herein.It should be understood that the description and specific examples are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The drawings described herein are for illustration purposes only and are not intended to limit the scope of the present disclosure in any way.
[0044] Figure 1 is a schematic diagram of a system for training a machine learning algorithm for a vehicle according to an exemplary embodiment;
[0045] Figure 2 is a flow chart of a method for training a machine learning algorithm according to an exemplary embodiment; and
[0046] Figure 3 is a flowchart of a method for receiving a data sample collection task and collecting a plurality of second input data samples according to an exemplary embodiment. DETAILED DESCRIPTION
[0047] The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses.
[0048] Machine learning algorithms have many applications in the fields of vehicle control, autonomous driving, driver assistance systems, etc. Training a machine learning algorithm may require a large amount of data. Collecting and / or storing a large amount of data may require a large amount of resources and / or time. Therefore, the present disclosure provides a new and improved system and method for collecting data for training a machine learning algorithm for a vehicle, which allows data collection to be optimized based on the performance of the machine learning algorithm, thereby reducing training time and / or resources.
[0049] Reference Figure 1 , a system for training a machine learning algorithm for a vehicle is shown and generally designated by the reference numeral 10. The system 10 is shown with an exemplary vehicle 12. Although a passenger vehicle is shown, it should be understood that the vehicle 12 may be any type of vehicle without departing from the scope of the present disclosure. The system 10 generally includes a vehicle system 20 that includes a vehicle controller 22, at least one vehicle sensor 24, and a vehicle communication system 26. The system 10 also includes a server system 30 that includes a server controller 32, a server storage device 34, and a server communication system 36.
[0050] The vehicle controller 22 is used to implement the method 100 for training a machine learning algorithm as described below. The vehicle controller 22 includes at least one processor 40 and a non-transitory computer-readable storage device or medium 42. The processor 40 can be a custom or commercially available processor, a central processing unit (CPU), a graphics processing unit (GPU), an auxiliary processor in a plurality of processors associated with the vehicle controller 22, a semiconductor-based microprocessor (in the form of a microchip or chipset), a macroprocessor, a combination thereof, or generally a device for executing instructions. The computer-readable storage device or medium 42 may include, for example, volatile and non-volatile storage devices in a read-only memory (ROM), a random access memory (RAM), and a keep-alive memory (KAM). KAM is a persistent or non-volatile memory that can be used to store various operating variables when the processor 40 is powered off. The computer readable storage device or medium 42 may be implemented using a plurality of storage devices, such as a programmable read only memory (PROM), an electrical PROM (EPROM), an electrically erasable PROM (EEPROM), flash memory, or other electrical, magnetic, optical, or combination storage device capable of storing data used by the vehicle controller 22 in controlling various systems of the vehicle 12, some of which represents executable instructions. The vehicle controller 22 may also be comprised of multiple controllers in electrical communication with each other. The vehicle controller 22 may be interconnected with additional systems and / or controllers of the vehicle 12, thereby allowing the vehicle controller 22 to access data, such as the speed, acceleration, braking, and steering angle of the vehicle 12.
[0051] The vehicle controller 22 is in electrical communication with at least one vehicle sensor 24 and a vehicle communication system 26. In an exemplary embodiment, the electrical communication is established using a network such as a CAN network, a FLEXRAY network, a local area network (such as WiFi, Ethernet, etc.), a serial peripheral interface (SPI) network, etc. It should be understood that various additional wired and wireless technologies and communication protocols for communicating with the vehicle controller 22 are within the scope of the present disclosure.
