One of the most fundamental, and most studied, human cognitive functions is working memory. Yet, it is currently unknown how working memory is unified. In other words, why does a healthy human brain have one integrated capacity of working memory, rather than one capacity per visual hemifield, for instance. Thus, healthy subjects can memorize roughly as many items, regardless of whether all items are presented in one hemifield, rather than throughout two visual hemifields. In this current research, we investigated two patients in whom either most, or the entire, corpus callosum has been cut to alleviate otherwise untreatable epilepsy. Crucially, in both patients the anterior parts connecting the frontal and most of the parietal cortices, are entirely removed. This is essential, since it is often posited that working memory resides in these areas of the cortex. We found that despite the lack of direct connections between the frontal cortices in these patients, working memory capacity is similar regardless of whether stimuli are all presented in one visual hemifield or across two visual hemifields. This indicates that in the absence of the anterior parts of the corpus callosum working memory remains unified. Moreover, it is important to note that memory performance was not similar across visual fields. In fact, capacity was higher when items appeared in the left visual hemifield than when they appeared in the right visual hemifield. Visual information in the left hemifield is processed by the right hemisphere and vice versa. Therefore, this indicates that visual working memory is not symmetric, with the right hemisphere having a superior visual working memory. Nonetheless, a (subcortical) bottleneck apparently causes visual working memory to be integrated, such that capacity does not increase when items are presented in two, rather than one, visual hemifield.

Modern Machine Learning (ML) has significantly advanced various research fields, but the opaque nature of ML models hinders their adoption in several domains. Explainable AI (XAI) addresses this challenge by providing additional information to help users understand the internal decision-making process of ML models. In the field of neuroscience, enriching a ML model for brain decoding with attribution-based XAI techniques means being able to highlight which brain areas correlate with the task at hand, thus offering valuable insights to domain experts. In this paper, we analyze human and Computer Vision (CV) systems in parallel, training and explaining two ML models based respectively on functional Magnetic Resonance Imaging (fMRI) and movie frames. We do so by leveraging the "StudyForrest" dataset, which includes functional Magnetic Resonance Imaging (fMRI) scans of subjects watching the "Forrest Gump" movie, emotion annotations, and eye-tracking data. For human vision the ML task is to link fMRI data with emotional annotations, and the explanations highlight the brain regions strongly correlated with the label. On the other hand, for computer vision, the input data is movie frames, and the explanations are pixel-level heatmaps. We cross-analyzed our results, linking human attention (obtained through eye-tracking) with XAI saliency on CV models and brain region activations. We show how a parallel analysis of human and computer vision can provide useful information for both the neuroscience community (allocation theory) and the ML community (biological plausibility of convolutional models).

The human brain processes visual input across various spatial frequency (SF) ranges to extract emotional cues. Prior studies have extensively explored SF processing in facial expressions, yielding partly conflicting results. However, bodily expressions, which provide complementary emotional and survival-relevant cues, remain unexplored. We investigated the neural mechanisms underlying the processing of low (LSF), high (HSF), and broad spatial frequency (BSF) components in fearful versus neutral bodily postures. Using functional Magnetic Resonance Imaging, we examined brain activity in 20 participants viewing SF-filtered images of bodily expressions in a semi-passive task. A multivariate “searchlight” analysis based on Multi-Voxel Pattern Analysis was employed to decode the non-linear activation patterns associated with each SF band. Our findings reveal that SF processing engages distinct neural networks in response to fearful bodily expressions. BSF stimuli activated a widespread network, including the amygdala, pulvinar, frontal, and temporal cortices. These findings suggest a general threat-detection system integrating information across all SFs. HSF stimuli engaged cortical regions associated with detailed emotional evaluation and motor planning, such as the orbitofrontal cortex, anterior cingulate cortex, and premotor areas, suggesting that processing fine-grained fear cues involves computationally demanding networks related to emotional resonance and action preparation. In contrast, LSF stimuli primarily activated motor-preparatory regions linked to rapid, action-oriented responses, highlighting the brain prioritization of quick readiness to low-detail threats. Notably, the amygdala showed no SF selectivity, supporting its role as a generalized “relevance detector” in emotional processing. The present study demonstrates that the brain flexibly adapts its SF processing strategy based on the visual details available in fearful bodily expressions, underscoring the complexity and adaptability of emotional processing from bodily signals.