AutoML is a new way in which the process of building a machine learning model is automated or mechanized. At first glance, it may seem that AutoML is supposed to replace the scientist and fire him/her. But the issue is not so simple and so tragic. In this article, we first answer the question “What is AutoML” in simple terms.
What is machine learning?
Before focusing on the question “What is AutoML”, it must first be clarified what is machine learning trying to achieve? To answer this question, it is best to first provide a definition of machine learning:
Machine learning is one of the applications of artificial intelligence (AI) that allows automatic learning and advancement through the experience without clear and precise programming for a system. Machine learning focuses on the development of computer programs that can access data and use it to learn on their own. The process of building a machine learning model. In this section of “What is AutoML”, the process of building a machine learning model is listed. This is to provide the necessary background to address the question of exactly what is to be automated in AutoML.
- Finding a Problem in Business (Business Problem)
- Translate a business issue into a “data science problem form”
- Find the necessary data set
- Defining the purpose and criteria for estimating.
- Explore the dataset
- Build and teach the model by doing the steps
- Model deployment
- Model evaluation (if there is no necessary consent, we can return to the before step)
- Model commercialization
- Use and application of the model
AutoML’s title for Automated Machine Learning means automatic machine learning. AutoML is an automation operation of the machine learning process with the aim of simplifying and speeding up its tasks. The question may be, what is to be mechanized and automated?
Automated machine learning is an emerging technology that still faces various challenges in its implementation. The following is an overview of “What is AutoML” and the main challenges of automatic machine learning.
What are the challenges of AutoML?
As stated, the purpose of using AutoML is to simplify and expedite the machine learning modeling process. However, automatic machine learning also poses new challenges. The first challenge is: “How to ensure the result of AutoML?” For many automated learning users, the question is, “Is this really the best model possible?”
Therefore, many AutoML technologies and solutions allow users to manipulate some of the super parameters. This way, users can create several different models and compare these models. Although it is understandable to manipulate super parameters, it does eliminate the advantages of automatic machine learning, which are simplicity and speed.
The second problem is that the usual AutoML technology will still depend on technical knowledge. In other words, AutoML often needs a data scientist to examine the constructed model and find the best model or models to implement. If there are several well-functioning models in the evaluation phase, the data scientist may use other technologies to build the final model for commercialization. This also eliminates the advantages of AutoML in simplifying and speeding up the machine learning process. One of the proposed solutions to address these challenges is the implementation of AutoML at various levels, which is discussed in the following article “What is AutoML”.
Different levels of AutoML
Some experts are now categorizing AutoML into different levels. Perhaps by adjusting the amount of AutoML usage depending on the type and features of a machine learning project, it is possible to take advantage of it simultaneously and meet its challenges as much as possible.
- Level zero: no automation; The application encrypts its machine learning algorithms in C ++ from scratch.
- Level 1: Using APIs and high-level algorithms available in technologies such as Sklearn, Keras, Pandas, H2O, XGBoost, and others
- Level two: Adjusting and assembling cloud parameters and selecting the base model automatically