[0052] At least one vehicle sensor 24 is used to determine performance data about the vehicle 12. In one exemplary embodiment, the at least one vehicle sensor 24 includes a motor speed sensor, a motor torque sensor, an electric drive motor voltage and / or current sensor, an accelerator pedal position sensor, a coolant temperature sensor, a cooling fan speed sensor, and a transmission oil temperature sensor. In another exemplary embodiment, the plurality of vehicle sensors also includes a sensor for determining information about the environment surrounding the vehicle 12, such as an ambient air temperature sensor, an air pressure sensor, and / or a camera and / or video camera positioned to observe the environment in front of the vehicle 12. In another exemplary embodiment, at least one of the at least one vehicle sensor 24 is capable of measuring distances in the environment surrounding the vehicle 12. In one non-limiting example, the at least one vehicle sensor 24 includes a video camera, and the at least one vehicle sensor 24 measures distances using an image processing algorithm configured to process images from the video camera and determine distances between objects. In another non-limiting example, the at least one vehicle sensor 24 includes a stereo camera with distance measurement capabilities. In one example, at least one of the at least one vehicle sensor 24 is fixed to the interior of the vehicle 12, for example, fixed to the roof lining of the vehicle 12, and its view is observed through the windshield of the vehicle 12. In another example, at least one of the at least one vehicle sensor 24 is fixed to the exterior of the vehicle 12, such as to the roof of the vehicle 12, with a view of the environment surrounding the vehicle 12. It should be understood that various additional types of vehicle sensors, such as LiDAR sensors, ultrasonic ranging sensors, radar sensors, and / or time-of-flight sensors are within the scope of the present disclosure.
[0053] The controller 14 uses the vehicle communication system 26 to communicate with other systems external to the vehicle 12. For example, the vehicle communication system 26 includes the ability to communicate with vehicles ("V2V" communication), infrastructure ("V2I" communication), remote systems at a remote call center (e.g., General Motors' ON-STAR), and / or personal devices. In general, the term "vehicle-to-everything communication" ("V2X" communication) refers to communication between the vehicle 12 and any remote system (e.g., vehicle, infrastructure, and / or remote system). In some embodiments, the vehicle communication system 26 is a wireless communication system that is configured to communicate via a wireless local area network (WLAN) using the IEEE 802.11 standard or by using cellular data communications (e.g., using GSMA standards such as SGP.02, SGP.22, SGP.32, etc.). Therefore, the vehicle communication system 26 may also include an embedded universal integrated circuit card (eUICC) that is configured to store at least one cellular connection configuration profile, such as an embedded subscriber identity module (eSIM) profile. The vehicle communication system 26 is also configured to communicate via a personal area network (e.g., Bluetooth) and / or near field communication (NFC). However, additional or alternative communication methods such as dedicated short-range communication (DSRC) channels and / or mobile telecommunication protocols based on the third generation partnership project (3GPP) standard are also considered to be within the scope of the present disclosure. DSRC channels refer to unidirectional or bidirectional short-range to medium-range wireless communication channels designed specifically for automotive use and a set of corresponding protocols and standards. 3GPP refers to a partnership between multiple standard organizations that develop mobile telecommunication protocols and standards. The structure "version" of the 3GPP standard. Therefore, communication methods based on 3GPP versions 14, 15, 16 and / or future 3GPP versions are considered to be within the scope of the present disclosure. Therefore, the vehicle communication system 26 may include one or more antennas and / or communication transceivers for receiving and / or transmitting signals, such as collaborative sensing messages (CSM). The vehicle communication system 26 is configured to wirelessly transmit information between the vehicle 12 and another vehicle. In addition, the vehicle communication system 26 is configured to wirelessly transmit information between the vehicle 12 and infrastructure or other vehicles. It should be appreciated that the vehicle communication system 26 may be integrated with the controller 14 (eg, on the same circuit board as the controller 14 or a portion of the controller 14 ) without departing from the scope of the present disclosure.
[0054] Continue to refer to Figure 1, a server system is shown and generally indicated by reference numeral 30. The server system 30 includes a server controller 32 in electrical communication with a server storage device 34 and a server communication system 36. In a non-limiting example, the server system 30 is located in a server farm, a data center, etc., and is connected to the Internet. The server controller 32 is used to implement the method 100 for training a machine learning algorithm as described below. The server controller 32 includes at least one server processor 50 and a server non-transitory computer-readable storage device or server medium 52. The description of the types and configurations given above for the vehicle controller 22 also applies to the server controller 32. In a non-limiting example, the server processor 50 and the server medium 52 of the server controller 32 are similar in structure and / or function to the processor 40 and the medium 42 of the vehicle controller 22 as described above. The server communication system 36 is used to communicate with an external system (e.g., the vehicle controller 22) via the vehicle communication system 26. In a non-limiting example, the server communication system 36 is similar in structure and / or function to the vehicle communication system 26 of the vehicle system 20 as described above.
