Two computational problems which are trivial for symbol-manipulating systems but which pose serious challenges to connectionist networks are the binding problem and the type-token problem. These difficulties arise because representations in connectionist networks do not automatically i) specify which features go with which object tokens, or ii^ distinguish between different tokens of the same type. Nevertheless, these processing shortcomings may constitute advantages when connectiomst networks are taken as models of human visual information processing. Perception research shows evidence not only of binding errors, for example in Treisman's illusory conjunctions (Treisman and Schmidt, 1982), but also of type-token errors, as seen in repetition blindness (Kanwisher, 1987) and other phenomena.

Faces are an important source of perceptual and social information. Multiple cortical regions including the fusiform faces area (FFA), the superior temporal sulcus (STS), and medial prefrontal cortex (MPFC) respond more to dynamic faces than videos of toy objects, human bodies, and pastoral scenes. Do face-preferring regions in FFA, STS, and MPFC have different functions? To address this question, we re-analyzed functional magnetic resonance imaging (fMRI) data from seven different experiments that included a dynamic faces versus objects localizer. Each of the seven experiments tested different perceptual and social features using dynamic videos, point light displays, narratives, and animated cartoons. Using a functional region of interest approach, we observed a significant condition by region interaction in four of the seven experiments. Thus, although FFA, STS, and MPFC respond more to dynamic faces than objects, bodies, and scenes, these three regions differ from each other functionally.

Humans demonstrate a remarkable ability to infer physical properties of objects and predict physical events in dynamicscenes. These abilities have been modeled as probabilistic simulations of a mental physics engine akin to 3D physicsengines used in computer simulations and video games (Battaglia, Hamrick & Tenenbaum 2013; Sanborn, Mansinghka &Griffiths 2013), but it is unknown if and how such a physics engine is implemented in the brain. Does the brain representquantities corresponding to the key latent variables of physical objects that contribute to their dynamics? To find out,we used multivariate pattern classification analyses of fMRI data from subjects viewing videos of dynamic objects. Themass of depicted objects could be decoded, across physical scenarios and object materials, from brain regions previouslyimplicated in intuitive physics. This invariant representation of mass may serve as a key variable in a generalized enginefor intuitive physics.

Humans are profoundly social creatures. We depend on others for survival and crave social interactions. Faces are thegateway to many typical social interactions. One of the most replicable results in cognitive neuroscience is the selectiveresponse of some cortical regions to faces in humans and other social primates. Specifically, the fusiform face area (FFA)is a region in the ventral temporal cortex (VTC) that is selectively responsive to faces. To determine whether infantsshow early cortical responses to faces, we recruited 86 human infants (2.1-11.9 months) to participate in an awake infantfunctional resonance imaging (fMRI) experiment. We obtained usable fMRI data from 49 infants (2.1-9.7 months) whilethey watched videos of faces, bodies, objects, and scenes, 30 of whom (2.5-9.4 months) had enough data for a functionalregion of interest (fROI) analysis. A group random effects analysis revealed significantly higher responses to faces thanobjects in the medial prefrontal cortex (MPFC), superior temporal sulcus (STS), ventral temporal cortex (VTC), andsubcortical areas of the infant brain. Additionally, the fROI analysis revealed face selective responses in the STS, MPFC,and VTC but not lateral occipital cortex or subcortical areas of the infant brain. Thus, we provide the first evidence offace selective responses in the infant brain and demonstrate that social regions (STS and MPFC) respond selectively at thesame time as perceptual regions (VTC).

Causal reasoning enables us to explain the past, predict the future, and intervene in the present. Does the brain allocate specialized cortical regions to causal reasoning? And if so, are they involved in reasoning about both physical and social causal relationships, or are they domain-specific? In a pre-registered experiment (Exp 1) we scanned adults using fMRI while they matched physical and social causes to effects (e.g., ‘The car swerved to avoid a crash' -> ‘Coffee spilled all over the car seat'; ‘He was late for work' -> ‘Tom was scolded by his boss') or physical and social descriptions of the same entity matched for difficulty and linguistic variables to the causal conditions (e.g., ‘The brightest object in the sky'-> ‘The closest star to earth'; ‘She works at a hotel' -> ‘She brings in guests' luggage'). A region in the left lateral prefrontal cortex (LPFC) responded significantly more strongly to causal than descriptive conditions in most subjects individually. Responses in this region in held-out data were high for both social and physical causal conditions, yet no greater than baseline for the two descriptive (non-causal) conditions. In a follow-up exploratory experiment (Exp 2), we tested a different task (answering causal versus non-causal questions about physical and social narratives, matched for linguistic variables). Again, we found that both the physical and social causal stimuli selectively engaged the LPFC region. Finally, in both experiments, we found that brain regions previously implicated in intuitive physical reasoning responded more to the physical causal than the physical non-causal stimuli. Collectively, these results suggest that a) a region in the LPFC is selectively engaged in causal reasoning independent of content domain and b) the hypothesized physics network (hPN) is selectively involved in physical causal reasoning across modalities (visual vs. linguistic).

The neuroscience of perception has recently been revolutionized with an integrative modeling approach in which computation, brain function, and behavior are linked across many datasets and many computational models. By revealing trends across models, this approach yields novel insights into cognitive and neural mechanisms in the target domain. We here present a systematic study taking this approach to higher-level cognition: human language processing, our species' signature cognitive skill. We find that the most powerful "transformer" models predict nearly 100% of explainable variance in neural responses to sentences and generalize across different datasets and imaging modalities (functional MRI and electrocorticography). Models' neural fits ("brain score") and fits to behavioral responses are both strongly correlated with model accuracy on the next-word prediction task (but not other language tasks). Model architecture appears to substantially contribute to neural fit. These results provide computationally explicit evidence that predictive processing fundamentally shapes the language comprehension mechanisms in the human brain.