Training method and training system for machine learning system

1. A method for a machine learning system, comprising:

allocating training data to a plurality of working machines;

obtaining a weight and a loss function value locally calculated by each working machine based on a subset of a respective data piece of the training data allocated to the respective working machine;

aggregating the weight and the loss function value locally calculated by each working machine to obtain a current weight and a current loss function value;

performing model abnormality detection using the current weight and/or the current loss function value;

inputting a weight and a loss function value of a previous aggregation to the machine learning system for training in response to a result of the model abnormality detection being a first type of abnormality; and

modifying the current weight and/or the current loss function value to a current weight and/or a current loss function value within a first threshold in response to the result of the model abnormality detection being a second type of abnormality, and inputting thereof to the machine learning system for training.

2. The method of claim 1 , wherein performing the model abnormality detection using the current weight and/or the current loss function value comprises:

determining whether the current loss function value is greater than k times of a loss function value of a previous aggregation; and

determining that a result of the model abnormality detection is a first type of abnormality in response to determining whether the current loss function value is greater than k times of the loss function value of the previous aggregation.

3. The method of claim 1 , wherein performing the model abnormality detection using the current weight and/or the current loss function value comprises:

determining whether an absolute value of a maximum value and/or a minimum value of elements in the current weight is greater than a preset second threshold; and

determining that the result of the model abnormality detection is a second type of abnormality in response to determining whether the absolute value of the maximum value and/or the minimum value of the elements in the current weight is greater than the preset second threshold.

4. The method of claim 3 , wherein modifying 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 in response to the result of the model abnormality detection being the second type of abnormality comprises subtracting elements having the absolute value of the maximum value and/or the minimum value being greater than the first threshold by the first threshold to obtain a current weight in which each element is located within the first threshold.

5. The method of claim 1 , wherein performing the model abnormality detection using the current weight and/or the current loss function value comprises:

determining whether an average value of loss function values of first n aggregations is greater than a loss function value of a first aggregation; and

determining that the result of the model abnormality detection is the first type of abnormality in response to determining that the average value of the loss function values of the first n aggregations is larger than the loss function value of the first aggregation.

6. The method of claim 1 , wherein: after performing the model abnormality detection using the current weight and/or the current loss function value, the method further comprises inputting the current weight and/or the current loss function value into the machine learning system for training in response to a result of the model abnormality detection being normal.

7. The method of claim 6 , wherein: after inputting the current weight and/or the current loss function value into the machine learning system for training in response to the result of the model abnormality detection being normal, the method further comprises:

determining whether the training is ended;

generating a model in response to determining that the training is ended; and

repeating the obtaining of the weight and the loss function value locally calculated by each working machine again in response to determining that the training is not ended.

8. The method of claim 6 , wherein determining whether the training is ended comprises determining that the training is ended in response to detecting that the current loss function value and the loss function value of the previous aggregation are smaller than a third threshold.

9. The method of claim 1 , wherein: before allocating the training data to the plurality of working machines, the method further comprises:

obtaining the training data; and

allocating the training data to the plurality of working machines in accordance with an identification of each working machine.

10. The method of claim 1 , wherein aggregating the weight and the loss function value locally calculated by each working machine to obtain the current weight and the current loss function value comprises:

summing weights locally calculated by each working machine to obtain the current weight; and

summing loss function values locally calculated by each working machine to obtain the current local loss function.

11. A system comprising:

one or more processors;

memory;

a training data allocation module stored in the memory and executable by the one or more processors to allocate the training data to a plurality of working machines;

an acquisition module stored in the memory and executable by the one or more processors to obtain a weight and a loss function value locally calculated by each working machine based on a subset of a respective data piece of the training data allocated to the respective working machine;

an aggregation module stored in the memory and executable by the one or more processors to aggregate the weight and the loss function value locally calculated by each working machine to obtain a current weight and a current loss function value;

an anomaly detection module stored in the memory and executable by the one or more processors to perform model abnormality detection using the current weight and/or the current loss function value;

a first processing module stored in the memory and executable by the one or more processors to input a weight and a loss function value of a previous aggregation to the system for training in response to a result of the model abnormality detection being a first type of abnormality; and

a second processing module stored in the memory and executable by the one or more processors to modify the current weight and/or the current loss function value to a current weight and/or a current loss function value within a first threshold, and input thereof to the machine learning system for training, in response to the result of the model abnormality detection being a second type of abnormality.

12. The system of claim 11 , wherein the anomaly detection module comprises:

a first determination sub-module configured to determine whether the current loss function value is greater than k times of a loss function value of a previous aggregation; and

a first judgment sub-module configured to determine that a result of the model abnormality detection is a first type of abnormality in response to determining whether the current loss function value is greater than k times of the loss function value of the previous aggregation.

13. The system of claim 11 , wherein the anomaly detection module comprises:

a second determination sub-module configured to determine whether an absolute value of a maximum value and/or a minimum value of elements in the current weight is greater than a preset second threshold; and

a second judgment sub-module configured to determine that the result of the model abnormality detection is a second type of abnormality in response to determining whether the absolute value of the maximum value and/or the minimum value of the elements in the current weight is greater than the preset second threshold.

14. The system of claim 13 , wherein: when the result of the model abnormality detection is the second type of abnormality, modifying 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 comprises subtracting elements having the absolute value of the maximum value and/or the minimum value being greater than a preset second threshold by the first threshold to obtain a current weight in which each element is located within the first threshold.

15. The system of claim 11 , wherein the anomaly detection module comprises:

a third determination sub-module configured to determine whether an average value of loss function values of first n aggregations is greater than a loss function value of a first aggregation; and

a third judgment sub-module configured to determine that the result of the model abnormality detection is the first type of abnormality in response to determining that the average value of the loss function values of the first n aggregations is larger than the loss function value of the first aggregation.

16. The system of claim 11 , further comprising a training module configured to input the current weight and/or the current loss function value into the machine learning system for training in response to the result of the model abnormality detection being normal.

17. The system of claim 16 , further comprising:

an ending determination module configured to determine whether the training is ended;

a model production module configured to generate a model in response to determining that the training is ended; and

a return execution module configured to perform obtaining the weight and the loss function value locally calculated by each working machine again in response to determining that the training is not ended.

18. The system of claim 17 , wherein the ending determination module is configured to determine that the training is ended in response to detecting that the current loss function value and the loss function value of the previous aggregation are smaller than a third threshold.

19. The system of claim 11 , further comprising a training data acquisition module configured to obtain the training data, the training data allocation module being configured to allocate the training data to the plurality of working machines according to an identification of each working machine.

20. One or more computer readable media storing executable instructions that, when executed by one or more processors of a system, cause the one or more processors to perform acts comprising:

allocating training data to a plurality of working machines;

obtaining a weight and a loss function value locally calculated by each working machine based on a subset of a respective data piece of the training data allocated to the respective working machine;

aggregating the weight and the loss function value locally calculated by each working machine to obtain a current weight and a current loss function value;

performing model abnormality detection using the current weight and/or the current loss function value;

inputting a weight and a loss function value of a previous aggregation to the system for training in response to a result of the model abnormality detection being a first type of abnormality; and

modifying the current weight and/or the current loss function value to a current weight and/or a current loss function value within a first threshold in response to the result of the model abnormality detection being a second type of abnormality, and inputting thereof to the machine learning system for training.