[0055] Reference Figure 2 , a flow chart of a method 100 for training a machine learning algorithm is shown. The method 100 starts at box 102 and proceeds to boxes 104 and 106. At box 104, the server controller 32 performs at least one exploratory training course of the machine learning algorithm. Within the scope of the present disclosure, the machine learning algorithm can be any type of supervised learning algorithm (i.e., a class of machine learning algorithms that learn to make predictions or classifications based on labeled input data), such as computer vision (CV) machine learning algorithms, language processing machine learning algorithms, etc. Within the scope of the present disclosure, at least one exploratory training course involves training the machine learning algorithm using an input data set. In a non-limiting example, the input data set includes a plurality of first input data samples. Within the scope of the present disclosure, the input data samples are training samples (e.g., images of road signs) and accompanying labels for the training samples (e.g., "stop sign" classification labels). One or more exploratory training courses can be performed to determine the accuracy of the machine learning algorithm, as will be discussed in more detail below. After box 104, the method 100 proceeds to boxes 108 and 110, as will be discussed in more detail below.
[0056] At box 106, the server controller 32 identifies at least one data set parameter by which the input data set is classified. In an exemplary embodiment, the at least one data set parameter is a parameter that can vary between individual input data samples. In a non-limiting example, the input data set includes a traffic signal image, and the at least one data set parameter includes ambient light level, traffic signal phase (i.e., traffic signal color), weather conditions (i.e., clear, foggy, rainy, etc.), distance between the vehicle and the traffic light, etc. In an exemplary embodiment, the at least one data set parameter for classifying the input data set is pre-determined and stored in the server storage device 34 for retrieval by the server controller 32. In another exemplary embodiment, the server controller 32 determines the at least one data set parameter by random selection. After box 106, the method 100 proceeds to box 112.
[0057] At box 112, the server controller 32 divides the input data set into a plurality of input data categories based on at least one data set parameter identified at box 106. In an exemplary embodiment, each of the plurality of input data categories includes a subset of a plurality of first input data samples of the input data set. In a non-limiting example, one of the plurality of input data categories is "low ambient light." Thus, the relevant data set parameter is "ambient light," and the input data category "low ambient light" includes a subset of a plurality of first input data samples of the input data set having ambient light below a predetermined ambient light threshold. Thus, at box 112, each of the plurality of first input data samples of the input data set is labeled as one of the plurality of input data categories, and this information is saved in the server storage device 34. After box 112, the method 100 proceeds to boxes 108 and 110.
[0058] At box 108, the server controller 32 generates an exploratory learning curve graph for each of the multiple input data categories determined at box 112 based on the at least one exploratory training session performed at box 104. Within the scope of the present disclosure, a learning curve graph is a graph of the model loss of the machine learning algorithm relative to the number of input data samples trained in the at least one exploratory training session. Within the scope of the present disclosure, the model loss of the machine learning algorithm represents the difference between the predicted output of the machine learning algorithm and the actual output of the given input data sample. At box 108, the server controller 32 generates a plurality of exploratory learning curve graphs, wherein one exploratory learning curve graph corresponds to each of the multiple input data categories. Therefore, the plurality of exploratory learning curve graphs describe the learning curves for each of the multiple input data categories after the at least one exploratory training session. After box 108, the method 100 proceeds to box 114.
