WO2017148314A1 - 一种机器学习系统的训练方法和训练系统 - Google Patents

一种机器学习系统的训练方法和训练系统 Download PDF

Info

Publication number
WO2017148314A1
WO2017148314A1 PCT/CN2017/074318 CN2017074318W WO2017148314A1 WO 2017148314 A1 WO2017148314 A1 WO 2017148314A1 CN 2017074318 W CN2017074318 W CN 2017074318W WO 2017148314 A1 WO2017148314 A1 WO 2017148314A1
Authority
WO
WIPO (PCT)
Prior art keywords
loss function
training
current
weight
value
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2017/074318
Other languages
English (en)
French (fr)
Inventor
周俊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Priority to JP2018546445A priority Critical patent/JP7004661B2/ja
Priority to EP17759165.8A priority patent/EP3425527A4/en
Publication of WO2017148314A1 publication Critical patent/WO2017148314A1/zh
Priority to US16/119,585 priority patent/US11257005B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/045—Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence

Definitions

  • the present application relates to the field of big data processing, and in particular, to a training method and a training system for a machine learning system.
  • each website can use the machine learning system to collect behavior data of users online, and provide different search results for users of different genders/ages, so as to provide services to the greatest extent according to the preferences of users.
  • the machine learning system described above can collect a large number of user behavior data, such as user browsing/clicking/purchasing/top posting/posting comments on posts, etc., and using some machine learning algorithms to train the machine learning system offline. After getting a predictive model, put it on the online system to provide services to users. For example, Amazon's recommendation system can recommend different products to different users after learning user preferences based on the user's past behavior logs (such as logs containing yesterday and yesterday and pushing forward for 2 weeks).
  • Online learning means that the model continuously updates the model itself as new training data is received. Therefore it can achieve better results.
  • the accuracy of online learning is relatively high, but because the time period of data collection is relatively short, the model is affected by the noise in the data, which makes the model training unstable.
  • the online learning adopts the method of asynchronous update model, the training data continuously flows into the model, and the model continues to Performing update learning often results in model training stuck due to some uncontrollable problems, resulting in training failures, resulting in failure to produce available models, affecting online services, and damaging the user experience.
  • embodiments of the present application have been made in order to provide a training method and training system for a machine learning system that overcomes the above problems or at least partially solves the above problems.
  • the present application discloses a training method of a machine learning system, which uses the training data to train the machine learning system, and the training method includes:
  • the last summarized weight and the last summarized loss function value are input into the machine learning system for training;
  • the current weight and/or the current loss function value is corrected to a current weight and/or a current loss function value within the first threshold, and input into the machine learning system for training. .
  • a training method and a training system for a machine learning system by processing each work machine only for one piece of data at a time, and summarizing the local weight and the local loss function value calculated according to the data piece The effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value, when the obtained current weight and/or the current loss function value is detected
  • the current weight and/or the current loss function value calculated this time may be discarded, and the weight and loss function values obtained from the previous summary may be input into the machine learning model, or the current calculation.
  • the current weight and/or the current loss function value are corrected, and the update of the weight and loss function values obtained by each calculation and summary is limited to the effective range, thereby avoiding the instability of the model training and enhancing the anti-data noise of the system.
  • the ability to improve the stability of the online learning machine learning system ensures the normal output of the machine learning system and improves the service capabilities of the machine learning system after the launch.
  • FIG. 1 is a flow chart of a training method of a machine learning system of a first embodiment of the present application.
  • FIG. 2 is a flow chart of a training method of the machine learning system of the second embodiment of the present application.
  • FIG. 3 is a flow chart of a training method of the machine learning system of the third embodiment of the present application.
  • FIG. 4 is a block diagram of a training system of a machine learning system of a fourth embodiment of the present application.
  • Figure 5 is a block diagram of a training system of a machine learning system of a fifth embodiment of the present application.
  • Figure 6 is a block diagram of a training system of a machine learning system of a sixth embodiment of the present application.
  • One of the core ideas of the present application is to propose a training method of a machine learning system, in which training data is first allocated to a plurality of working machines; and training data allocated by each working machine is divided into a plurality of data.
  • the function value performs model anomaly detection; when the detection result is the first type of abnormality, the last summed weight and the last summarized loss function value are input into the machine learning system for training; when the detection result is The second type of abnormality corrects the current weight and/or the current loss function value to the current weight and/or the current loss function value within the first threshold, and inputs the machine learning system for training.
  • FIG. 1 is a flowchart of a training method of a machine learning system according to a first embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, including a parameter server, which may include, for example, a plurality of workers (workers or slaves), a plurality of servers (servers), and a coordinator.
  • a parameter server which may include, for example, a plurality of workers (workers or slaves), a plurality of servers (servers), and a coordinator.
  • the training method includes the following steps:
  • Step S101 the training data is allocated to a plurality of working machines
  • each work machine can read its own training data according to its identification, and the data between the working machines does not overlap.
  • This step may, for example, be divided by the coordinator into training data belonging to each working machine according to the id of each working machine, and the training data is read by the corresponding working machine.
  • Step S102 dividing the training data allocated by each working machine into a plurality of data pieces
  • each work machine divides its assigned training data into a plurality of pieces of data based on its own system configuration, and each processing of each working machine is processed based on one piece of data. That is, each work machine processes only one piece of data at a time.
  • Step S103 obtaining a local weight and a local loss function value calculated by each working machine based on each piece of data
  • the local weight and the local loss function value corresponding to the calculation of the working machine are obtained;
  • the learning system includes a plurality of working machines, and in this step, local weights and local loss function values obtained by the plurality of working machines in the calculation can be obtained.
  • Step S104 summarizing the local weight and the local loss function value calculated by each working machine based on each piece of data, and obtaining the current weight and the current loss function value;
  • each working machine sends the calculated local weight and local loss function values to the server, and the server pairs the plurality of local weights and local loss function values calculated by the plurality of working machines at the current time. Summarize and get the current weight and current loss function values.
  • the process of aggregation includes, but is not limited to, addition, averaging after addition, and the like, and the present invention is not particularly limited.
  • each of the working machines calculates the local weight and the local loss function value in step S103, it may be asynchronously transmitted to the server, and in step S104, the coordinator driver server may perform aggregation of the plurality of local weights and local loss functions.
  • Step S105 performing model anomaly detection by using a current weight and/or a current loss function value
  • the summarized current weight and/or the current loss function value may be input to the model anomaly detecting device to detect whether the model is abnormal according to a preset condition.
  • the result of the detection may be, for example, a normal, a first type of abnormality, and a second type of abnormality, and corresponding processing is performed separately for the three cases.
  • the input values of the model anomaly detection device are the current weight, the current loss function value, the last summarized weight, and the last summarized loss function value.
  • the last summed weight is obtained by summarizing the last local weight calculated by each work machine based on the data piece used in the previous calculation; the last summary loss function value is used by each work machine based on the last calculation.
  • the last local loss function value calculated from the data slice is obtained by summarizing.
  • Step S106 when the result of the model abnormality detection is a first type of abnormality, input the weight of the last summary and the value of the last summarized loss function into the machine learning system for training;
  • the local weight and local loss function values obtained by the current calculation of each working machine and the current weight and current loss function values after the summary are ignored or deleted.
  • the summed weights and the last summarized loss function values are entered into the machine learning system for training.
  • Step S107 when the result of the model abnormality detection is a second type of abnormality, correct the current weight and/or the current loss function value to a current weight and/or a current loss function value within the first threshold, and input the machine learning.
  • the system is trained.
