Researchers at Freie Universität Berlin and the University of Plymouth (UK) are developing biologically plausible neural network models to study human cognition
A team made up of researchers from Freie Universität Berlin and the University of Plymouth are investigating whether neural networks and algorithms based on artificial intelligence can help us to understand the foundations of human cognition. The researchers concluded that neural networks are only able to convincingly simulate and provide explanations for well-known cognitive phenomena when the networks demonstrate similarities with the human brain with regard to specific features. The results of the study were published in the scientific.
Neural networks and artificial intelligence algorithms have advanced leaps and bounds in the last decade. In some areas they have managed to replicate the capabilities of human beings: they can recognize faces, compose texts, produce images, and even drive cars. While artificial neural networks are designed with the human brain in mind, they are not a realistic neuroscientific representation of the brain in many respects. The human brain is a complex system in which different areas communicate by means of neurons. This is why properties of the brain that are crucial for a realistic model can range from the microscopic to the macroscopic level - from the individual neuron on the cell level to the way in which larger regions of the brain are connected. Researchers who want to model certain characteristics of the brain have to decide how detailed and realistic each level should be compared to the original.
Neuroscience professor Friedemann Pulvermüller and his team investigated different types of neural models and identified several factors that have characterized models up to now. On the microscopic cell level, this includes the choice of model neurons and mechanisms of synaptic plasticity and learning, while on the macroscopic level this can be the use of local and global control mechanisms for neural activity. Further aspects include the structure of the modeled areas of the brain, local within-area connectivity, and global between-area connectivity.
However, it is particularly important that the different levels are integrated into one individual model. "Models that form a bridge between the microscale and macroscales are an invaluable resource in neuroscience," explains Professor Pulvermüller. "Researchers are able to compare these models with physiological data from experiments." Models that have been empirically validated and are biologically plausible could be used to make predictions on how cognitive functions, like language processing and attention, are reorganized after a lesion. Personalized neural models that are constrained by the specific properties of individual brains could also be a useful aid in planning individual therapies and operations, for example, for people with brain tumors.
Recent research at Freie Universität Berlin has involved using neuroscientific modeling to explain well-known phenomena, such as the takeover of visual areas during language and cognitive processing in congenitally blind individuals.
On the basis of these findings, Professor Pulvermüller’s research team will further develop existing neurobiologically plausible neural networks within the European Research Council’s project "Material Constraints Enabling Human Cognition," which has received 2.5 million euros in funding. The group’s work is also receiving funding from the German Research Foundation through the Cluster of Excellence "Matters of Activity. Image Space Material."