[0059] At block 114, the server controller 32 determines a regression curve equation for each of the plurality of input data categories. Within the scope of the present disclosure, a regression curve equation is a mathematical expression that represents the relationship between a dependent variable (i.e., model loss) and an independent variable (i.e., the number of trained input data samples). In an exemplary embodiment, the server controller 32 first determines a mathematical model for the regression curve equation. In a non-limiting example, the mathematical model is predetermined to be a power law equation having the following form:
[0060]
[0061] Where ε is the model loss of the machine learning algorithm for one of the multiple input data categories; α is the first constant factor for the one of the multiple input data categories; m is the number of input data samples trained for the one of the multiple input data categories; β g is the steepness of the regression curve described by the regression curve equation of the one input data category among the multiple input data categories; and γ is the lower bound model loss (i.e., the second constant factor) of the machine learning algorithm for the one input data category among the multiple input data categories.
[0062] In an exemplary embodiment, to determine the regression curve equation for each of the multiple input data categories, the server controller 32 uses a curve fitting method on each of the multiple exploratory learning curve graphs determined in block 108 to fit each of the multiple exploratory learning curve graphs to the above regression curve equation (1). In an exemplary embodiment, the curve fitting method is configured to find a tuple (α, β g ,γ), which minimizes the curve fitting error between the regression curve described by the regression curve equation (1) and the exploratory learning curve graphs for each of the multiple input data categories. In a non-limiting example, the curve fitting method includes at least one of the following methods: least squares method, maximum likelihood method, Bayesian inference method. It should be understood that additional methods for determining a mathematical expression representing the relationship between the model loss and the number of input data samples trained based on multiple exploratory learning curve graphs are within the scope of the present disclosure. After box 114, method 100 proceeds to boxes 116 and 118, as will be discussed in more detail below.
[0063] At block 110, the controller 14 identifies a first subset of the plurality of input data categories. In an exemplary embodiment, the first subset of the plurality of input data categories includes input data categories for which sufficient input data samples have not been collected. At an initial execution of the method 100, in order to determine the first subset of the plurality of input data categories, the server controller 32 compares the number of input data samples in each of the plurality of input data categories with a predetermined sample number threshold (e.g., one hundred samples). If the number of input data samples in a given one of the plurality of input data categories is less than the sample number threshold, the given one of the plurality of input data categories is determined to be included in the first subset of the plurality of input data categories. As will be discussed in more detail below, the method 100 may be repeatedly executed, with some or all of the data from a previous execution of the method 100 being persistently stored in the server storage device 34. At a subsequent execution of the method 100, the server controller 32 compares the number of input data samples in each of the plurality of input data categories with a previous data collection quota. The previous data collection quota is a data collection quota determined during a previous execution of the method 100 and stored in the server storage device 34. The data collection quota is defined and discussed in further detail below. If the number of input data samples in a given one of the plurality of input data categories is less than a previous data collection quota, it is determined that the given one of the plurality of input data categories is included in the first subset of the plurality of input data categories. After block 110, the method 100 proceeds to block 120.
[0064] At block 120, the server controller 32 identifies a second subset of the plurality of input data categories. In an exemplary embodiment, the second subset of the plurality of input data categories includes input data categories for which sufficient input data samples have been collected. In an exemplary embodiment, the second subset of the plurality of input data categories includes each of the plurality of input data categories that are not in the first subset of the plurality of input data categories determined at block 110. After block 120, the method 100 proceeds to blocks 116 and 118.
[0065] At block 116, the server controller 32 determines an average steepness of the plurality of regression curve equations corresponding to each input data category in the second subset of the plurality of input data categories. In other words, the average steepness is a plurality of steepness values of the input data categories for which sufficient input data samples have been collected (ie, a second subset of the plurality of input data categories) The average value (ie, the arithmetic mean) of the values determined at block 114. After block 116, the method 100 proceeds to block 122, as discussed in greater detail below.
[0066] At block 118, the server controller 32 determines a constant factor α′ and a lower bound model loss γ′ for each input data category in the first subset of the plurality of input data categories based on a plurality of regression curve equations corresponding to each input data category in the first subset of the plurality of input data categories determined at block 114. After block 118, the method 100 proceeds to block 122.
