Shulou(Shulou.com)06/02 Report--

Background introduction

If you have ever used machine learning algorithms, you must experience the horrors of being dominated by algorithm parameters. In the face of complex algorithm parameters, algorithm users often have to spend endless night trying, just like looking for a needle in a haystack. Sometimes I work late at night and finally find a reliable combination of parameters, but is the combination of parameters really the best? God knows.

However, in the process of building machine learning links, it is often time-consuming and time-consuming to adjust parameters. It is not easy to generate the algorithm model, how to deploy the model as a service for mobile phones, PC and other terminals to call is also a difficult problem for developers. Sometimes, in order to get through such links, it takes all night to debug the associations between models and servers in different formats.

While artificial intelligence services facilitate human life, can they also provide a humanized development environment for the majority of algorithm engineers? It is the common wish of algorithm engineers to reduce dark circles under the eyes. PAI gave the answer to this question. Today, PAI released a full set of automated machine learning engine, which solves the problem of machine learning process by machine learning.

An overall introduction to AutoML

Let's take a look at what PAI-AutoML is. AutoML, as its name implies, automates the whole process of machine learning. The process after uploading machine learning data can be divided into three steps: model training, model evaluation and model deployment.

PAI automatic parameter adjustment

PAI automatic parameter adjustment feature is of great value to both experienced algorithm users and algorithm rookies:

For novice users: novice users do not know the mathematical principle of each algorithm parameter in the process of algorithm calculation, and are often confused about parameter adjustment, so automatic parameter adjustment can quickly help these users solve the problem.

For experienced users: experienced users often have some experience in parameter adjustment, but this experience can only guide the parameter adjustment work in the general direction, and some detailed parameters still need to be tried repeatedly. For example, for a parameter range of 0,100, experienced users can use experience to determine the result when the parameter is set to 90 or 80, but for smaller granularity, such as 81 or 82, which is better for the result, experienced users also need to experiment manually. The self-defined parameter adjustment function can replace this part of repetitive work.

At present, the mainstream idea of parameter adjustment in the industry is mainly based on Parallel Search, represented by grid search and random search. Through the random principle, the system constantly samples possible parameter combinations, and tries to find the optimal parameter group through non-stop iteration. Each exploration process is independent of each other. The advantage is that it is not easy to fall into the local optimal solution and can be explored in a broader parameter space. The disadvantage is that each exploration is random, lack of information accumulation process, and consume computing resources.

PAI provides the original Evolutionary Optimizer evolutionary parameter adjustment method, so that each iteration of the model is automatically developed in the last round of better parameter set interval, and the built-in efficient algorithm can quickly help you find the most appropriate parameter combination, greatly reducing the consumption of computing resources and the number of parameter exploration. All you have to do is make a pot of tea and wait for a miracle to come.

PAI evolutionary parameter adjustment iterative effect diagram, you can clearly see the improvement of each round of iteration:

Automatic evaluation of PAI model

PAI AutoML provides a multi-dimensional algorithm evaluation method. As long as you select the evaluation indicators you need in F1Score, Precision, Recall and AUC, the system will automatically complete the model evaluation and send the services to the downstream training environment. All evaluation processes do not require human participation.

Model sorting table:

Model distribution configuration:

One-click release of PAI model

The model is generated and can be published as an API service on the PAI platform with one click. As long as you click the deploy button, you will list the models that can be deployed in the current experiment, and you can complete the deployment with one click by selecting the desired model, which is not very simple.

After the deployment is completed, it will automatically jump to the online service management platform, where you can do all the model management related work.

Customer case

PAI-AutoML looks great, can it really help users' business? let's take a look at the feedback of PAI users after using it on Aliyun platform. First introduce customers: coconut Media is a company specializing in mobile native interactive video advertising. It has been deeply cultivated in the reward video industry for more than 2 years. With the growth of business scale, multi-platform and multi-channel, the problem of intelligent delivery efficiency in multi-mode is becoming more and more prominent.

The person in charge of coconut technology said: Ali PAI platform provides a low threshold, quick-start service capability, so that the business can be quickly docked to the machine learning platform based on big data, effectively promoting the rapid development of the company's business. Based on PAI AutoML engine, we can locate target users in different platforms and modes more quickly.

Coconut Media uses PAI AutoML engine to adjust parameters to help improve the accuracy of the model by 40%. Automated deployment is expected to achieve tens of millions of times after all businesses are online, saving manpower by 20% and 30%. The most important thing is to shorten the time for business to be built on machine learning service platform by at least half a year.

Architecture diagram:

Summary

The PAI AutoML engine aims to minimize the cost of building machine learning services. At present, the online model training parameter tuning and model one-click automatic deployment service have provided help in saving labor costs. In the future, PAI platform will continue to invest in this direction, so that machine learning is no longer a high threshold technology, so that artificial intelligence is within reach.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.