  • the detection result is a second type of abnormality
  • the model fine adjustment device can be entered, and the current weight and/or the current loss function value is subtracted from the preset first threshold, and the value of each element is obtained first.
  • the current weight and/or current loss function value within the threshold is entered into the machine learning system for training.
  • the processing of subtracting the first threshold may be performed one or more times. Until the value of each element is within the first threshold.
  • each work machine is processed only for one piece of data at a time, and is summarized according to the local weight and the local loss function value calculated by the data piece.
  • the effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value.
  • the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • FIG. 2 is a flowchart of a training method of the machine learning system according to the second embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, as shown in FIG. 2, the training method includes the following steps:
  • S205 Perform model anomaly detection by using a current weight and/or a current loss function value
  • step S205 that is, the step of performing model anomaly detection by using the current weight and/or the current loss function value comprises:
  • Sub-step S205a determining whether the current loss function value is greater than k times the value of the loss function after the last summary
  • the value k can be set in advance as a basis for determining whether the current loss function value is abnormal.
  • Sub-step S205b when it is determined whether the current loss function value is greater than k times the value of the loss function that was last summarized, it is determined that the result of the model abnormality detection is the first type of abnormality.
  • the current loss function value calculated and summarized in this time is the first type of abnormality, that is, Serious anomaly.
  • the current weight and the current loss function value calculated this time can be directly discarded.
  • step S205 that is, the step of performing the model abnormality detection by using the current weight and/or the current loss function value may further include:
  • Sub-step S205c determines whether the absolute value of the maximum value and/or the minimum value of the element in the current weight is greater than a preset second threshold
  • the current weight is a vector consisting of N elements, and after summarizing the current weights, the absolute value of the maximum value and/or the minimum value of each element of the vector (ie, each dimension) can be determined. Whether it is greater than a preset second threshold; the second threshold may be specified in advance.
  • Sub-step S205d when it is determined whether the absolute value of the maximum value and/or the minimum value of the element in the current weight is greater than a preset second threshold, determining that the result of the model abnormality detection is a second type of abnormality.
  • the model when it is determined that the absolute value of at least one of the maximum value and the minimum value of the elements in the current weight is greater than a preset second threshold, the model is considered to have a second type of abnormality, ie Relatively minor anomalies.
  • the current weight and the current loss function value calculated by the current calculation may be corrected to be corrected to the current weight and/or the current loss function value within the first threshold, and input into the machine learning system for training.
  • the step of modifying the current weight and/or the current loss function value to a local weight and/or a local loss function value within the first threshold may include:
  • Subtracting the element whose absolute value of the maximum value and/or the minimum value is greater than the preset second threshold value by the second threshold value obtains a current weight in which each element is located within the second threshold.
  • step S205 that is, the step of performing the model abnormality detection by using the current weight and/or the current loss function value may further include:
  • Sub-step S205e determining whether the average value of the loss function values after the first n times is greater than the loss function value after the first summary;
  • the loss function values obtained by the first n times are stored, for example, in a specific storage area, and the average value of the loss function values after the first n times of aggregation can be calculated, and whether the value is greater than the first summary is determined. The resulting loss function value.
  • Sub-step S205f when it is determined that the average value of the loss function values after the first n times is larger than the loss function value after the first summary, the result of the model abnormality detection is determined to be the first category often.
  • the model has the first type of abnormality, that is, a more serious abnormality.
  • the current weight and the current loss function value calculated this time can be directly discarded.
  • each work machine is processed only for one piece of data at a time, and is summarized according to the local weight and the local loss function value calculated by the data piece.
  • the effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value.
  • the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • FIG. 3 is a flowchart of a training method of a machine learning system according to a third embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, as shown in FIG. 3, the training method includes the following steps:
  • the training data allocated by each working machine is divided into multiple pieces of data.
  • S305 Perform model anomaly detection by using a current weight and/or a current loss function value
  • the training method may further include:
  • the current weight and/or the current loss function value may be used as a valid calculation result, and the The machine learning system is trained.
  • step S308 that is, when the result of the model abnormality detection is normal, the current weight and/or the current loss function value are input into the machine learning system for training, the The method also includes:
  • the third threshold may be a threshold detected by the system.
  • the judging step is not limited to being performed by comparing with the difference and the third threshold, and there are various verification trainings in the field whether the training can be ended. The method will not be described here.
  • the model is produced, for example, the model can be applied online.
  • the step of obtaining the local weight and the local loss function value calculated by each work machine based on each piece of data may be repeatedly performed, that is, each working machine continues to calculate the local weight and the local loss function based on other pieces of data.
  • the value, and the current weight and current loss function values are obtained in subsequent summaries, and the training method is repeated until a model that satisfies the usage requirements is obtained.
  • the method further includes:
  • the step of allocating the training data to a plurality of working machines further includes:
  • the training data is distributed to a plurality of working machines in accordance with the identification of each working machine.
  • the summary each work machine obtains the current weight and the current loss function based on the local weight and the local loss function value calculated for each piece of data.
  • the steps of value can include:
  • the local loss function values obtained by each work machine based on each piece of data are added to obtain a current local loss function.
  • each The working machine only processes one piece of data at a time, and summarizes the local weight and the local loss function value calculated by the data piece, and achieves the effect of accurately updating the weight and loss function value after each calculation. It avoids the problem that the prior art online learning is susceptible to noise in the data.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value. When it is detected that the obtained current weight and/or the current loss function value belongs to the first type of abnormality or the second type of abnormality, the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • FIG. 4 is a block diagram of a training system of a machine learning system according to a fourth embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, as shown in FIG. 4, the training system 400 includes:
  • a training data distribution module 401 configured to allocate the training data to a plurality of working machines
  • the training data dividing module 402 is configured to divide the training data allocated by each working machine into a plurality of data pieces;
  • the obtaining module 403 is configured to obtain a local weight and a local loss function value calculated by each working machine based on each piece of data;
  • a summary module 404 configured to summarize the local weight and local loss function values calculated by each working machine based on each piece of data, to obtain a current weight and a current loss function value;
  • An abnormality detecting module 405, configured to perform model anomaly detection by using a current weight and/or a current loss function value
  • the first processing module 406 is configured to: when the result of the model abnormality detection is a first type of abnormality, input the last aggregated weight and the last aggregated loss function value into the machine learning system for training;
  • a second processing module 407 configured to: when the result of the model abnormality detection is a second type of abnormality, correct the current weight and/or the current loss function value to a current weight and/or a current loss function value within the first threshold, and Enter the machine learning system for training.
  • each work machine is processed only for one piece of data at a time, and is summarized according to the local weight and the local loss function value calculated by the data piece.
  • the effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value.
  • the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • the fifth embodiment of the present application provides a training system for a machine learning system
  • FIG. 5 is a block diagram of a training system of the machine learning system of the fifth embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, as shown in FIG. 5, the training system 500 includes:
  • a training data distribution module 501 configured to allocate the training data to a plurality of working machines
  • the training data dividing module 502 is configured to divide the training data allocated by each working machine into a plurality of data pieces;