[0067] At block 122, the server controller 32 determines a plurality of data collection quotas, wherein each of the plurality of data collection quotas corresponds to an input data category in a first subset of the plurality of input data categories. Within the scope of the present disclosure, the data collection quota for a given input data category is the total number of input data samples for the given input data category that will result in a reduction in model loss for the given input data category by a predetermined loss reduction amount (e.g., fifty percent) when further trained. In an exemplary embodiment, each of the plurality of data collection quotas is determined using a predetermined equation:
[0068]
[0069] Where m i+1 is the one data collection quota among multiple data collection quotas; m i is the number of input data samples for the input data class in the first subset of the plurality of input data classes; α′ is a first constant factor for the one input data class in the first subset of the plurality of input data classes determined at block 118; γ′ is a lower bound model loss for the one input data class in the first subset of the plurality of input data classes determined at block 118; and is the average steepness of the regression curve equation for each input data category in the second subset of the plurality of input data categories determined in block 116. Therefore, it should be understood that in order to reduce the model loss for a given input data category in the first subset of input data categories by half, a certain number of additional input data samples should be collected for the given input data category in the first subset of input data categories:
[0070] s=m i+1 -m i (3)
[0071] Where s is the number of additional input data samples to be collected for one input data category in the first subset of the plurality of input data categories; m i+1 is a data collection quota for the one input data category in the first subset of the plurality of input data categories in the plurality of data collection quotas; and m iis the number of input data samples in the one input data category in the first subset of the plurality of input data categories. Thus, after completion of block 122, method 100 has calculated the number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories in order to reduce the model loss for each input data category in the first subset of the plurality of input data categories by half. It should be appreciated that by collecting additional input data samples (as will be discussed in more detail below) and repeatedly performing method 100, the model loss will be iteratively reduced. After block 122, method 100 proceeds to block 124.
[0072] At box 124, the server controller 32 sends a data sample collection task for each of the first subset of the plurality of input data categories to the vehicle 12 using the server communication system 36. Within the scope of the present disclosure, a data sample collection task is a message containing information for enabling the vehicle controller 22 to collect additional input data samples. In an exemplary embodiment, the data sample collection task includes a description of a validation algorithm for one of the plurality of input data categories, a task priority (i.e., an indication of the priority of the data sample collection task relative to other data sample collection tasks), an expected reduction in model loss (e.g., determined by extrapolation of a regression curve equation), a number of additional input data samples to be collected, a task identifier (i.e., a numeric code that uniquely identifies the data sample collection task), and a sensor identifier (i.e., a numeric code that uniquely identifies which of the at least one vehicle sensor 24 should be used for data collection).
[0073] Within the scope of the present disclosure, a verification algorithm is a mathematical, logical, programming or similar algorithm that is configured to determine whether a particular collected data sample falls into a desired one of a plurality of input data categories. In a non-limiting example, the verification algorithm is a simple logic statement, such as "if light_level is greater than 200, it is true, otherwise it is false". In the above example, the verification algorithm will result in only selecting input data with a light level greater than 200. In other examples, the verification algorithm can be a more complex construction, such as a series of nested logic statements, a set of computer executable instructions (i.e., a computer program or script), or a mathematical function. The verification algorithm will be discussed in more detail below. It should be understood that the data sample collection task may include additional information and / or omit a portion of the above information without departing from the scope of the present disclosure. After box 124, method 100 proceeds to box 126.
[0074] At box 126, the vehicle 12 receives the data sample collection task sent at box 124 and collects a plurality of second input data samples, as will be referred to in Figure 3After frame 126 , the method 100 proceeds to a standby state entering frame 128 .
[0075] In an exemplary embodiment, the server controller 32 repeatedly exits the standby state 128 and restarts the method 100 at block 102. In a non-limiting example, the server controller 32 exits the standby state 128 and restarts the method 100 on a timer, such as once every three hundred milliseconds. By repeatedly executing the method 100, the model loss of each input data category in the first subset of the plurality of input data categories is iteratively reduced.
