RealmNN

Recurrent Expressive Adaptive Leaky Memory Neural Network: Harnessing Cortical Neuron Insights for Next-Generation AI Architectures.

01/12/2025 to 30/11/2028

Modern artificial intelligence mostly relies on very simple computing units that add up their inputs and then
apply a nonlinear transformation. In contrast, real brain cells behave in a much more complex way: they
process signals on multiple levels and over different time scales. This raises the question of whether
artificial intelligence could become more powerful, efficient, and robust if it mimics these biological
properties.
In this project, we aim to develop networks made of particularly expressive artificial neurons. Such networks
could solve difficult tasks more effectively, especially those that require linking information over long periods
of time, such as language processing or decision-making. We have already developed a new type of
artificial neuron that closely reproduces the behavior of brain cells and performs well on tasks with long time
dependencies.
Building on this, we will first study the properties of real brain cells and how these can be transferred to
artificial neurons. Next, we will integrate these neurons into larger networks and explore how the
composition and structure of these networks affect their performance. Finally, we will apply these networks
in scenarios where robots or virtual agents need to learn and act over long periods of time.
With this approach, we aim to uncover fundamental insights into how neurons process information and to
lay the foundation for a new generation of artificial intelligence that is more closely inspired by the
functioning of the brain.