AutoML

Building an AI model normally means: choosing an approach, trying out dozens of settings, comparing results, and starting over again. This costs experts many hours. AutoML – short for Automated Machine Learning – takes over exactly this process. You feed in data and get a finished model out. The software decides on its own which approach fits best and how it should be configured.

The bottleneck AutoML removes

Building models requires a specific profession: the data scientist, someone who combines statistics, programming, and domain knowledge. Such experts are scarce and expensive worldwide. For many companies, AI was therefore simply unaffordable.

AutoML lowers this hurdle considerably. A marketing department can use it to train a model that predicts which customers will churn – without writing a single line of code. This shifts AI from the lab into the everyday operations of ordinary companies. At the same time, experts benefit too: instead of hours of trial and error, they can use AutoML to quickly find a good starting point, which they can then refine in a targeted way.

What AutoML tries out behind the scenes

At the core of AutoML is a systematic search. The system tests many combinations: different approaches, different preprocessing steps for the data, different values for so-called hyperparameters – settings that are fixed before training, such as how fast a model learns or how complex it is allowed to become.

Early AutoML systems simply tried out every combination, which took a very long time. More modern approaches are smarter: they remember which combinations have performed well so far and search specifically in their vicinity. This method is called Bayesian optimization. It works similarly to an experienced cook who, after a failed attempt at a recipe, doesn’t randomly try something else but deliberately targets the suspected weak point.

For image data, there is also Neural Architecture Search: the software automatically designs and tests various neural network architectures – that is, their fundamental structure – until it finds a high-performing variant. This used to be pure research work and took weeks.

AutoML in products and headlines

Google, Microsoft, and Amazon offer AutoML as a cloud service. Google Cloud AutoML lets you train an image recognition model simply by uploading and labeling photos – the rest happens automatically. Microsoft’s Azure Machine Learning and Amazon’s SageMaker Autopilot work in a similar way.

In financial news, AutoML often comes up in connection with the “democratization of AI” – the promise that AI should not be usable only by tech giants. Critics point out that while AutoML delivers a good model, it provides no understanding of why it is good. Anyone who doesn’t understand a model has a hard time recognizing its errors. This is a particular problem in regulated fields such as medicine or lending, where decisions must be explainable.

For beginners and small teams, AutoML is nevertheless a genuine advance. It makes AI experiments faster and cheaper – creating time for what no algorithm can take over: asking the right questions.