Bio-inspired neural networks for AI applications

There has been an enormous development in artificial intelligence (AI) recently. However, while AI supersedes human skills in some fields, the brain is still far superior in many areas - it is more robust, it is more energy efficient, it needs fewer examples to learn, it can learn complex tasks easier, and it can transfer knowledge from one task to another. Recently, there has been significant progress in reinforcement learning and deep learning, which in principle are inspired by neurological processes, but in practice there are large differences between the way artificial and biological neural networks work. Our hypothesis is that artificial neural networks based on the structure, dynamics and learning rules in the brain will be significantly more effective and robust than current artificial neural network models. We will transfer recent knowledge from neuroscience to develop new machine learning methods that open for new scientific and technological developments. We will focus on three main approaches: (i) how representations such as grid-cells from biological neural networks can improve artificial network models, (ii) how to develop and apply effective learning rules in biologically inspired neural networks, (iii) and how spiking neural networks can define a new model for artificial networks and processing. To address these challenges, we have organized a cross-disciplinary team of leading experts in neurophysiology, machine learning, statistical physics, and computational science. Our goal is to build a strong interdisciplinary group that can impact AI research for a long time with rapid transfer from neuroscience to AI with a fully integrated approach with experiments, computations and AI. This will put us in an excellent position to develop biologically inspired machine learning methods for the future, and to educate the next generation of interdisciplinary researchers and technologists that will develop the next generation of AI science and applications. In 2023 we have three full time PhD-students and one adjunct professor employed on the project. The project has been developed in three main directions: (1) We have developed more robust artificial neural network model inspired by biological learning rules. (2) We have developed and studied models for navigation based on recurrent neural networks. In these networks, we show that grid-cells appear spontaneously, but that the cells the play the most prominent role in navigation are cells that are mapped onto a torus when the dimensionality of the system is reduced to three dimensions. Our first results from this work were published in 2023.

There has been an enormous development in artificial intelligence (AI) recently. However, while AI supersedes human skills in some fields, the brain is still far superior in many areas - it is more energy efficient, it needs fewer examples to learn, it can learn complex tasks easier, and it can transfer knowledge from one task to another. Recent progress has come in reinforcement learning and deep learning, which in principle are inspired by neurological processes, but in practice there are large differences between the way artificial and biological neural networks work. Our hypothesis is that artificial neural networks based on the structure, dynamics and learning rules in the brain will be significantly more effective than current models. We will transfer recent knowledge from neuroscience to develop new machine learning methods that open for new scientific and technological developments. We will address (i) how representations in biological networks can improve artificial network models, transfer and unsupervised learning models, (ii) how to develop and apply effective learning rules in biologically inspired neural networks, (iii) and how spiking neural networks can define a new model for artificial networks and processing, (iv) and develop a synergetic model that incorporates these features. To address these challenges, we have organized a cross-disciplinary team of leading experts in neurophysiology, machine learning, statistical physics, and computational science. We have experience in building and supervising cross-disciplinary breakthrough research, and have a broad international network of collaborators spanning neurophysiology, psychology, computational neuroscience and AI. This puts us in an excellent position to develop biologically inspired machine learning methods for the future, and to educate the next generation of interdisciplinary researchers and technologists that will develop the next generation of AI science and applications.