[0076] Reference Figure 3 , a flow chart of an exemplary embodiment of block 126 is shown. The exemplary embodiment of block 126 begins at block 302. At block 302, the vehicle communication system 26 receives a plurality of data sample collection tasks from the server system 30. In an exemplary embodiment, the plurality of received data sample collection tasks are stored in the medium 42 of the vehicle controller 22. After block 302, the exemplary embodiment of block 126 proceeds to block 304.
[0077] At block 304, the vehicle controller 22 determines a priority for each of the plurality of data sample collection tasks received at block 302. In an exemplary embodiment, the priority is determined based on a task priority of each of the plurality of data sample collection tasks. In an example where the plurality of data sample collection tasks have equal task priorities, the expected reduction in model loss and / or the number of additional input data samples to be collected are used to determine the priority of each of the plurality of data sample collection tasks. After block 304, the exemplary embodiment of block 126 proceeds to block 306.
[0078] At block 306, the vehicle controller 22 uses the at least one vehicle sensor 24 to record a plurality of unverified input data samples. Within the scope of the present disclosure, the unverified input data samples are input data samples recorded by the at least one vehicle sensor 24, but they may not necessarily correspond to any one of the first subset of the plurality of input data categories. After block 306, the exemplary embodiment of block 126 proceeds to block 308.
[0079] In box 308, the vehicle controller 22 uses the verification algorithm as described above to determine multiple second input data samples. Within the scope of the present disclosure, the multiple second input data samples are a subset of multiple unverified input data samples for which the verification algorithm of at least one of the multiple data sample acquisition tasks returns a true value. In other words, the verification algorithm of each of the multiple data sample acquisition tasks is executed with multiple unverified input data samples as input. A subset of the multiple unverified input data samples for which at least one of the verification algorithms of each of the multiple data sample acquisition tasks returns "true" is considered to be the multiple second input data samples. After box 308, the exemplary embodiment of box 126 proceeds to box 310.
[0080] At block 310, the vehicle controller 22 transmits the plurality of second input data samples determined at block 308 to the server communication system 36 using the vehicle communication system 26. In an exemplary embodiment, after receiving the plurality of second input data samples, the server system 30 adds the plurality of second input data samples to the input data set discussed above, generating an updated input data set. The method 100 is then repeated with the updated input data set. After block 310, the exemplary embodiment of block 126 ends, and the method 100 continues as described above.
[0081] The system 10 and method 100 of the present disclosure provide multiple advantages. The system 10 allows communication between the vehicle system 20 and the server system 30 to optimize data collection for training a machine learning algorithm. Using the method 100, the characteristics of the training data collected by the vehicle system 20 are adjusted based on the performance of the machine learning algorithm, thereby reducing training time and reducing resource usage.
[0082] The description of the present disclosure is merely exemplary in nature, and variations that do not depart from the gist of the present disclosure are intended to fall within the scope of the present disclosure. These variations should not be regarded as departing from the spirit and scope of the present disclosure.
Claims
1. A method for training a machine learning algorithm, the method include: performing at least one exploratory training session of the machine learning algorithm using an input data set, wherein the input data set comprises a plurality of first input data samples; dividing the input data set into a plurality of input data categories; determining a regression curve equation for each of the plurality of input data categories based at least in part on the at least one exploratory training session; collecting a plurality of second input data samples based at least in part on a regression curve equation for each of the plurality of input data categories; as well as The machine learning algorithm is trained using the plurality of second input data samples and the input data set.
2. The method according to claim 1, in, Dividing the input data set into a plurality of input data categories further comprises: identifying at least one data set parameter by which to classify the input data set; and The input data set is divided into a plurality of input data categories based on the at least one data set parameter.
3. The method according to claim 1, in, Determining a regression curve equation for each of the plurality of input data categories further comprises: generating an exploratory learning curve plot of the model loss of the machine learning algorithm versus the number of input data samples trained for each of the plurality of input data categories in the at least one exploratory training session; and A regression curve equation for each of the plurality of input data categories is determined based at least in part on the exploratory learning curve graph for each of the plurality of input data categories.
