The No-code AI movement continues to grow, and quickly. In the past few days, we’ve got new announcements from AWS and Google with new tools that will make AI even more accessible for everyone. It is starting to become a very crowded space, with a lot of options and new innovations.
In this post, I would like to clarify a misconception that I see often which assumes that No-code AI is the same as AutoML. Both terms have boomed in the last few years and while there is a relationship, it is important to understand the differences that will help you choose the best tool for your projects.
What is No-code AI?
A few months ago we posted a blog about the definition of No-code AI. A simple way to put it is:
No-code AI is a way of building AI solutions without writing a line of code. It is a great way to test ideas quickly, build new projects, start businesses and new products faster.
I also identified different types of No-code AI tools, classified as follows:
- Drag and Drop or Flow Builder tools
- AutoML
- Pre-trained APIs
- Transfer Learning Tools
- AI App & Dashboard Builders
As you can see, AutoML is just one type of No-code AI tool available and more and more we see these types of tools being combined into a single solution. Automation, whether it is for model selection, data preparation, feature engineering, or hyper-parameter optimization can provide a great acceleration to your work.
What is AutoML?
To build a Machine Learning solution users need to follow a set of steps to create an ML pipeline. A simple AutoML definition is:
AutoML is the process of automating the tasks of applying Machine Learning to real-world problems, covering the complete pipeline from the raw dataset to deployable machine learning model.
Usually, there is a lot of manual work required to build a good pipeline, with a lot of difficult skills that users need to master to get good performing results. A lot of research and innovation went into automating some of those steps, making the technology available to non-experts, and democratizing the creation of AI Models.
At the end of the day...it is all about the DATA!
Machine Learning involves learning from Data, and Model Performance depends on the quality of that data. This is often summarized as “garbage in, garbage out”.
When you start a project, the top item is data access and feature engineering. You transform raw data into features that represent the underlying problem and algorithms learn from it. Having access to the right data, in the right format is critical to getting good results no matter which tool you are using.
Be careful using AutoML
Understanding the domain of your project and the data is one of the most critical aspects of any Machine Learning solution. That’s why Subject Matter experts play such a big role, and most of the time, they don’t have the Data Science Skills. AutoML tools, in theory, solve that problem because they allow anybody to build models without having the model development expertise. But don’t get confused, you still need the Model Interpretability skills, the Data skills, the feature skills, and if you are serious about putting your model into production, you also need Model Deployment skills.
Most Machine Learning use cases are too custom to get fully automated and SMEs are busy doing other tasks. Data understanding is usually done by Data Scientists having conversations and brainstorming sessions with multiple stakeholders. AutoML tools are very good at automating tasks but have a hard time understanding the business context.
Explainability plays also a big role for AutoML. For example, if you automate the Feature Transformation steps, it is very important to understand how those are done and which impact they will have. Those advanced capabilities are not always available in every tool.
The biggest danger with AutoML is that you will often get odd results and you won’t have a full path to iterate because the outputs of AutoML tools are not always actionable or your skills are limited.
Don’t get me wrong, I love AutoML, but more as a complementary tool for data scientists than a tool for non-experts. It provides a good baseline and can help to compare and validate model performance. It is also useful to introduce quick results to clients and stakeholders for fast prototyping. Nowadays, some tools also generate Python code that a data scientist can continue to tweak and convert into a Production Model.
To summarize:
- AutoML is just one type of No-code AI tool available, that automates parts of the workflow
- It is powerful but has to be used carefully and with good domain knowledge, especially on the data and its features
- It is perfect for quick prototyping or to get a baseline model that can be improved to a production model
- Explainability is one of the key capabilities to help understand the automation created by AutoML tools
- Enables to bring more people to the AI field and lower the skills barrier which will provide more innovation to the space
My favorite No-code AI tool would be the one that combines a flow interface, with full control on every step of the process and AutoML capabilities embedded in it as an option. Continue the discussion in the forum! What have you built with AutoML?