Talk abstracts > Jean Daunizeau (ICM, Paris)

Go back to Homepage.

Jean Daunizeau - ICM, Paris

Thursday, June 1^st

Talk Session 1: How artificial neural networks make decisions

11h30 - 12h

 Does synaptic plasticity in the OFC explain how risk attitude adapts to the range of risk prospects?

Is irrational behavior the incidental outcome of biological constraints imposed on neural information processing? In particular, can biological constraints alter decisions, even when people know the consequences of alternative actions? Recent studies indicate that orbitofrontal neurons encode decision value in relative terms, i.e. value signals in OFC neurons are normalized with respect to the context. Value-based decisions may thus exhibit irrational context-dependence effects. A candidate explanation is "efficient coding": OFC neurons may mitigate information loss by adapting their (bounded) output firing properties to the recent value range. This is seducing, because it suggests that relative value coding is the brain’s best attempt to produce rational behavior, given its own hard-wired biological constraints. However, whether the behavioral implications of this scenario are met, how it generalizes to realistic situations in which OFC neurons construct value from multiple decision-relevant attributes, and what its neurophysiological bases are, is unclear. Here, we address these issues by re-analyzing two open fMRI datasets from the OpenNeuro.org initiative, where people have to integrate prospective gains and losses to decide whether to gamble or not. First, we show that peoples' risk attitudes critically depend on the range of gain prospects they are exposed to. Importantly, counter to simple efficient coding scenarios, differences in gain ranges induce progressive changes in peoples’ sensitivity to both gains and losses. Second, we use artificial neural network models to show that hebbian rewiring processes between attribute-specific and attribute-integration neurons predict (out-of-sample) both context-dependence effects in peoples’ risk attitude and multivariate patterns of fMRI activity in the OFC. Under mild conditions on distributed population codes for decision attributes, hebbian plasticity eventually translates the distribution of reweighted attribute signals towards the responsive range of integration neurons. In turn, integration units exhibit the known features of range adaptation, including (but not limited to), relative value coding. Our results demonstrate how hebbian plasticity within brain networks may result in range adaptation, eventually yielding complex though predictable irrational behavior.