4. The method according to claim 3, in, Determining a regression curve equation for each of the plurality of input data categories further comprises: Determining a regression curve equation for each of the plurality of input data categories based at least in part on an exploratory learning curve graph for each of the plurality of input data categories, wherein the regression curve equation for one of the plurality of input data categories is a power law equation having the following form: Where ε is the model loss of the machine learning algorithm for the one of the multiple input data categories; α is a first constant factor for the one of the multiple input data categories; m is the number of input data samples trained for the one of the multiple input data categories; β g is the steepness of the regression curve described by the regression curve equation of the one of the multiple input data categories; and γ is the lower bound model loss of the machine learning algorithm for the one of the multiple input data categories.
5. The method according to claim 4, in, Collecting a plurality of second input data samples further includes: identifying a first subset of the plurality of input data categories based at least in part on a number of input data samples in each of the plurality of input data categories; determining a plurality of data collection quotas based at least in part on a regression curve equation for each of the plurality of input data categories, wherein each of the plurality of data collection quotas corresponds to an input data category in a first subset of the plurality of input data categories; and A second plurality of input data samples are collected based at least in part on the plurality of data collection quotas.
6. The method according to claim 5, in, Identifying a first subset of the plurality of input data categories further comprises: comparing the number of input data samples in each of the plurality of input data categories to a previous data collection quota, wherein the previous data collection quota is one of the plurality of data collection quotas determined during a previous execution of the method; and Each of the plurality of input data categories having a number of input data samples less than the previous data collection quota is determined as an input data category in a first subset of the plurality of input data categories.
7. The method according to claim 5, in, Determining a data collection quota from the plurality of data collection quotas corresponding to an input data category in the first subset of the plurality of input data categories further includes: identifying a second subset of the plurality of input data categories, wherein the second subset of the plurality of input data categories includes each of the plurality of input data categories that is not in the first subset of the plurality of input data categories; determining an average steepness of the regression curve equation for each of the second subset of the plurality of input data categories based at least in part on the regression curve equation for each of the second subset of the plurality of input data categories determining a constant factor α′ and a lower bound model loss γ′ for the one input data category in the first subset of the plurality of input data categories based at least in part on a regression curve equation for the one input data category in the first subset of the plurality of input data categories; and Based at least in part on the average steepness The constant factor α′ and the lower bound model loss γ′ determine the one data collection quota among the plurality of data collection quotas using a predetermined equation.
8. The method according to claim 7, in, Determining one of the plurality of data collection quotas using the predetermined equation further includes: The one data collection quota among the plurality of data collection quotas is determined using the predetermined equation, wherein the predetermined equation comprises: Where m i+1 is the one data collection quota among the multiple data collection quotas; m i is the number of input data samples of the one input data category in the first subset of the plurality of input data categories; α′ is a first constant factor of the one input data category in the first subset of the plurality of input data categories; γ′ is the lower bound model loss of the one input data category in the first subset of the plurality of input data categories; and is the average steepness of the regression curve equation for each input data category in the second subset of the plurality of input data categories.
9. The method according to claim 8, in, Collecting the plurality of second input data samples further includes: The number of additional input data samples to be collected for each input data category in the first subset of the plurality of input data categories is determined using the following equation: s=m i+1 -m i wherein s is the number of additional input data samples to be collected for the one input data category in the first subset of the plurality of input data categories; m i+1 is the one data collection quota of the plurality of data collection quotas for the one input data category of the first subset of the plurality of input data categories; and m i is the number of input data samples in the one input data category in the first subset of the plurality of input data categories; and A data sample collection task is sent to the vehicle, wherein the data sample collection task includes at least a number of additional input data samples to be collected for each input data category in a first subset of the plurality of input data categories.
10. The method according to claim 9, in, Training the machine learning algorithm using the plurality of second input data samples and the input data set further comprises: receiving the plurality of second input data samples from the vehicle; generating an updated input data set, wherein the updated input data set includes the plurality of second input data samples and the input data set; and The method is performed using the updated input data set.
