The Free Energy Theory of Emotions – Fr-E-Emo

The theory we adopt in our project is a formulation of human emotions in terms of free energy (Friston, Kilner, & Harrison, 2006). We propose that emotions are caused by the dynamic (mis)match between expected and observed sensorial inputs. In a world of uncertainty, environmental quantities (external states) and internal model’s parameters (internal states) are continuously changing over time. Biological agents are actively seeking to bring these two states close together when suppressing free energy. In the present project, we propose a definition of emotions based on the rate of change of free energy over time. The free energy theory offers a unifying approach to action, perception and learning. Despite the intimate relationship between emotions and action-perception-learning, emotions have not yet been defined under this same formulation. With our project we intend to fill this gap.

In all experiments we found that subjects significantly learn to direct their choices in order to maximize the payoff. Learning was not affected by reinforcer’s valance, meaning that subjects learnt equally well to avoid punishments and to seek rewards. On the contrary, and as expected, the level of information at the feedback significantly affected learning, indicating that subject successfully integrated the fictive information. Leaning performances were significantly affected by outcome probability. Reinforcer magnitude also affected learning performances, but in a less extend compared to probability. In the subsequent test in which subjects has to assess the relative “goodness” of a symbol, subjects were able to discriminate the most rewarding and the most punishing stimulus. The “intermediate” stimuli, the less rewarding and the most punishing, were not or poorly discriminated and rated around zero, suggesting that subjected encoded not only the value (reward vs. punishment), but also the correctness of a symbol (good vs, bad). The behavioral results of the imaging studies (fMRI and MEG), very nicely replicated the results of the psychophysical study. Analysis of the imaging data with standard reinforcement learning model (Q-Learning) replicated previous results implicated the ventromedial prefrontal cortex and the ventral striatum in positive value learning and the anterior insula and the dorsal anterior cingulated cortex in the negative value learning.

Next analyses will be focused on more advanced computational modeling (e.g., Free-energy model) of the data.

-Joffily, M. and Coricelli G. (2013). Emotional valence and the free-energy principle. PLoS Computational Biology. Jun;9(6):e1003094

-Coricelli, G. and Joffily, M. (2012). Cognitive-based emotions: theory and evidence from the brain. In F. Paglieri, L. Tummolini, M. Miceli, R. Falcone (Eds.), The goals of cognition: essays in honor of Cristiano Castelfranchi. London: College Publications. 203-209.

Human decisions cannot be explained solely by rational imperatives but are strongly influenced by emotions. Here we intend to combine decision theory, psychology and neuroscience in order to study the emergence of emotions in decision-making using computational and neuroimaging methods. The theory we adopt in our project is a formulation of human emotions in terms of free energy (Friston, Kilner, & Harrison, 2006). We propose that emotions are caused by the dynamic (mis)match between expected and observed sensorial inputs. In a world of uncertainty, environmental quantities (external states) and internal model’s parameters (internal states) are continuously changing over time. Biological agents are actively seeking to bring these two states close together when suppressing free energy. In the present project, we propose a definition of emotions based on the rate of change of free energy over time. The free energy theory offers a unifying approach to action, perception and learning. Despite the intimate relationship between emotions and action-perception-learning, emotions have not yet been defined under this same formulation. With our project we intend to fill this gap.