Exhibitors & Products
Events

Automated Machine Learning (AutoML) should make AI more accessible and also find wider application in industry. But where do we stand today? In this interview, Prof. Dr. Marius Lindauer from the University of Hannover provides insights into the current development, challenges and future prospects of AutoML.

What is AutoML and why is it so difficult to implement in industry?

AutoML promises to automate the development of AI models by optimizing design decisions such as the choice of network architecture, hyperparameters and learning methods. In practice, however, it turns out that AutoML cannot be used without prior knowledge.

One key problem is data quality: without sufficient high-quality data, AutoML cannot generate meaningful models. Furthermore, there is often a lack of understanding of the respective application and the economic framework conditions. This leads to misjudgements, as an example from industry shows: a company wanted to use AutoML for predictive maintenance, but only provided data from a single machine. When used on other machines, the model failed – a fundamental miscalculation that AutoML cannot correct.

AutoML: between euphoria and disillusionment

The hype surrounding AutoML began around 2018, when Google and other tech giants promised that it would enable everyone to apply machine learning. However, the reality is different: Although many companies such as Siemens, Bosch or Beckhoff use AutoML, they do not use it in the way it was originally promised. Instead of mechanical engineers working directly with AutoML, it remains a tool for data scientists and AI experts. The discrepancy between expectations and reality has led to disappointment in the industry.

Is a unified platform for AutoML missing?

Another obstacle to the widespread adoption of AutoML is the lack of a unified, easily accessible platform. While there are established frameworks in the AI community, such as PyTorch or TensorFlow, there is no comparable solution for AutoML. Open-source projects such as AutoSklearn or Optuna do exist, but they lack the necessary broad acceptance and support from industry.

Competition from large language models (LLMs)

Some see large language models such as GPT or Gemini as competition for AutoML, but Lindauer disagrees: LLMs are more of an addition than a replacement. They help with code generation, for example, and can also provide support within AutoML processes. In addition, many challenges, such as optimizing hyperparameters or choosing the model architecture, remain relevant.

New applications and future prospects for AutoML

Lindauer sees several potential fields of application for AutoML:

1. rapid prototyping: AutoML can help data science teams to arrive at valid AI prototypes more quickly.

2. Certification and accountability: AutoML systems could play a central role in the EU AI Act, as they can automatically generate documentation and evidence for AI applications.

3. Energy efficiency: AutoML can help to train AI models in a more resource-efficient way and make them more efficient.

The future: Human-Centered AutoML and sustainability

Lindauer's research focuses on three main areas:

  • Human-Centered AutoML: Greater user involvement in the AutoML process, instead of expecting a “one-click miracle”.
  • Green AutoML: Optimization of computing resources to develop energy-efficient AI models.
  • AutoML for Reinforcement Learning: Support for the development of autonomous agents and robotic systems.
  • Conclusion: AutoML needs an ecosystem

    Lindauer emphasizes that AutoML is more than just hype – but it needs a better infrastructure. Stronger collaboration between research and industry is essential to develop a viable platform. Companies interested in such an initiative are invited to network and actively shape the future of AutoML.