  • the obtaining module 503 is configured to obtain a local weight and a local loss function value calculated by each working machine based on each piece of data;
  • a summary module 504 configured to summarize the local weight and the local loss function value calculated by each working machine based on each piece of data, to obtain a current weight and a current loss function value;
  • An abnormality detecting module 505, configured to perform model anomaly detection by using a current weight and/or a current loss function value
  • the first processing module 506 is configured to: when the result of the model abnormality detection is a first type of abnormality, input the last aggregated weight and the last aggregated loss function value into the machine learning system for training;
  • a second processing module 507 configured to: when the result of the model abnormality detection is a second type of abnormality, correct the current weight and/or the current loss function value to a current weight and/or a current loss function value within the first threshold, and Enter the machine learning system for training.
  • the abnormality detecting module 505 includes:
  • the first determining sub-module 505a is configured to determine whether the current loss function value is greater than k times the value of the loss function after the last summary;
  • the first determining sub-module 505b is configured to determine that the result of the model abnormality detection is a first type of abnormality when it is determined that the current loss function value is greater than k times of the last summarized loss function value.
  • a second determining sub-module 505c configured to determine whether an absolute value of a maximum value and/or a minimum value of an element in the current weight is greater than a preset second threshold
  • the second determining sub-module 505d is configured to determine that the result of the abnormality detection of the model is a second type of abnormality when it is determined that the absolute value of the maximum value and/or the minimum value of the element in the current weight is greater than a preset second threshold. .
  • the abnormality detecting module 505 further includes:
  • the third determining sub-module 505e is configured to determine whether the average value of the loss function values after the first n times is greater than the loss function value after the first summary;
  • the third determining sub-module 505f is configured to determine that the result of the model abnormality detection is a first type of abnormality when it is determined that the average value of the loss function values after the first n times of aggregation is greater than the loss function value after the first summary.
  • the step of modifying the current weight and/or the current loss function value to the current weight and/or the current loss function value within the second threshold comprises:
  • Subtracting the element whose absolute value of the maximum value and/or the minimum value is greater than the preset second threshold value by the second threshold value obtains a current weight in which each element is located within the second threshold.
  • each work machine is processed only for one piece of data at a time, and is summarized according to the local weight and the local loss function value calculated by the data piece.
  • the effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value.
  • the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • FIG. 6 is a block diagram showing a training system of a machine learning system according to a sixth embodiment of the present application.
  • the machine learning system is preferably a distributed machine learning system, as shown in FIG. 6, the training system includes:
  • a training data distribution module 601, configured to allocate the training data to a plurality of working machines
  • the training data dividing module 602 is configured to divide the training data allocated by each working machine into a plurality of data pieces;
  • the obtaining module 603 is configured to obtain a local weight and a local loss function value calculated by each working machine based on each piece of data;
  • a summary module 604 configured to summarize the local weight and local loss function values calculated by each working machine based on each piece of data, to obtain a current weight and a current loss function value;
  • An abnormality detecting module 605, configured to perform model anomaly detection by using a current weight and/or a current loss function value
  • the first processing module 606 is configured to: when the result of the model abnormality detection is a first type of abnormality, input the last aggregated weight and the last aggregated loss function value into the machine learning system for training;
  • a second processing module 607 configured to: when the result of the model abnormality detection is a second type of abnormality, correct the current weight and/or the current loss function value to a current weight and/or a current loss function value within the first threshold, and Enter the machine learning system for training.
  • system further includes:
  • the training module 608 is configured to input the current weight and/or the current loss function value into the machine learning system for training when the result of the model abnormality detection is normal.
  • system further includes:
  • End determination module 609 configured to determine whether the training ends
  • a model output module 610 configured to output the model when the training is determined to be ended
  • the execution module 611 is configured to, when determining that the training is not finished, repeat the step of acquiring the local weight and the local loss function value calculated by each working machine based on each piece of data.
  • the end determination module is configured to determine that the training ends when the value of the current loss function is detected and the value of the loss function is less than the third threshold after the last summary.
  • system may further include:
  • a training data acquisition module is used to acquire training data.
  • the training data distribution module is configured to allocate the training data to a plurality of working machines according to the identifier of each working machine.
  • the summary module comprises:
  • a weight accumulating module configured to add each local machine weight calculated by each working machine based on each piece of data to obtain a current weight
  • the loss function value accumulating module is configured to add the local loss function values calculated by each working machine based on each piece of data to obtain a current local loss function.
  • each work machine is processed only for one piece of data at a time, and is summarized according to the local weight and the local loss function value calculated by the data piece.
  • the effect of accurately updating the weight and loss function values after each calculation is achieved, and the problem that the prior art online learning is susceptible to noise in the data is avoided.
  • the abnormality detection is performed after obtaining the current weight and/or the current loss function value.
  • the calculated calculation may be discarded.
  • the current weight and/or the current loss function value, and the weight and loss function values obtained from the previous summary are input into the machine learning model, or the current weight and/or the current loss function value calculated this time are corrected, and each calculation is performed.
  • the update of the weight and loss function values obtained by the summary is limited to the effective range, avoiding the instability of the model training, enhancing the system's ability to resist data noise, improving the stability of the online learning machine learning system, and ensuring the machine. Learning the normal output of the system improves the service capabilities of the machine learning system after the launch.
  • the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment.
  • embodiments of the embodiments of the present application can be provided as a method, apparatus, or computer program product. Therefore, the embodiments of the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Moreover, embodiments of the present application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.
  • computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • the computer device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
  • the memory may include non-persistent memory, random access memory (RAM), and/or non-volatile memory in a computer readable medium, such as read only memory (ROM) or flash memory.
  • RAM random access memory
  • ROM read only memory
  • Memory is an example of a computer readable medium.
  • Computer readable media including both permanent and non-persistent, removable and non-removable media may be implemented by any method or technology for signal storage.
  • the signals can be computer readable instructions, data structures, modules of programs, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory. (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disk read only memory (CD-ROM), digital versatile disk (DVD) or other optical storage,
  • PRAM phase change memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • RAM random access memory
  • ROM read only memory
  • EEPROM electrically erasable programmable read only memory
  • flash memory or other memory technology
  • compact disk read only memory CD-ROM
  • DVD digital versatile disk
  • a magnetic tape cartridge, magnetic tape storage or other magnetic storage device or any other non-transporting medium can be used to store signals that can be accessed by a computing device.
  • computer readable media does not include non-persistent computer readable media, such as modulated data signals and carrier waves.
  • Embodiments of the present application are described with reference to flowcharts and/or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block of the flowchart illustrations and/or FIG.
  • These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing terminal device to produce a machine such that instructions are executed by a processor of a computer or other programmable data processing terminal device
  • Means are provided for implementing the functions specified in one or more of the flow or in one or more blocks of the flow chart.
  • the computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device.
  • the instruction device implements the functions specified in one or more blocks of the flowchart or in a flow or block of the flowchart.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Databases & Information Systems (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Numerical Control (AREA)

