When AI cannot explain its own steps

The centipede was doing perfectly well until the frog asked how. Which leg moves first? Which one follows? The question, posed as a joke, unsettles the animal so deeply that it can no longer walk. What had seemed effortless suddenly becomes mysterious.

Dr. Linara Adilova begins her science communication essay Generalization in Deep Learning: From Theory to Practice with this old dilemma. Written as a submission to the KlarText Prize for Science Communication, the text opens a door into the doctoral research she completed at Ruhr University Bochum - research that now informs her work as a Postdoctoral Researcher at TU Dortmund University and the Research Center Trustworthy Data Science and Security (RC Trust): why do deep neural networks work so well, and what do we actually understand about the way they do it?

Modern AI systems already “walk” with remarkable confidence. They classify images, generate fluent text, support medical assessments, and detect patterns in complex data. From the outside, their performance can look almost natural. But Linara’s essay reminds us that apparent competence is not the same as understanding. She connects the centipede’s dilemma with John Searle’s Chinese Room, the famous thought experiment in which a system appears to understand Chinese while merely following very detailed rulesbook for answering any possible question from the inside.

For small everyday tasks, that difference may seem harmless. In high-risk contexts, it is not. When AI systems enter medicine, finance, autonomous mobility, or other sensitive areas, successful output alone is too thin a basis for trust. We need to know how such systems reach their conclusions, when they fail, and what can justify confidence in their behavior.

This is the central question behind Linara Adilova’s dissertation, “Generalization in Deep Learning: From Theory to Practice.” Her research investigates how deep neural networks generalize from training data to new, unseen examples. Instead of accepting neural networks as powerful but opaque machines, she asks how their behavior can be described with mathematical precision.

To make this challenge tangible, her essay turns to bridge building. Long before structural physics, bridges were built through experience, imitation, and incremental improvement. Some endured for centuries. Others failed. Only mathematical theory made it possible to calculate forces, model risks, and design structures that once seemed impossible.

Deep learning, Linara suggests, still resembles that earlier phase. Progress often comes from larger models, more data, and more computing power. But trustworthy AI requires more than scaling. It needs theoretical foundations.

Her work focuses on two such foundations: high-dimensional geometry and information theory. Geometry helps describe neural networks as objects in spaces with many parameters, offering insight into stability and robustness. Information theory, first developed to understand communication channels, helps analyze how information moves from input to output inside a model.

The strength of Linara’s essay lies in making this abstract problem visible. It shows why trustworthy AI is not only a question of better results. It is also a question of whether we understand the rules defining the “walking”.