ScaDS.AI Dresden/Leipzig and Leipzig University are pleased to welcome Prof. Alexander Binder as the new  Professor for Multimodal Machine Learning and Principal Investigator. In this position, he will research the explainability of predictions made by multimodal AI models.

On the trail of AI thinking: explainability of multimodal models

Multimodal AI models offer new possibilities for numerous applications, including medical diagnostics, virtual assistants and autonomous vehicles. These models integrate a variety of data types, such as speech, text, audio, images and information from external sources. One can think of household robots offering assistance, agents identifying inclarities in reports. But also of drones in agriculture which can drive away wild birds from plantations, or detect invasive insect species and their nests.

At the same time, the decision-making processes of multimodal models are often difficult to interpret due to their non-linear accumulation of information and data compression. A known consequence are adversarial attacks. Also, this allows for attacks on their reasoning for example with poisoned training data. Occasionally models learn shortcuts to decision making due to non-causal correlations in the data. The well known hallucinations are an inevitable consequence of similarity-based representations and expected loss minimization. Building on these traits and his previous successful research, Professor Alexander Binder will continue exploring ways to obtain insights from explanation for deep neural networks.

Stages of the academic career

Prof. Binder studied mathematics at the Humboldt University in Berlin. He received his Ph.D. in Computer Science from the Technical University of Berlin in 2013. Under the supervision of Prof. Klaus-Robert Müller, he worked on explanations of predictions from bag-of-words models and deep neural networks. This collaboration resulted in the development of Layer-wise Relevance Propagation (LRP), a method primarily used for explaining decisions made by neural networks (XAI). Following research and teaching position in Oslo, he held an assistant professorship at the Singapore University of Technology and Design (SUTD) from 2015 to 2020. This was followed by an associate professorship at the Singapore Institute of Technology from 2021.

“My assistant professorship at the SUTD in Singapore was very important for my career. I wrote numerous fundamental papers there on  LRP explanations and outlier detection. SUTD was a newly founded university at the time, and I was also able to participate in its university-wide committees.”

Prof. Binder as new AI Professor for Multimodal Machine Learning at Leipzig University

As the new Professor of Multimodal Machine Learning, Prof. Alexander Binder is tackling the problem of explaining multimodal models. He is building his own research team and contributing to the curriculum by designing a new lecture on his area of research and for general topics in deep learning.

“I strive to contribute insights to readers of my research which might be to some extent unexpected yet relevant. In the past XAI was able to repeatedly discover unexpected effects even in smaller models. If there is a shortcut which humans cannot sense, a deep learning model will discover it. Large DL models contain the potential to replicate all kinds of human behaviour because of the training paradigm employed and their memory capacity. There is more to be found out.”