Towards a unifying account of generalization in cognitive neuroscience

Abstract

In a dynamic environment, the ability to adapt to changes is a key factor for survival. This cognitive ability is called generalization and the fact that humans excel at it is a major reason for our evolutionary success. Commensurate with this, the literature on human generalization is vast, but our understanding suffers from a lack of a unifying framework. Instead, generalization in associative learning, in reinforcement learning and inductive reasoning are implicitly treated as separate entities. Stimulus generalization, i.e. generalization of associative learning is most often studied in the context of fear conditioning and the explanatory focus is typically strongly concerned with the role of perception. Generalization in reinforcement learning is usually assumed to rely on abstractions of state spaces, e.g. via dimensionality reduction. Lastly, research on inductive reasoning proposes different Bayesian mechanisms. To arrive at a more general mechanism of generalization, I propose a Bayesian model of generalization that integrates dimensionality reduction into a probabilistic framework and is applicable to probabilistic reinforcement and therefore the typical study designs in stimulus generalization. To test the predictions of the model and to find common ground between stimulus generalization and abstraction in reinforcement learning with respect to their neural signature, I conducted a series of experiments. Importantly, I used face stimuli that differed on facial identity and facial expression, which allowed me to investigate dimensional preferences and their relationship with prior knowledge. In addition, the study design contained time-resolved ratings to characterize the temporal dynamics. Using the proposed model I could then make specific predictions for the data that I expected to arise from those studies. Behavioral ratings closely followed the predictions of the model in all three studies, independently of the value of the reinforcement. Initial ratings were dependent on the value and strength of the emotional expression but became increasingly dependent on the perceptual similarity to the reinforced stimulus. The latter effect was stronger along the emotion dimension in all studies. Model comparison confirmed that the data followed all of the predictions of the model. Using fMRI data from one of the studies, I found positive correlations with behavioral generalization in the frontoparietal attention network and the salience network and negative correlations in the default mode network. Importantly, generalization along the emotion dimension was only associated with the frontoparietal attention network and the salience network, which mirrors results from reinforcement learning. In a last step, I found that representations in the middle frontal gyrus mirrored the behavioral relevance of the different dimensions. Taken together, I present coherent theoretical considerations and empirical evidence for a common mechanism of generalization that can be well explained as a Bayesian model and suggest that low-dimensional representations of stimuli are one key neural mechanism underlying generalization.