Abstract

一种机器学习系统的训练方法和系统，利用训练数据对所述机器学习系统进行训练，所述训练方法包括：将训练数据分配给多个工作机（S101）；将每一个工作机分配得到的训练数据划分为多个数据片(S102)；获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值(S103)；汇总这些本地权重和本地损失函数值，获得当前权重和当前损失函数值(S104)；利用当前权重和/或当前损失函数值执行模型异常检测(S105)；当检测结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练(S106)；当检测结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入机器学习系统进行训练(S107)。

Description

一种机器学习系统的训练方法和训练系统 技术领域
本申请涉及大数据处理领域，尤其涉及一种机器学习系统的训练方法和训练系统。
背景技术
现如今，服务好用户是各互联网公司都希望实现的目标。为此，绝大部分公司都采用机器学习的方式捕捉用户的偏好和习惯，提供个性化的服务。举例来说，各网站可以利用机器学习系统，收集用户上网的行为数据，为不同性别/年龄的用户提供不同的搜索结果，做到最大程度地根据用户的喜好提供服务。
上述的机器学习系统可以收集大量用户的行为数据，例如用户浏览/点击/购买/顶帖子/发表对帖子的意见等等，并利用某些机器学习算法，离线对机器学习系统进行训练。在得到一个预测模型之后，放在线上系统，对用户提供服务。例如亚马逊的推荐系统，其可以根据用户以往的行为日志(比如包含昨天以及昨天再向前推2周的日志)，学习用户偏好之后，给不同用户推荐不同的商品。
这种利用以往用户行为数据，离线进行训练，得到模型的方式称为离线学习。然而正常情况下用户行为和兴趣可能在一天之内都会发生显著变化，由于离线学习系统中用户当下产生的日志并不能马上加入到系统中去训练，离线学习显然不能捕获这种变化。
在线学习就是模型随着接收的新训练数据，不断更新模型本身。因而它能取得更好的效果。在线学习精度比较高，但由于数据采集的时间段比较短，模型受数据中噪声影响比较大，造成模型训练不稳定；同时由于在线学习采用异步更新模型的方式，训练数据不断流入模型，模型不断进行更新学习，经常由于某些不可控的问题导致模型训练卡死，出现训练失败，导致无法产出可用的模型、影响线上服务、损害用户体验。
发明内容
鉴于上述问题，提出了本申请实施例以便提供一种克服上述问题或者至少部分地解决上述问题的机器学习系统的训练方法和训练系统。
为解决上述问题，本申请公开一种机器学习系统的训练方法，利用训练数据对所述机器学习系统进行训练，所述训练方法包括：
将所述训练数据分配给多个工作机；
将每一个工作机分配得到的训练数据划分为多个数据片；
获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
利用当前权重和/或当前损失函数值执行模型异常检测；
当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
本申请实施例至少具有以下优点：
根据本申请实施例提出的机器学习系统的训练方法和训练系统，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属 于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
附图说明
图1是本申请第一实施例的机器学习系统的训练方法的流程图。
图2是本申请第二实施例的机器学习系统的训练方法的流程图。
图3是本申请第三实施例的机器学习系统的训练方法的流程图。
图4是本申请第四实施例的机器学习系统的训练系统的方框图。
图5是本申请第五实施例的机器学习系统的训练系统的方框图。
图6是本申请第六实施例的机器学习系统的训练系统的方框图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员所获得的所有其他实施例，都属于本申请保护的范围。
本申请的核心思想之一在于，提出一种机器学习系统的训练方法，在该方法中，首先将训练数据分配给多个工作机；将每一个工作机分配得到的训练数据划分为多个数据片；获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；汇总这些本地权重和本地损失函数值，获得当前权重和当前损失函数值；利用当前权重和/或当前损失函数值执行模型异常检测；当检测结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；当检测结果为第 二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入机器学习系统进行训练。
第一实施例
本申请第一实施例提出一种机器学习系统的训练方法，如图1所示为本申请第一实施例的机器学习系统的训练方法的流程图。该机器学习系统优选为分布式机器学习系统，包括参数服务器(parameter server)，该参数服务器例如可以包括多个工作机(worker或者slaver)、多个服务器(server)、一个协调器(coordinator)。如图1所示，该训练方法包括如下步骤：
步骤S101，将所述训练数据分配给多个工作机；
在这一步骤中，例如，每个工作机可以根据其标识读取属于自己的训练数据，各工作机之间的数据不重叠。这一步骤例如可以由协调器按照每个工作机的id将训练数据划分为属于每一个工作机的训练数据，并由对应的工作机读取该训练数据。
步骤S102，将每一个工作机分配得到的训练数据划分为多个数据片；
在这一步骤中，例如，每一个工作机基于其自身的系统配置将其分配得到的训练数据划分为多个数据片，每一个工作机的每次处理是基于一个数据片进行处理。即，每个工作机每次只处理一个数据片。
步骤S103，获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
在这一步骤中，例如，在每一个工作机中，基于该工作机获得的训练数据中的一个数据片，计算获得该工作机该次计算对应的本地权重和本地损失函数值；由于该机器学习系统包括多个工作机，在这一步骤中，可以获得多个工作机在该次计算中获得的本地权重和本地损失函数值。
步骤S104，汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
在这一步骤中，例如，每一个工作机将计算出的本地权重和本地损失函数值发送至服务器，由服务器对由多个工作机在本次计算出的多个本地权重和本地损失函数值进行汇总，获得当前权重和当前损失函数值。汇总的过程包括但不限于相加、相加后平均等，本发明并不特别限制。
在步骤S103中每一个工作机计算出本地权重和本地损失函数值之后，可以异步地发送给服务器，在步骤S104中，可以由协调器驱动服务器进行多个本地权重和本地损失函数的汇总。
步骤S105，利用当前权重和/或当前损失函数值执行模型异常检测；
在这一步骤中，例如，可以将汇总后的当前权重和/或当前损失函数值输入模型异常检测装置，根据预设的条件对模型是否异常进行检测。检测的结果例如可以为正常、第一类异常和第二类异常，并在后续分别针对这三种情况进行对应处理。
在一实施例中，模型异常检测装置的输入值为当前权重、当前损失函数值、上一次汇总后的权重和上一次汇总后的损失函数值。上一次汇总后的权重由每一个工作机基于上一次计算使用的数据片计算得出的上一次本地权重进行汇总获得；上一次汇总后的损失函数值由每一个工作机基于上一次计算使用的数据片计算得出的上一次本地损失函数值进行汇总获得。
步骤S106，当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
在这一步骤中，例如检测结果为第一类异常，则忽略或删除每一个工作机本次计算获得的本地权重和本地损失函数值以及汇总后的当前权重和当前损失函数值，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练。
步骤S107，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
在这一步骤中，例如检测结果是第二类异常，可以进入模型微调装置，将当前权重和/或当前损失函数值减去预设的第一阈值，获得每一个元素的数值均位于第一阈值内的当前权重和/或当前损失函数值，并输入机器学习系统进行训练。特别地，当第一次减去预设的第一阈值仍不能得到位于第一阈值内的当前权重和/或当前损失函数值时，则可以再进行一次或多次减去第一阈值的处理，直到每一个元素的数值均位于第一阈值内。
在本申请第一实施例提出的机器学习系统的训练方法中，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
第二实施例
本申请第二实施例提出一种机器学习系统的训练方法，如图2所示为本申请第二实施例的机器学习系统的训练方法的流程图。该机器学习系统优选为分布式机器学习系统，如图2所示，该训练方法包括如下步骤：
S201，将所述训练数据分配给多个工作机；
S202，将每一个工作机分配得到的训练数据划分为多个数据片；
S203，获取每一个工作机基于每一个数据片计算出的本地权重和本地 损失函数值；
S204，汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
S205，利用当前权重和/或当前损失函数值执行模型异常检测；
S206，当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
S207，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
上述步骤S201至S207与第一实施例中的步骤S101至S107相同的部分在此不再赘述。
在本实施例中，步骤S205，即所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤包括：
子步骤S205a，判断当前损失函数值是否大于上一次汇总后的损失函数的值的k倍；
在这一子步骤中，可以预先设置数值k，作为判定当前损失函数值是否异常的判定依据。
子步骤S205b，当判断出当前损失函数值是否大于上一次汇总后的损失函数的值的k倍时，判定所述模型异常检测的结果为第一类异常。
在这一子步骤中，当判断出当前损失函数值大于上一次汇总后的损失函数的值的k倍时，可以认为本次计算并汇总出的当前损失函数值为第一类异常，即较为严重的异常。当判断出模型出现这一类异常，则可以直接抛弃本次计算出的当前权重和当前损失函数值。
在本实施例中，步骤S205，即所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤还可以包括：
子步骤S205c判断当前权重中的元素的最大值和/或最小值的绝对值是否大于预设的第二阈值；
在这一子步骤中，当前权重为包括N个元素的向量，在汇总得出当前权重后，可以判断该向量的每一个元素(即每一个维度)的最大值和/或最小值的绝对值是否大于预设的第二阈值；第二阈值可以预先指定。
子步骤S205d，当判断出当前权重中的元素的最大值和/或最小值的绝对值是否大于预设的第二阈值时，判定所述模型异常检测的结果为第二类异常。
在这一子步骤中，当判断出当前权重中的元素的最大值和最小值中的至少一者的绝对值大于预设的第二阈值时，则认为所述模型存在第二类异常，即相对较为轻微的异常。可以通过修正保留本次计算出的当前权重和当前损失函数值，将其修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
优选地，将当前权重和/或当前损失函数值修正为第一阈值内的本地权重和/或本地损失函数值的步骤可以包括：
将所述最大值和/或最小值的绝对值大于预设的第二阈值的元素减去所述第二阈值，获得每一个元素均位于所述第二阈值内的当前权重。
在本实施例中，步骤S205，即所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤还可以包括：
子步骤S205e，判断前n次汇总后的损失函数值的平均值是否大于第一次汇总后的损失函数值；
在这一子步骤中，前n次汇总得出的损失函数值例如存储在特定的存储区域，可以计算前n次汇总后的损失函数值的平均值，并判断该值是否大于第一次汇总得出的损失函数值。
子步骤S205f，当判断出前n次汇总后的损失函数值的平均值大于第一次汇总后的损失函数值时，判定所述模型异常检测的结果为第一类异 常。
在这一子步骤中，当判断出前n次汇总后的损失函数值的平均值大于第一次汇总后的损失函数值时，则可以认为该模型存在第一类异常，即较为严重的异常，当判断出模型出现这一类异常，则可以直接抛弃本次计算出的当前权重和当前损失函数值。
在本申请第二实施例提出的机器学习系统的训练方法中，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
第三实施例
本申请第三实施例提出一种机器学习系统的训练方法，如图3所示为本申请第三实施例的机器学习系统的训练方法的流程图。该机器学习系统优选为分布式机器学习系统，如图3所示，该训练方法包括如下步骤：
S301，将所述训练数据分配给多个工作机；
S302，将每一个工作机分配得到的训练数据划分为多个数据片；
S303，获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
S304，汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
