Decision-making and learning – neural mechanisms and pharmacology
The overarching question of this part of our work is how different brain networks learn in parallel with different computational advantages, and how the brain then integrates this plethora of information into a single coherent behaviour. We tackle this question by measuring brain activity in humans, causally manipulating different neurotransmitter systems, and using computational modelling to tease apart and precisely measure different cognitive processes.
Multiple prefrontal learning systems (Scholl et al., 2015)
Until recently models and experiments on learning have often focused on very simple scenarios such as associations of single stimuli with outcomes. This however neglects that the decision problems humans face in real life are often far more complex. In fact, often multiple pieces of information need to be learnt about simultaneously with the brain keeping these representations separate and preventing other experiences or emotions from interfering with them. To test how the brain solves this problem, I combined computational modelling and brain imaging (FMRI). In fact, people found keeping all information separate difficult, misattributing luck to their choices. One reason such a bias might emerge is because of the strong reward-related brain signals in many regions making it hard to keep this information suppressed even when it is irrelevant. Thus, I wondered whether other brain areas carried signals compensating for this bias. Anterior prefrontal cortex (aPFC) was as a key candidate for this given its role in emotional control. Indeed, I found that aPFC was crucial in overcoming the bias.
Serotonin’s impact on learning and plasticity (Scholl et al., 2017 )
It has been proposed that how plastic the brain is and how easily it can learn is regulated through neuromodulators. Work in animals in particular highlighted the role of tonic levels of the serotonin. However, it is unclear how well these plasticity effects translate to humans. Coincidentally, serotonin is currently the primary pharmacological treatment target for depression, via serotonin re-uptake inhibitors (SSRIs). Here, we tested whether serotonin increases, over prolonged time, improved learning in humans, in a cognitive domain relevant for depression, namely reward and effort learning. We found that sub-chronic administration (14 days) of a standard clinical dose of an SSRI (vs. placebo) in healthy increased learning signals for both reward and effort in several brain areas. Thus, my findings suggested that changes in tonic serotonin levels affect participants’ general plasticity rather than simply emphasizing positive over negative value signals. This contrasted with a common view at the time that serotonin is relevant only for learning about aversive stimuli.
NMDA receptors and decision-making (Scholl et al., 2014 )
After learning, we have to decide how to use what we have learnt and weight it up against alternative sources of information. Human (and animals) can do that in various ways that are more or less optimal given the situation. For example, if you are asked whether you prefer to spin a wheel of fortune with a 90% chance of winning 10 euros (and otherwise zero) or one with a 5% chance of winning 100 euros (and otherwise zero), rationally, you should prefer the first one (as 90%*10 euros is more than 5%*100 euros). However, if you are not quite sure that the probabilities are really 90% and 5% - maybe you never see the wheel of fortune, but have to learn them from a series of outcomes – your preferences might change. Here, with Andrea Reinecke, we found that brain receptors (NMDA) can change what strategies people use this in a pharmacological study. We found that the partial NMDA-agonist d-cycloserine (medical use: antituberculosis medication, off-label use for anxiety disorder) changed people's strategy from linear (i.e. comparing the two options separately on probabilities of winning and amount to be won) to non-linear (i.e. comparing the multiplied together probabilities and magnitudes).
Excitation-inhibition balance and using what you have learnt to make decisions (Scholl et al., 2017)
Non-linear integration could theoretically allow up- or down-regulating of how much learnt relative to new information is used. However, it was still unclear which brain region was particularly relevant for this. Here, I tested the hypothesis that natural variations in dorsal anterior cingulate cortex (dACC) excitatory-inhibitory (E-I) neurotransmitter balance regulates how much learnt information is used. I measured the neurotransmitters glutamate and GABA in dACC in humans using magnetic resonance spectroscopy at baseline, before they did the same learning task as in the studies above.
And indeed, higher levels of glutamate and lower levels of GABA in dACC were associated with both increased strength of the value to be learnt in dACC and increased use of learnt information over new information when making choices. Taken together, these results suggest that dACC plays a crucial role in allowing information that has been learnt to influence behaviour and that, on a molecular level, this may be realised by regulating the E-I balance.
