Termin

28.08.2025

RealmNN Neuro-AI project

RealmNN, collaboration with Georg Martius was funded by BMFTR

Our project brings together computational neuroscience and machine learning to investigate a fundamental question: how complex should the basic computational units of artificial neural networks be? While modern deep learning relies on simple, memoryless units, biological neurons perform far richer computations—integrating information across multiple timescales and processing inputs through elaborate dendritic structures. RealmNN explores whether incorporating these biologically inspired mechanisms into artificial networks can lead to AI systems that are more efficient, more robust, and better suited to tasks requiring long-range temporal reasoning—such as language understanding, audio processing, and decision-making in embodied agents. Building on the recently introduced Expressive Leaky Memory (ELM) neuron, which faithfully reproduces the dynamics of cortical neurons, the project pursues three interconnected directions: characterizing the computational repertoire of diverse cortical cell types, designing recurrent architectures around expressive units, and applying these networks as policy and world models in reinforcement learning. Potential applications range from more capable language and audio models to embodied agents that learn and act over extended time horizons with greater efficiency.