S305，利用当前权重和/或当前损失函数值执行模型异常检测；
S306，当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
S307，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
上述步骤S301至S307与第一实施例中的步骤S101至S107相同的部分在此不再赘述。
在本实施例中，在步骤S305即所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤之后，所述训练方法还可以包括：
S308，当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练。
在这一实施例中，当检测出本次计算并汇总后的当前权重和当前损失函数值下该模型正常，则可以将当前权重和/或当前损失函数值作为有效的计算结果，输入所述机器学习系统进行训练。
进一步地，在该实施例中，在步骤S308即所述当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练的步骤之后，所述方法还包括：
S309，判断训练是否结束；
在该步骤中，例如，当检测到当前损失函数的值与上一次汇总后损失函数的值小于第三阈值时，判定训练结束。该第三阈值可以是系统检测到的阈值。
然而，本领域技术人员可以明确，该判断步骤并不限定于通过与差值与第三阈值的比较的方式进行，本领域还有多种验证训练是否可以结束的 方法，在此不再赘述。
S310，当判定训练结束，产出所述模型；
在这一步骤中，当判断出训练结束，即该模型已可以使用，则产出该模型，例如可以使该模型上线进行应用。
S311，当判定训练未结束，重复执行获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值的步骤。
在这一步骤中，例如，当检测到当前损失函数的值与上一次汇总后损失函数的值大于第三阈值时，判定训练未结束。在此情况下，可以重复执行获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值的步骤，即让每一个工作机继续基于其他数据片计算出本地权重和本地损失函数值，并在后续汇总获得当前权重和当前损失函数值，重复该训练方法，直到获得满足使用要求的模型。
另外，在第一至第三实施例公开的方案中，优先地，在将所述训练数据分配给多个工作机的步骤之前，所述方法还包括：
获取训练数据；
将所述训练数据分配给多个工作机的步骤还包括：
按照每一个工作机的标识，将所述训练数据分配给多个工作机。
另外，在第一至第三实施例公开的方案中，优先地，所述汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值的步骤可以包括：
将每一个工作机基于每一个数据片计算出的所述本地权重相加，获得当前权重；
将每一个工作机基于每一个数据片计算出的所述本地损失函数值相加，获得当前本地损失函数。
在本申请第三实施例提出的机器学习系统的训练方法中，通过使每一 个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
第四实施例
本申请第四实施例提出一种机器学习系统的训练系统，如图4是本申请第四实施例的机器学习系统的训练系统的方框图。该机器学习系统优选为分布式机器学习系统，如图4所示，该训练系统400包括：
训练数据分配模块401，用于将所述训练数据分配给多个工作机；
训练数据划分模块402，用于将每一个工作机分配得到的训练数据划分为多个数据片；
获取模块403，用于获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
汇总模块404，用于汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
异常检测模块405，用于利用当前权重和/或当前损失函数值执行模型异常检测；
第一处理模块406，用于当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
第二处理模块407，用于当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
在本申请第四实施例提出的机器学习系统的训练系统中，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
第五实施例
本申请第五实施例提出一种机器学习系统的训练系统，如图5是本申请第五实施例的机器学习系统的训练系统的方框图。该机器学习系统优选为分布式机器学习系统，如图5所示，该训练系统500包括：
训练数据分配模块501，用于将所述训练数据分配给多个工作机；
训练数据划分模块502，用于将每一个工作机分配得到的训练数据划分为多个数据片；
获取模块503，用于获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
汇总模块504，用于汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
异常检测模块505，用于利用当前权重和/或当前损失函数值执行模型异常检测；
第一处理模块506，用于当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
第二处理模块507，用于当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
在该实施例中，所述异常检测模块505包括：
第一判断子模块505a，用于判断当前损失函数值是否大于上一次汇总后的损失函数值的k倍；
第一判定子模块505b，用于当判断出当前损失函数值大于上一次汇总后的损失函数值的k倍时，判定所述模型异常检测的结果为第一类异常。
第二判断子模块505c，用于判断当前权重中的元素的最大值和/或最小值的绝对值是否大于预设的第二阈值；
第二判定子模块505d，用于当判断出当前权重中的元素的最大值和/或最小值的绝对值大于预设的第二阈值时，判定所述模型异常检测的结果为第二类异常。
所述异常检测模块505还包括：
第三判断子模块505e，用于判断前n次汇总后的损失函数值的平均值是否大于第一次汇总后的损失函数值；
第三判定子模块505f，用于当判断出前n次汇总后的损失函数值的平均值大于第一次汇总后的损失函数值时，判定所述模型异常检测的结果为第一类异常。
优选地，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第二阈值内的当前权重和/或当前损失函数值的步骤包括：
将所述最大值和/或最小值的绝对值大于预设的第二阈值的元素减去所述第二阈值，获得每一个元素均位于所述第二阈值内的当前权重。
在本申请第五实施例提出的机器学习系统的训练系统中，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
第六实施例
本申请第六实施例提出一种机器学习系统的训练系统，如图6所示为本申请第六实施例的机器学习系统的训练系统的方框图。该机器学习系统优选为分布式机器学习系统，如图6所示，该训练系统包括：
训练数据分配模块601，用于将所述训练数据分配给多个工作机；
训练数据划分模块602，用于将每一个工作机分配得到的训练数据划分为多个数据片；
获取模块603，用于获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
汇总模块604，用于汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
异常检测模块605，用于利用当前权重和/或当前损失函数值执行模型异常检测；
第一处理模块606，用于当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
第二处理模块607，用于当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
在本实施例中，所述系统还包括：
训练模块608，用于当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练。
在本实施例中，所述系统还包括：
结束判定模块609，用于判断训练是否结束；
模型产出模块610，用于当判定训练结束时，产出所述模型；
返回执行模块611，用于当判定训练未结束，重复执行获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值的步骤。
在本实施例中，所述结束判定模块用于当检测到当前损失函数的值与上一次汇总后损失函数的值小于第三阈值，判定训练结束。
在本实施例中，所述系统还可以包括：
训练数据获取模块，用于获取训练数据。所述训练数据分配模块用于按照每一个工作机的标识，将所述训练数据分配给多个工作机。
优选地，所述汇总模块包括：
权重累加模块，用于将每一个工作机基于每一个数据片计算出的所述本地权重相加，获得当前权重；
损失函数值累加模块，用于将每一个工作机基于每一个数据片计算出的所述本地损失函数值相加，获得当前本地损失函数。
在本申请第六实施例提出的机器学习系统的训练系统中，通过使每一个工作机一次只针对一个数据片进行处理、并根据该数据片计算得出的本地权重和本地损失函数值进行汇总，达到了每一次计算后能够准确地小幅度更新权重和损失函数值的效果，避免了现有技术的在线学习中容易受到数据中噪声影响的问题。同时，在获得当前权重和/或当前损失函数值之后执行异常检测，当检测出所获得的当前权重和/或当前损失函数值属于第一类异常或第二类异常，可以抛弃本次计算出的当前权重和/或当前损失函数值，而将上一次汇总得到的权重和损失函数值输入机器学习模型，或者对本次计算出的当前权重和/或当前损失函数值进行修正，将每次计算和汇总获得的权重和损失函数值的更新限制在有效范围内，避免了模型训练的不稳定性，增强了系统抗数据噪声的能力，提升了在线学习的机器学习系统的稳定性，确保了机器学习系统的正常产出，提高了上线后的机器学习系统的服务能力。
对于装置实施例而言，由于其与方法实施例基本相似，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见即可。
本领域内的技术人员应明白，本申请实施例的实施例可提供为方法、装置、或计算机程序产品。因此，本申请实施例可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请实施例可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
在一个典型的配置中，所述计算机设备包括一个或多个处理器(CPU)、输入/输出接口、网络接口和内存。内存可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/或非易失性内存等形式，如只读存储器(ROM)或闪存(flash RAM)。内存是计算机可读介质的示例。计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信号存储。信号可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信号。按照本文中的界定，计算机可读介质不包括非持续性的电脑可读媒体(transitory media)，如调制的数据信号和载波。
本申请实施例是参照根据本申请实施例的方法、终端设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理终端设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理终端设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理终端设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理终端设备上，使得在计算机或其他可编程终端设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程终端设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本申请实施例的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例做出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本申请实施例范围的所有变更和修改。
最后，还需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者终端设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者终端设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者终端设备中还存在另外的相同要素。
以上对本申请所提供的一种机器学习系统的训练方法和系统，进行了详细介绍，本文中应用了具体个例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想；同时，对于本领域的一般技术人员，依据本申请的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本申请的限制。

