A training method and training system for a machine learning system

Abstract

The present application discloses a training method and system for a machine learning system, using training data to train the machine learning system. The training method includes: distributing the training data to multiple working machines; Divide the training data into multiple data slices; obtain the local weight and local loss function value calculated by each worker based on each data slice; summarize these local weights and local loss function values ​​to obtain the current weight and current loss function value; use The current weight and / or current loss function value performs model anomaly detection; when the detection result is the first type of anomaly, the weight after the last summary and the loss function value after the last summary are input into the machine learning system for training; when the detection The result is the second type of abnormality, the current weight and / or current loss function value is corrected to the current weight and / or current loss function value within the first threshold, and input to the machine learning system for training.

Description

technical field [0001] 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. Background technique [0002] Nowadays, serving users well is the goal that every Internet company hopes to achieve. To this end, most companies use machine learning to capture user preferences and habits and provide personalized services. For example, various websites can use machine learning systems to collect user behavior data online, provide different search results for users of different genders / ages, and provide services based on user preferences to the greatest extent. [0003] The above-mentioned machine learning system can collect a large amount of user behavior data, such as user browsing / clicking / purchasing / upvoting / posting opinions on posts, etc., and use certain machine learning algorithms to train the machine learning system offline. After obtaining a predictive model, put it o...

AI Technical Summary

Problems solved by technology

The accuracy of online learning is relatively high, but due to the short period of data collection, the model is greatly affected by the noise in the data, resulting in unstable model training; at the same time, because online learning uses asynchronous model update, training data continues to flow into the model, and the model continues to For update learning, model training is often stuck due to some uncontrollable problems, and training failures occur, resulting in the inability to produce usable models, affecting online services, and damaging user experience

Images