“We are making deep learning more robust”

Dr. Arens, AI processes, for instance those used to classify images, are very efficient. In spite of this, they are only hesitantly making their way into practical applications. Why is this?

Michael Arens: Despite their renowned performance, deep learning processes also continuously attract attention when they make serious mistakes: AI processes are often overwhelmed by specifically selected input data.
This is why a key issue for us is robustness: How can we make sure that AI will deliver the desired outcome for any situation and that it will not, for example, think that a picture of a butterfly is actually a plane because of a certain
cloud texture in the background? Something like this could happen as a result of a paradigm change borne by machine learning: Rather than on algorithms, results depend on training data. When the AI system is learning, millions of parameters are set based on the training data, resulting in a type of black box that will determine – in complex and barely comprehensible ways – which outputs are caused by specific inputs at a later date.

So, can the key to robustness be found in the training data?

Arens: Exactly. The training data must describe the problem comprehensively. Ideally, they cover the whole range of what might be input at a later date. There are, however, other aspects which play a defining part in our work. Such
as: Can we find weak spots and use them to outwit the AI, for example to avoid detection? We are covering this in a project on camouflage and camouflage assessment (see next page). Or, another important aspect, what will happen when the input data systematically shift because the sensor gets old, for example – will the process still be robust? To address this kind of question and systematically make the process more resilient, we are also investigating various tools that allow for a look into the black box.

What approaches could ensure that there is sufficient training data?

Arens: In our field of business, a general lack of training data poses a challenge. This data doesn’t accumulate on a daily basis to the same degree that it does for most other fields of application. This is especially the case for multispectral data, which plays an important role in defense research. Data produced by simulations may help. So may artificial variations of real data. A third approach is to transform existing data. In such a way, infrared data can be generated from conventional images acquired in the visible spectrum, though an appropriate AI process is required for this. We are exploring and developing tools for all of these approaches.

Computer scientist Dr. Michael Arens is the spokesperson for the Defense business unit and head of the Object Recognition department (OBJ).