Claims (20)

  1. 一种机器学习系统的训练方法，利用训练数据对所述机器学习系统进行训练，其特征在于，所述训练方法包括：
    将所述训练数据分配给多个工作机；
    将每一个工作机分配得到的训练数据划分为多个数据片；
    获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
    汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
    利用当前权重和/或当前损失函数值执行模型异常检测；
    当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
    当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
  2. 如权利要求1所述的训练方法，其特征在于，所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤包括：
    判断当前损失函数值是否大于上一次汇总后的损失函数值的k倍；
    当判断出当前损失函数值大于上一次汇总后的损失函数值的k倍时，判定所述模型异常检测的结果为第一类异常。
  3. 如权利要求1所述的训练方法，其特征在于，所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤包括：
    判断当前权重中的元素的最大值和/或最小值的绝对值是否大于预设 的第二阈值；
    当判断出当前权重中的元素的最大值和/或最小值的绝对值大于预设的第二阈值时，判定所述模型异常检测的结果为第二类异常。
  4. 如权利要求3所述的训练方法，其特征在于，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值的步骤包括：
    将所述最大值和/或最小值的绝对值大于预设的第一阈值的元素减去所述第一阈值，获得每一个元素均位于所述第一阈值内的当前权重。
  5. 如权利要求1所述的训练方法，其特征在于，所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤包括：
    判断前n次汇总后的损失函数值的平均值是否大于第一次汇总后的损失函数值；
    当判断出前n次汇总后的损失函数值的平均值大于第一次汇总后的损失函数值时，判定所述模型异常检测的结果为第一类异常。
  6. 如权利要求1所述的训练方法，其特征在于，所述利用当前权重和/或当前损失函数值执行模型异常检测的步骤之后，所述方法还包括：
    当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练。
  7. 如权利要求6所述的训练方法，其特征在于，所述当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练的步骤之后，所述方法还包括：
    判断训练是否结束；
    当判定训练结束，产出所述模型；
    当判定训练未结束，返回执行获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值的步骤。
  8. 如权利要求6所述的训练方法，其特征在于，判断训练是否结束的步骤包括：
    当检测到当前损失函数的值与上一次汇总后损失函数的值小于第三阈值，判定训练结束。
  9. 如权利要求1所述的训练方法，其特征在于，在将所述训练数据分配给多个工作机的步骤之前，所述方法还包括：
    获取训练数据；
    将所述训练数据分配给多个工作机的步骤还包括：
    按照每一个工作机的标识，将所述训练数据分配给多个工作机。
  10. 如权利要求1所述的训练方法，其特征在于，所述汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值的步骤包括：
    将每一个工作机基于每一个数据片计算出的所述本地权重相加，获得当前权重；
    将每一个工作机基于每一个数据片计算出的所述本地损失函数值相加，获得当前本地损失函数。
  11. 一种机器学习系统的训练系统，利用训练数据对所述机器学习系统进行训练，其特征在于，所述训练系统包括：
    训练数据分配模块，用于将所述训练数据分配给多个工作机；
    训练数据划分模块，用于将每一个工作机分配得到的训练数据划分为多个数据片；
    获取模块，用于获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值；
    汇总模块，用于汇总每一个工作机基于每一个数据片计算出的所述本地权重和本地损失函数值，获得当前权重和当前损失函数值；
    异常检测模块，用于利用当前权重和/或当前损失函数值执行模型异常检测；
    第一处理模块，用于当所述模型异常检测的结果为第一类异常，将上一次汇总后的权重和上一次汇总后的损失函数值输入所述机器学习系统进行训练；
    第二处理模块，用于当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值，并输入所述机器学习系统进行训练。
  12. 如权利要求11所述的训练系统，其特征在于，所述异常检测模块包括：
    第一判断子模块，用于判断当前损失函数值是否大于上一次汇总后的损失函数值的k倍；
    第一判定子模块，用于当判断出当前损失函数值大于上一次汇总后的损失函数值的k倍时，判定所述模型异常检测的结果为第一类异常。
  13. 如权利要求11所述的训练系统，其特征在于，所述异常检测模块包括：
    第二判断子模块，用于判断当前权重中的元素的最大值和/或最小值的绝对值是否大于预设的第二阈值；
    第二判定子模块，用于当判断出当前权重中的元素的最大值和/或最小值的绝对值大于预设的第二阈值时，判定所述模型异常检测的结果为第二类异常。
  14. 如权利要求13所述的训练系统，其特征在于，当所述模型异常检测的结果为第二类异常，将当前权重和/或当前损失函数值修正为第一阈值内的当前权重和/或当前损失函数值的步骤包括：
    将所述最大值和/或最小值的绝对值大于预设的第二阈值的元素减去所述第一阈值，获得每一个元素均位于所述第一阈值内的当前权重。
  15. 如权利要求11所述的训练系统，其特征在于，所述异常检测模块包括：
    第三判断子模块，用于判断前n次汇总后的损失函数值的平均值是否大于第一次汇总后的损失函数值；
    第三判定子模块，用于当判断出前n次汇总后的损失函数值的平均值大于第一次汇总后的损失函数值时，判定所述模型异常检测的结果为第一类异常。
  16. 如权利要求11所述的训练系统，其特征在于，所述系统还包括：
    训练模块，用于当所述模型异常检测的结果为正常，将当前权重和/或当前损失函数值输入所述机器学习系统进行训练。
  17. 如权利要求16所述的训练系统，其特征在于，所述系统还包括：
    结束判定模块，用于判断训练是否结束；
    模型产出模块，用于当判定训练结束时，产出所述模型；
    返回执行模块，用于当判定训练未结束，重复执行获取每一个工作机基于每一个数据片计算出的本地权重和本地损失函数值的步骤。
  18. 如权利要求16所述的训练系统，其特征在于，所述结束判定模块用于：
    当检测到当前损失函数的值与上一次汇总后损失函数的值小于第三阈值，判定训练结束。
  19. 如权利要求11所述的训练系统，其特征在于，所述系统还包括：
    训练数据获取模块，用于获取训练数据；
    所述训练数据分配模块用于按照每一个工作机的标识，将所述训练数据分配给多个工作机。
  20. 如权利要求11所述的训练系统，其特征在于，所述汇总模块包括：
    权重累加模块，用于将每一个工作机基于每一个数据片计算出的所述本地权重相加，获得当前权重；
    损失函数值累加模块，用于将每一个工作机基于每一个数据片计算出的所述本地损失函数值相加，获得当前本地损失函数。
PCT/CN2017/074318 2016-03-04 2017-02-21 一种机器学习系统的训练方法和训练系统 Ceased WO2017148314A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2018546445A JP7004661B2 (ja) 2016-03-04 2017-02-21 機械学習システムのトレーニング方法及びトレーニングシステム
EP17759165.8A EP3425527A4 (en) 2016-03-04 2017-02-21 METHOD FOR TRAINING MACHINEER SYSTEMS AND TRAINING SYSTEM
US16/119,585 US11257005B2 (en) 2016-03-04 2018-08-31 Training method and training system for machine learning system

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610125904.6 2016-03-04
CN201610125904.6A CN107153630B (zh) 2016-03-04 2016-03-04 一种机器学习系统的训练方法和训练系统

