Sensory perception is an active process: we change the way we sense the world by actively changing the physical properties of our sensory organs, but also our brain itself continually changes its properties in order to optimally adapt to the task at hand or the behavioural state. The process in which neuron and network properties are regulated by other brain areas, neuromodulation, is typically done in a top-down fashion: ‘higher’ brain areas, involved in complex and context-dependent forms of perception, influence the ‘lower’ brain areas, involved in the bottom-up processing of incoming sensory signals (inference), via specific chemical substances called neuromodulators. Only recently have the effects of these neuromodulators been measured in cortex and the data needed to analyze the structure and function of neural networks been recorded/simulated. So what the effects of neuromodulators are on the computational properties of cortical single cells, microcircuits and neural networks is still largely unknown. Therefore, the overarching research question of this project is:
How does top-down neuromodulation influence bottom-up sensory computations?
This question will be studied by building computational network models using three frameworks: 1) Balanced Networks, in which network activity can be studied as a function of connectivity and neuron properties; 2) FORCE learning, a form of machine learning in which balanced cortical networks learn a task by training an output and 3) Efficient Coding, in which a recurrent network is set up to perform a task close to optimally using a cooperative code. Scientists will make use of cellular and behavioural data gathered in our department over the previous 5 years on dopamine, acetylcholine and serotonin, to bridge the gap between single cell and behavioural effects. This will result in fundamental insights into the network computations responsible for perception and inference, needed both for understanding the brain and for AI applications.
