Shortening Machine Learning Development Cycles With AutoML
If you have used Machine Learning algorithms, you must have experienced the difficulties of parameter tuning.
Join the DZone community and get the full member experience.Join For Free
If you have used Machine Learning algorithms, you must have experienced the difficulties of parameter tuning. Facing complex algorithm parameters, algorithm users always end up spending countless nights constantly trying. They may finally find a satisfying parameter combination after working overnight. However, is the found parameter combination really the best? No one knows.
During the Machine Learning link establishment, parameter tuning is not the only time-consuming and labor-consuming job. After an algorithm model is generated, developers also have to rack their brains to figure out how they deploy the model to a service to be called by terminals such as mobile phones and PCs. Sometimes, to connect such links, developers need to spend a whole night debugging the association between models of different formats and the server.
Artificial intelligence (AI) services bring convenience to people's lives. Will they also provide a user-friendly development environment for algorithm engineers? Shortening the development time is a common wish of algorithm engineers. Today, Platform of Artificial Intelligence (PAI) releases the automatic Machine Learning (AutoML) engine to solve problems in the Machine Learning process through Machine Learning methods.
Introduction to AutoML
What is PAI AutoML? AutoML means automation of the whole Machine Learning process. After Machine Learning data is uploaded, the Machine Learning process can be classified into three steps, that is, model training, evaluation, and deployment.
PAI Automatic Parameter Tuning
The PAI automatic parameter tuning feature has great value to both senior and new algorithm users.
- New algorithm users do not know the mathematical principle of each algorithm parameter during the algorithm computing process and have no idea about parameter tuning. Automatic parameter tuning helps them solve the problem quickly.
- Senior algorithm users have experience in parameter tuning. However, their experience is about parameter tuning guiding. For detailed parameters, they still need repeated attempts. For example, for a parameter with its value ranging from 0 to 100, senior algorithm users can determine the results when the parameter is set to 90 or 80 based on their experience. But in smaller granularity, for example, when the parameter is set to 81 or 82, they also need to manually test which is better for the result. Fortunately, the automatic parameter tuning feature can save the time spent in such repeated attempts.
Currently, the mainstream parameter tuning method in the industry is based on parallel search led by grid search and random search. The system repeatedly samples possible parameter combinations based on the random principle and tries to find the best parameter combination through repeated iterations. Each exploration is independent. One advantage is that algorithm users can explore the best solutions in a broader parameter space instead of a local space. The disadvantage is that each exploration is random, lacking an information accumulation process and wasting computing resources.
PAI provides Evolutionary Optimizer, a creative parameter tuning method, to ensure that each iteration of a model is automatically developed in an optimized parameter set range of the previous round, and the embedded efficient algorithm helps you find the most suitable parameter combination quickly, greatly reducing computing resource consumption and the number of parameter explorations. You just need to wait patiently for the miracle to come.
The following figure shows the parameter tuning iteration effect of Evolutionary Optimizer. You can clearly view effect improvement of each iteration.
Automatic Evaluation of PAI Models
PAI AutoML provides multi-dimensional algorithm evaluation methods. After you select required evaluation indicators from F1Score, Precision, Recall, and AUC, the system will automatically complete model evaluation and deliver services to the downstream training environment. All evaluation processes do not need manual intervention.
Model ordering table:
Model delivery configuration:
Quick Release of PAI Models
After a model is generated, you can quickly release the model to an API service on PAI. After you click Deploy, the system will list deployable models in the current experiment. You can select the required models and deploy them quickly.
After deployment is completed, the online service control platform automatically appears. You can manage all the models on the platform.
Example Use Case
Is PAI AutoML really helpful for users' businesses? Let's find it out through users' feedback after they use PAI on the Alibaba Cloud platform. Customer YZ STAR GAME focuses on mobile native and interactive video advertising and has been engaged in rewarded video advertising for over two years. With the growth of its businesses, the company is facing a growing challenge in intelligent advertising efficiency on multiple platforms, across multiple channels, and in multiple modes.
The technical director of YZ STAR GAME said that Alibaba Cloud PAI provides service capabilities featuring a low threshold and easy start. In this case, businesses can be quickly aligned with the big data-based Machine Learning platform, boosting the company's business development. Based on the PAI AutoML engine, the customer can quickly locate target users on different platforms and in different modes.
With the PAI AutoML engine, YZ STAR GAME has improved the model precision by 40% during parameter tuning, and expects over 10 million automatic deployments after all businesses are launched, which is expected to save human labor costs by 20% to 30%. The most important is that the time spent in establishing businesses on the Machine Learning service platform is shortened by at least half a year.
The PAI AutoML engine is designed to minimize the costs of Machine Learning business establishment. The current available model training parameter optimization and quick model deployment services have already helped save human labor costs. In the future, PAI will continue to invest in this part, with an aim to make Machine Learning and AI more accessible.
Published at DZone with permission of Leona Zhang. See the original article here.
Opinions expressed by DZone contributors are their own.