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/119,585 Continuation US11257005B2 (en) 2016-03-04 2018-08-31 Training method and training system for machine learning system

Publications (1)

Publication Number Publication Date
WO2017148314A1 true WO2017148314A1 (zh) 2017-09-08

Family

ID=59743465

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/074318 Ceased WO2017148314A1 (zh) 2016-03-04 2017-02-21 一种机器学习系统的训练方法和训练系统

Country Status (6)

Country Link
US (1) US11257005B2 (zh)
EP (1) EP3425527A4 (zh)
JP (1) JP7004661B2 (zh)
CN (1) CN107153630B (zh)
TW (1) TWI740896B (zh)
WO (1) WO2017148314A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114282624A (zh) * 2021-12-29 2022-04-05 北京天融信网络安全技术有限公司 对象标注方法、装置、设备及介质
KR20230111056A (ko) * 2022-01-17 2023-07-25 한국전자통신연구원 시계열 데이터 기반 미래 상태와 신뢰도 예측 장치 및 방법

Families Citing this family (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754105B (zh) 2017-11-07 2024-01-05 华为技术有限公司 一种预测方法及终端、服务器
US11354301B2 (en) 2017-11-13 2022-06-07 LendingClub Bank, National Association Multi-system operation audit log
US10042879B1 (en) 2017-11-13 2018-08-07 Lendingclub Corporation Techniques for dynamically enriching and propagating a correlation context
CN107944990A (zh) * 2017-12-29 2018-04-20 山大地纬软件股份有限公司 一种基于机器学习的精准化推送的综合柜员装置及方法
US12299547B1 (en) * 2018-04-30 2025-05-13 Amazon Technologies, Inc. Edge filtering of machine learning training data
CN108875776B (zh) * 2018-05-02 2021-08-20 北京三快在线科技有限公司 模型训练方法和装置、业务推荐的方法和装置、电子设备
CN110689133B (zh) * 2018-06-20 2023-09-05 深信服科技股份有限公司 一种训练机器学习引擎的方法、系统及相关装置
CN109902167B (zh) * 2018-12-04 2020-09-01 阿里巴巴集团控股有限公司 嵌入结果的解释方法和装置
US12039458B2 (en) * 2019-01-10 2024-07-16 Visa International Service Association System, method, and computer program product for incorporating knowledge from more complex models in simpler models
CN111507483A (zh) * 2019-01-30 2020-08-07 鸿富锦精密电子(天津)有限公司 返修板检测装置、方法及计算机可读存储介质
CN110084380A (zh) * 2019-05-10 2019-08-02 深圳市网心科技有限公司 一种迭代训练方法、设备、系统及介质
CN110263294B (zh) * 2019-05-23 2020-08-04 阿里巴巴集团控股有限公司 损失函数取值的确定方法、装置和电子设备
US11755884B2 (en) 2019-08-20 2023-09-12 Micron Technology, Inc. Distributed machine learning with privacy protection
US11392796B2 (en) 2019-08-20 2022-07-19 Micron Technology, Inc. Feature dictionary for bandwidth enhancement
CN110837657B (zh) * 2019-11-06 2023-03-14 腾讯云计算（北京）有限责任公司 数据处理方法、客户端、服务器及存储介质
CN111159961B (zh) * 2019-12-13 2024-02-02 杭州广立微电子股份有限公司 一种基于曲线数据的异常检测方法及系统
WO2021213614A1 (en) * 2020-04-20 2021-10-28 Abb Schweiz Ag A fault state detection apparatus
CN114067379A (zh) * 2020-07-28 2022-02-18 北京君正集成电路股份有限公司 一种人脸角度检测模型训练中损失函数的设计方法
CN111881983B (zh) * 2020-07-30 2024-05-28 平安科技（深圳）有限公司 基于分类模型的数据处理方法、装置、电子设备及介质
CN112527631A (zh) * 2020-11-18 2021-03-19 武汉迈威通信股份有限公司 bug定位方法、系统、电子设备及存储介质
US11310563B1 (en) * 2021-01-07 2022-04-19 Dish Network L.L.C. Searching for and prioritizing audiovisual content using the viewer's age
CN113298127B (zh) * 2021-05-12 2024-08-06 深圳前海微众银行股份有限公司 训练异常检测模型的方法及电子设备
US12387085B2 (en) 2021-05-21 2025-08-12 Samsung Electronics Co., Ltd. Deep learning device and system including the same
CN113378914B (zh) * 2021-06-08 2023-06-30 上海壁仞智能科技有限公司 虚假关联性去除方法及装置
CN114021464A (zh) * 2021-11-09 2022-02-08 京东科技信息技术有限公司 数据处理方法、装置和存储介质
CN114970864A (zh) * 2022-04-29 2022-08-30 珠高智能科技(深圳)有限公司 模型更新方法、装置、电子设备及存储介质
CN115972214B (zh) * 2023-02-28 2025-04-04 广东电网有限责任公司 绑扎作业机器人控制方法、装置、机器人和存储介质
CN117668719B (zh) * 2023-11-14 2024-08-16 深圳大学 一种自适应阈值的隧道监测数据异常检测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070233679A1 (en) * 2006-04-03 2007-10-04 Microsoft Corporation Learning a document ranking function using query-level error measurements
CN102176698A (zh) * 2010-12-20 2011-09-07 北京邮电大学 一种基于迁移学习的用户异常行为检测方法
CN102737126A (zh) * 2012-06-19 2012-10-17 合肥工业大学 云计算环境下的分类规则挖掘方法
CN103853614A (zh) * 2012-12-06 2014-06-11 财团法人资讯工业策进会 工作负载分配装置及其工作负载分配方法
CN104008420A (zh) * 2014-05-26 2014-08-27 中国科学院信息工程研究所 一种基于自动编码机的分布式离群点检测方法及系统
WO2014160648A1 (en) * 2013-03-28 2014-10-02 Alibaba Group Holding Limited Ranking product search results

