What is AutoML in Machine Learning? 🍅 How to Learn Machine Learning

What AutoML automates—and what you still need to decide

Automated machine learning (AutoML) searches parts of a modeling workflow, such as preprocessing, model selection, hyperparameters, or ensembles. Some tools provide a graphical interface; others are Python libraries.

Before a first run: define the target, remove information unavailable at prediction time, choose a split that matches future use, set an evaluation metric, and cap the training budget. Compare the result with a simple baseline.

A winning validation score is not a deployment decision. Inspect errors, inference cost, and performance on an untouched test set. Our model-evaluation guide covers the checks that automation does not remove.

Put the fundamentals into practice. Read the free Starter Kit excerpt on test sets to preview the €19 handbook. No email required.

What is AutoML?

Before explaining what AutoML is, lets see what a traditional end-to-end Machine Learning pipeline looks like:

As we see, this pipeline has many steps and requieres a deep understanding of the Machine Learning and Data Science world in order to be able to run through it correctly.

It requires analytical capabilities for exploring and preparing the data and building the models, and also software engineering capacities for deploying the model and putting it on a productive environment.

Not very many people have either of these, so the question is:

How can we use the awesome power of Machine Learning without having to be an expert on the field?

The answer to this question is AutoML.

AutoML stands for automated machine learning. It can help both specialists and less experienced users search modeling options within a defined budget and evaluation setup.

Some AutoML tools offer no-code interfaces; others require code. The interface does not determine whether the resulting experiment is valid.

Aside from being used by non Data Scientist, AutoML can also be used by individuals that already have knowledge of the field of Machine Learning to speed up their processes and make them more efficient.

AutoML platforms usually have the following characteristics:

Automated preprocessing and model or hyperparameter search, depending on the tool.
Graphical interfaces, Python APIs, or both.
Evaluation reports and, in some products, export or deployment options.
Limits on supported tasks, search space, compute budget, and customization.

The normal functionalities of an AutoML platform are the shown in the following diagram:

The diagram summarizes steps an AutoML system may automate. You still choose the target or dependent variable, verify the data, and define how the model will be evaluated; automation does not make those choices correct.

It is worth highlighting that an AutoML platform not only provides model or algorithm selection and hyper-parameters optimisation, returning only a predictive model. It goes further than this, allowing us to do feature selection, raw data analysis, and gives us an interface to make data-driven decisions.

Supported data types depend on the tool. AutoML is not limited to tabular data: for example, AutoGluon documents models combining tabular, text, and image features. Check task support and resource requirements before choosing a system.

Also, depending on the platforms, developers have more or less customisation capabilities. The least advanced platforms can be restrictive on what you can do, stopping individuals with enough skills to perform complex pipelines from doing so.

Who is AutoML for?

The beautiful thing about AutoML is that it can be used by a wide range of end-users: from people with little or null Machine Learning expertise to experts.

Citizen Data Scientists, business managers that are confortable with technology but have no ML expertise, and analyst can all use AutoML platforms to exploit the benefits of Machine Learning and Big Data.

A no-code interface can lower the implementation barrier, but users still need enough understanding of the data and evaluation to detect leakage, misleading metrics, and unsuitable predictions.

Machine Learning experts can also use AutoML to quickly implement and study models, however our feeling is that these kind of platforms, while avoiding some tedious scripting, don’t allow you to go very deep inside the guts of the algorithms, and many times have restrictions that for very specific or complex models might constraint what you can do versus if your were coding.

However, they do make finding optimal solutions easier by quickly trying a wide range of pipelines that would maybe not even occur to the data scientists, and they usually have a really neat and simple way of deploying models to production.

AutoML serves beginners and experienced practitioners. Its value is reducing repetitive search work while keeping the experiment and its limitations understandable.

Lets be honest, ML and AI seem like magic to people and tools like this will have a huge role in making them more human like.

Will AutoML replace Data Scientists?

One question that we could ask ourself is:

Does AutoML do a better job than human experts?

There is no universal winner between AutoML and a manually developed model. Compare both on the same data splits, metric, and resource budget. AutoML can produce strong candidates while humans remain responsible for the problem definition and evaluation.

The useful question is how much time AutoML saves at an acceptable quality and cost. Include search runtime, inference latency, debugging, and maintenance in that comparison.

Awesome, lets see some examples AutoML platforms and finish off with where we think this is heading and some further resources.

Automated Machine Learning platforms

Examples illustrate different interfaces and capabilities; this is not a ranking. Check each project’s current installation requirements and supported tasks:

auto-sklearn: a separate AutoML toolkit with a scikit-learn-style interface, not a built-in scikit-learn module.
AutoGluon: documents workflows using tabular, text, and image inputs.

There are many more both, paid and open source, so feel free to do your own research find the one that suits you best. Lets end with where we think the future of AutoML is going.

The future of AutoML

AutoML will not replace, but boost productivity and make the work of data scientists much more interesting and challenging. Also, domain knowledge of a specific sector will become more important: knowing what data to use and when, and coming up with automated AI solutions.

As more implementation steps become automated, reproducible experiments, data-quality checks, and clear documentation remain useful skills. The balance between manual code and automation depends on the project.

However, what you will always need is people that are qualified enough to choose the right data, do the right analysis and make the right decisions.

Evaluate an AutoML system through a small reproducible pilot: record the dataset version, split strategy, metric, search budget, and final test result. Avoid choosing a tool solely from a headline benchmark.

Summary and further resources

That is it, we hope that you have learnt what AutoML is all about, and that we have lit a small spark in your mind so that you will research this technology by yourself. Some further resources are:

arxiv.org paper: Taking the Human Out of Learning applications: A Survey on Automated Machine Learning.
The Hundred Page Machine Learning Book is the perfect companion for an AutoML platform – short, concise, with little code, it is perfect to have by your side when using one of these tools.
Other books like Python Machine Learning will take you a step further and use these platforms to the maximum of their capabilities.

Lastly, this awesome Statquest video on the topic