[Service Development Team, Hyun-Jun Han]
What is AutoML
AutoML literally means Auto + ML, the process of automating the task of developing machine learning models. By automating the process that took a lot of time in research and development of the model, it not only reduces the required manpower, cost, and time, but also helps the overall process to find the optimal performance. It is mainly used as a concept for modeling and automating using separate artificial intelligence to create artificial intelligence.
AutoML is complementing many parts of ML research through automation, and related research is steadily progressing. Here are some of the benefits you can expect from AutoML.
Save time and money
There are many processes that go through research/development of ML models and actual service. It would be good if only the data and purpose were informed, and the model would be modeled on its own, and the model would solve all the problems. It takes a lot of effort and time by the experts. In addition, a lot of repetition and trial and error are inevitably encountered in this process, which further increases the cost. But AutoML helps save time and money by automating parts of HPO, NAS, etc.
ML implementation possible without extensive expertise
Developing ML models requires many experts, including data scientists, AI engineers, and domain experts. However, the reality is that the supply of experts is still insufficient compared to the demand, and the smaller the company, the more difficult it is to find AI-related experts.
AutoML can create a model by repeating the analysis and correction necessary for development on its own without the help of experts even in this asymmetric situation of supply and demand. There is.
AutoML-Zero, released by Google, is designed to effectively develop high-efficiency models by minimizing the intervention of model researchers. It's based on basic high school-level math concepts and uses an algorithm that evolves automatically. AutoML-Zero generates candidate algorithms by randomly combining mathematical operations and selects the best algorithm from among the candidate algorithms with a simple test. And a part of the code of the algorithm is added, deleted, and replaced to create a mutated version, and the redundant algorithm is deleted. If the generated algorithm has good performance, it is saved, otherwise it is removed. Algorithms found in this way were able to design models with better performance than human intervention algorithms.
Currently, AutoML is replacing courses in the overall AI field. In the image field, by applying AutoML, a model with less data and better performance like EfficientNet has emerged, and companies are increasingly applying AutoML. In the future, as AutoML develops, companies will become more accessible, and AI will penetrate more deeply into our lives for various services. Attention is paid to what level it will reach in the future in the ever-changing AI field.