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001344590A (ja) * 2000-05-31 2001-12-14 Fuji Electric Co Ltd ニューラルネットワーク及びその学習方法、解析方法並びに異常判定方法
KR100640264B1 (ko) * 2002-03-02 2007-02-28 김용대 앙상블 모형을 이용한 데이터 마이닝 모형 구축 장치 및그 방법
US7467118B2 (en) * 2006-01-12 2008-12-16 Entelos Inc. Adjusted sparse linear programming method for classifying multi-dimensional biological data
JP2009288933A (ja) * 2008-05-28 2009-12-10 Sony Corp 学習装置、学習方法、及びプログラム
JP5206197B2 (ja) 2008-07-28 2013-06-12 富士通株式会社 規則学習方法、プログラム及び装置
JP5584914B2 (ja) * 2010-07-15 2014-09-10 株式会社日立製作所 分散計算システム
CN103425996B (zh) * 2013-08-01 2017-08-25 华南理工大学 一种并行分布式的大规模图像识别方法
US9953271B2 (en) 2013-11-22 2018-04-24 California Institute Of Technology Generation of weights in machine learning
CN105187411B (zh) * 2015-08-18 2018-09-14 福建省海峡信息技术有限公司 一种分布式异常检测网络数据流的方法
US11087234B2 (en) 2016-01-29 2021-08-10 Verizon Media Inc. Method and system for distributed deep machine learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070233679A1 (en) * 2006-04-03 2007-10-04 Microsoft Corporation Learning a document ranking function using query-level error measurements
CN102176698A (zh) * 2010-12-20 2011-09-07 北京邮电大学 一种基于迁移学习的用户异常行为检测方法
CN102737126A (zh) * 2012-06-19 2012-10-17 合肥工业大学 云计算环境下的分类规则挖掘方法
CN103853614A (zh) * 2012-12-06 2014-06-11 财团法人资讯工业策进会 工作负载分配装置及其工作负载分配方法
WO2014160648A1 (en) * 2013-03-28 2014-10-02 Alibaba Group Holding Limited Ranking product search results
CN104008420A (zh) * 2014-05-26 2014-08-27 中国科学院信息工程研究所 一种基于自动编码机的分布式离群点检测方法及系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3425527A4 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114282624A (zh) * 2021-12-29 2022-04-05 北京天融信网络安全技术有限公司 对象标注方法、装置、设备及介质
KR20230111056A (ko) * 2022-01-17 2023-07-25 한국전자통신연구원 시계열 데이터 기반 미래 상태와 신뢰도 예측 장치 및 방법
KR102875045B1 (ko) * 2022-01-17 2025-10-22 한국전자통신연구원 시계열 데이터 기반 미래 상태와 신뢰도 예측 장치 및 방법

Also Published As

Publication number Publication date
TW201734871A (zh) 2017-10-01
EP3425527A4 (en) 2019-11-27
EP3425527A1 (en) 2019-01-09
CN107153630A (zh) 2017-09-12
JP7004661B2 (ja) 2022-01-21
US20180365595A1 (en) 2018-12-20
JP2019507442A (ja) 2019-03-14
CN107153630B (zh) 2020-11-06
TWI740896B (zh) 2021-10-01
US11257005B2 (en) 2022-02-22

Similar Documents

Publication Publication Date Title
TWI740896B (zh) 機器學習系統的訓練方法和訓練系統
CN109558295B (zh) 一种性能指标异常检测方法及装置
JP6841910B2 (ja) ユーザー異常行動検出方法、装置及びシステム
US20170329856A1 (en) Method and device for selecting data content to be pushed to terminal, and non-transitory computer storage medium
CN107464132B (zh) 一种相似用户挖掘方法及装置，电子设备
EP3270316B1 (en) Method and system for user device identification
US20140149418A1 (en) Method and system for measuring social influence and receptivity of users
CN109561052B (zh) 网站异常流量的检测方法及装置
US20190197071A1 (en) System and method for evaluating nodes of funnel model
US8898808B1 (en) System and method for assessing effectiveness of online advertising
CA3062119A1 (en) Method and device for setting sample weight, and electronic apparatus
JP2019512126A (ja) 機械学習システムをトレーニングする方法及びシステム
CN107203883B (zh) 一种风险控制方法和设备
US20150332315A1 (en) Click Through Ratio Estimation Model
CN105469279A (zh) 一种商品的质量评价方法和装置
CN109977177A (zh) 生成用户画像的方法及装置
Mukhlasin Sistem Pendukung Keputusan Pemilihan Smartphone Menggunakan Metode Simple Additive Weighting (SAW) Berbasis Web
WO2016177280A1 (zh) 记录及还原网页中点击位置的方法和装置
CN110020025A (zh) 一种数据处理方法及装置
CN108366012A (zh) 一种社交关系建立方法、装置及电子设备
CN105898835B (zh) 生成无线接入点的接入点属性信息的方法与设备
JP2011227720A (ja) 推薦システム、推薦方法、及び推薦プログラム
CN110020332B (zh) 一种基于圈选元素的事件生成方法及装置
CN110633804A (zh) 金融产品进件评价方法、服务器和存储介质
CN104156457B (zh) 一种基于MapReduce框架的网页排序方法和系统

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2018546445

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2017759165

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2017759165

Country of ref document: EP

Effective date: 20181004

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17759165

Country of ref document: EP

Kind code of ref document: A1