DeepSync - Oszillatorische Synchronisation als biologisch inspirierter Ansatz für tiefe neuronale Netze

01/12/2025 to 30/11/2028

Deep learning has revolutionized artificial intelligence, but most successful approaches rely on static,
feedforward neural network architectures. In contrast, brain activity exhibits rich temporal dynamics
distributed within and across brain areas. Transient oscillatory synchronization has been put forward
as a means of coupling information across space and time, leading to efficient and robust, yet flexible
and adaptive information processing.
DeepSync aims to integrate neural principles of oscillations into deep learning, developing neural
network models that leverage synchronization dynamics for sequential and multimodal tasks. We
propose to design and evaluate new AI architectures that exploit oscillatory couplings between
neurons to capture temporal dependencies, improve binding of features across modalities, and
enhance reasoning. We will benchmark these models on a variety of tasks. In particular, we will
evaluate their potential on challenging time-series tasks from neuroscience, e.g. the prediction and
classification of seizures in epilepsy patients. Finally, we will reverse-engineer trained oscillatory
networks using neural data analysis tools and mechanistic interpretability methods from AI research,
to investigate whether they develop topographic and functional specialization akin to brain networks.
By systematically bridging dynamical principles from neuroscience with machine learning, DeepSync
aims to provide new insights into both artificial intelligence and neural computation. DeepSync will
contribute to the development of AI systems that are more adaptable, interpretable, and aligned with
the computational strategies of the brain, and demonstrate their potential on applications in medical
neuroscience and beyond.